Gilles Pesant (Ed.

)
LNCS 9255

Principles and Practice

of Constraint Programming
21st International Conference, CP 2015
Cork, Ireland, August 31 – September 4, 2015
Proceedings

123
Lecture Notes in Computer Science 9255
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen

Editorial Board
David Hutchison
Lancaster University, Lancaster, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Friedemann Mattern
ETH Zurich, Zürich, Switzerland
John C. Mitchell
Stanford University, Stanford, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
C. Pandu Rangan
Indian Institute of Technology, Madras, India
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/7408
Gilles Pesant (Ed.)

Principles and Practice

of Constraint Programming
21st International Conference, CP 2015
Cork, Ireland, August 31 – September 4, 2015
Proceedings

123
Editor
Gilles Pesant
École Polytechnique de Montréal
Montréal, Québec
Canada

ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Computer Science
ISBN 978-3-319-23218-8 ISBN 978-3-319-23219-5 (eBook)
DOI 10.1007/978-3-319-23219-5

Library of Congress Control Number: 2015946574

Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made.

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
 Preface

This volume contains the proceedings of the 21st International Conference on the
Principles and Practice of Constraint Programming (CP 2015), which was held in Cork,
Ireland, from August 31 to September 4, 2015. Detailed information about the con-
ference is available at http://cp2015.a4cp.org. The CP conference is the annual inter-
national conference on constraint programming. It is concerned with all aspects of
computing with constraints, including theory, algorithms, environments, languages,
models, systems, and applications such as decision making, resource allocation,
scheduling, configuration, and planning.
For the purpose of the conference’s scientific programming, we invited submissions
to the technical, application, and published journal tracks. We received 80, 25, and 14
submissions to these tracks respectively. Authors chose to submit either long (15 pages)
or short (8 pages) papers to the technical and application tracks. The review process for
the technical track relied on a two-level Program Committee and on additional reviewers
recruited by Program Committee members. Each submission to the technical track was
assigned to one member of the Senior Program Committee and three members of the
track’s Program Committee. Submissions to the application track were each assigned to
three members of its Program Committee. Every paper received at least three reviews.
Once the initial reviews were in, authors were given an opportunity to respond before a
detailed discussion was undertaken at the level of the Program Committees, overseen by
the Program Chair and the Senior Program Committee member or the Application Track
Chair, as appropriate. For the first time this year the Senior Program Committee did not
meet in person but deliberated by video conference instead. The published journal track
gives an opportunity to discuss important results in the area of constraint programming
that appeared recently in relevant journals, but had not been previously presented to the
community at conferences. Submissions were evaluated by a separate Program Com-
mittee for relevance and significance.
At the end of the reviewing process, we accepted 39 papers from the technical track,
15 papers from the application track, and 10 papers from the published journal paper
track. The Senior Program Committee awarded the Best Paper Prize to John N. Hooker
for “Projection, Consistency, and George Boole” and the Best Student Paper Prize to
Alexander Ivrii, Sharad Malik, Kuldeep S. Meel, and Moshe Vardi for “On Computing
Minimal Independent Support and Its Applications to Sampling and Counting.” The
Application Track Program Committee awarded the Best Application Paper Prize to
Tommaso Urli and Philip Kilby for “Long-Haul Fleet Mix and Routing Optimisation
with Constraint Programming and Large Neighbourhood Search.” The Program Chair
and the Constraints journal editor-in-chief, Michela Milano, also invited six papers
from the technical and application tracks for direct publication in that journal. These
were presented at the conference like any other paper and they appear in the pro-
ceedings as a one-page abstract.
VI Preface

This edition of the conference was part of George Boole 200, a celebration of the
life and work of George Boole who was born in 1815 and worked at the University
College of Cork. It was also co-located with the 31st International Conference on Logic
Programming (ICLP 2015). The conference program featured three invited talks by
Claire Bagley, Gerhard Friedrich (joint with ICLP), and Douglas R. Smith. This vol-
ume includes one-page abstracts of their talks. The conference also featured four
tutorials and six satellite workshops, whose topics are listed in this volume. The
Doctoral Program gave PhD. students an opportunity to present their work to more
senior researchers, to meet with an assigned mentor for advice on their research and
early career, to attend special tutorials, and to interact with one another. The winners
of the 2015 ACP Research Excellence Award and Doctoral Research Award presented
their award talks. The results of the annual MiniZinc Challenge were announced and a
joint ICLP/CP Programming Competition was held. For the first time an Industry
Outreach effort, headed by Helmut Simonis, provided formal opportunities for CP
researchers and representatives from industry to interact: The former were presented
with industrial problems to solve in a friendly competition and the latter were presented
with samples of industrial problems solved using CP technology. Following the con-
ference, a CSPlib sprint event took place to update this library of constraint satisfaction
problems. All these contributed to a very exciting program.
I am grateful to many people who made this conference such a success. First of all, to
the authors who provided excellent material to select from. Then to the members of the
Program Committees and additional reviewers who worked hard to provide constructive,
high-quality reviews. To members of the Senior Program Committee who helped me
ensure that each paper was adequately discussed, wrote meta-reviews for their assigned
papers, and participated in live remote deliberations — for some, quite early or late in the
day. To Christian Schulte in particular, who agreed to step in where I had a conflict of
interest. Of course there is a whole team standing with me, who chaired various aspects
of the conference: Ken Brown, Barry O’Sullivan (Conference Chairs) and their own local
team (Barry Hurley as webmaster; Kathy Bunney who provided coordinating support; the
technical and administrative support staff of the Insight Centre for Data Analytics at
UCC), Louis-Martin Rousseau (Application Track Chair), Thomas Schiex (Published
Journal Track Chair), David Bergman and Marie Pelleau (Doctoral Program Chairs),
Willem-Jan van Hoeve (Workshop and Tutorial Chair), Helmut Simonis (Industry Out-
reach Chair), and Ian Miguel (Publicity Chair). Thank you for your dedication! I
acknowledge and thank our sponsors for their generous support: they include, at the time
of writing, the Association for Constraint Programming, the Association for Logic Pro-
gramming, ECCAI — the European Coordinating Committee for Artificial Intelligence,
the Insight Centre for Data Analytics, Science Foundation Ireland, Springer, and Uni-
versity College Cork. Finally, I thank the ACP Executive Committee for the trust they
showed me in asking me to serve as Program Chair. It has been an honor.

June 2015 Gilles Pesant

 Tutorials and Workshops

Tutorials

Constraints and Bioinformatics: Results and Challenges

Agostino Dovier
Lagrangian Relaxation for Domain Filtering
Hadrien Cambazard
Towards Embedded Answer Set Solving
Torsten Schaub
XCSP3
Frédéric Boussemart, Christophe Lecoutre, and Cédric Piette

Workshops

Workshop on Constraint-Based Methods in Bioinformatics (WCB 2015)

Alessandro Dal Palù and Agostino Dovier
6th International Workshop on Bin Packing and Placement Constraints (BPPC 2015)
Nicolas Beldiceanu and François Fages
5th International Workshop on the Cross-Fertilization Between CSP and SAT
(CSPSAT 2015)
Yael Ben-Haim, Valentin Mayer-Eichberger, and Yehuda Naveh
14th International Workshop on Constraint Modelling and Reformulation
(ModRef 2015)
Ozgur Akgun and Peter Nightingale
Workshop on Teaching Constraint Programming
Alan Frisch, Ciaran McCreesh, Karen Petrie, and Patrick Prosser
CP and Analytics
Youssef Hamadi and Willem-Jan van Hoeve
 Conference Organization

Conference Chairs
Ken Brown University College Cork, Ireland
Barry O’Sullivan University College Cork, Ireland

Program Chair
Gilles Pesant École Polytechnique de Montréal, Canada

Application Track Chair

Louis-Martin Rousseau École Polytechnique de Montréal, Canada

Published Journal Track Chair

Thomas Schiex INRA Toulouse, France

Doctoral Program Chairs

David Bergman University of Connecticut, USA
Marie Pelleau Université de Montréal, Canada

Workshop and Tutorial Chair

Willem-Jan van Hoeve Carnegie Mellon University, USA

Industry Outreach Chair

Helmut Simonis Insight Centre for Data Analytics, UCC, Ireland

Publicity Chair
Ian Miguel University of St Andrews, UK

Senior Program Committee

Chris Beck University of Toronto, Canada
Nicolas Beldiceanu TASC (CNRS/Inria), Mines Nantes, France
Christian Bessiere CNRS, France
Yves Deville UCLouvain, Belgium
X Conference Organization

Pierre Flener Uppsala University, Sweden

John Hooker Carnegie Mellon University, USA
Peter Jeavons University of Oxford, UK
Christophe Lecoutre CRIL, University of Artois, France
Jimmy Lee The Chinese University of Hong Kong, Hong Kong,
SAR China
Michela Milano DISI Università di Bologna, Italy
Jean-Charles Régin University of Nice Sophia Antipolis / CNRS, France
Christian Schulte KTH Royal Institute of Technology, Sweden
Peter J. Stuckey NICTA and the University of Melbourne, Australia
Pascal Van Hentenryck NICTA and ANU, Australia
Willem-Jan van Hoeve Carnegie Mellon University, USA
Toby Walsh NICTA and UNSW, Australia
Roland Yap National University of Singapore, Singapore

Technical Track Program Committee

Carlos Ansótegui Universitat de Lleida, Spain
Fahiem Bacchus University of Toronto, Canada
Pedro Barahona Universidade Nova de Lisboa, Portugal
Roman Bartak Charles University in Prague, Czech Republic
David Bergman University of Connecticut, USA
Hadrien Cambazard Grenoble INP, CNRS, Joseph Fourier University, France
Hubie Chen Universidad del País Vasco and Ikerbasque, Spain
Geoffrey Chu NICTA and the University of Melbourne, Australia
David Cohen Royal Holloway, University of London, UK
Remi Coletta University of Montpellier, France
Martin Cooper IRIT - Universitíé Paul Sabatier, France
Sophie Demassey CMA, MINES ParisTech, France
François Fages Inria Paris-Rocquencourt, France
Alan Frisch University of York, UK
Maria Garcia De La Monash University, Australia
Banda
Arnaud Gotlieb SIMULA Research Laboratory, Norway
Stefano Gualandi Università di Pavia, Italy
Emmanuel Hebrard LAAS, CNRS, France
Philippe Jégou LSIS - UMR CNRS 7296 - Aix-Marseille University,
France
George Katsirelos INRA, Toulouse, France
Zeynep Kiziltan Università di Bologna, Italy
Lars Kotthoff University of British Columbia, Canada
Philippe Laborie IBM, France
Michele Lombardi DISI Università di Bologna, Italy
Xavier Lorca Ecole des Mines de Nantes, France
Inês Lynce Inst. Superior Técnico INESC-ID Lisboa, Portugal
Arnaud Malapert University of Nice Sophia Antipolis / CNRS, France
 Conference Organization XI

Joao Marques-Silva University College Dublin, Ireland

Ian Miguel University of St. Andrews, UK
Justin Pearson Uppsala University, Sweden
Justyna Petke University College London, UK
Steve Prestwich Insight, UCC, Ireland
Patrick Prosser Glasgow University, UK
Claude-Guy Quimper Université Laval, Canada
Andrea Rendl NICTA and Monash University, Australia
Michel Rueher University of Nice Sophia Antipolis / CNRS, France
Marius Silaghi Florida Institute of Technology, USA
Stephen Smith Carnegie Mellon University, USA
Christine Solnon LIRIS CNRS UMR 5205 / INSA Lyon, France
Kostas Stergiou University of Western Macedonia, Greece
Guido Tack Monash University, Australia
Charlotte Truchet LINA, UMR 6241, Université de Nantes, France
Nic Wilson Insight, UCC, Ireland

Application Track Program Committee

Claire Bagley Oracle Corporation, USA
Thierry Benoist Innovation 24 - LocalSolver, France
Mats Carlsson SICS, Sweden
Jean-Guillaume Fages COSLING S.A.S., France
Carmen Gervet Université de Savoie, LISTIC, France
Laurent Perron Google, France
Ashish Sabharwal AI2, USA
Pierre Schaus UCLouvain, Belgium
Paul Shaw IBM, France
Helmut Simonis Insight, UCC, Ireland
Peter van Beek University of Waterloo, Canada
Mark Wallace Monash University, Australia
Tallys Yunes University of Miami, USA

Published Journal Track Program Committee

Simon de Givry MIAT-INRA, France
Yves Deville UCLouvain, Belgium
John Hooker Carnegie Mellon University, USA
Michela Milano DEIS Università di Bologna, Italy
Nina Narodytska Carnegie Mellon University, USA
Patrick Prosser Glasgow University, UK
Francesca Rossi University of Padova, Italy
Roland Yap Hock Chuan National University of Singapore, Singapore
Invited Talks
 Constraint-based Problems and Solutions
in the Global Enterprise

Claire Bagley

Advanced Constraint Technology

Oracle Corporation, Burlington, MA 01803, USA
{claire.bagley@oracle.com}

Oracle is a large global technology organization, whose product offerings have grown
organically and through the acquisition of first class companies. Its leadership has
expanded to the entire technology stack, to span the full range of computer hardware,
operating systems, programming languages, databases, enterprise software, collabora-
tion management tools, and into the cloud.
With thousands of products and tools running the full imaginable breadth and depth
of a technology stack, it is then not surprising that Oracle is exposed to a vast number
of complex combinatorial problems, as well as the different types of technology to
solve them. Indeed, many of the classical applications and variations of constraint
problems are represented: planning, scheduling, rostering, vehicle routing, configura-
tion, networking, grid optimization, logistics, analytics, and cloud management. As
expected with the development of products and the acquisition of companies operating
in these domains, a large number of technologies come into play including Constraint
Programming (CP), Mathematical Programming (MP), local search, heuristics,
knowledge-based reasoning, genetic algorithms, machine learning, and many more.
The Advanced Constraint Technology (ACT) group at Oracle is tasked with
identifying, understanding and solving the complex combinatorial problems that arise
in a diverse field of application environments. Our expertise and industry proven ACT
products are available to assist Oracle development teams on how to best model and
solve their problems using CP, MP, Heuristics, and Hybrid solutions.
In this talk we examine some of the successful solutions to constrained problems
within such a large corporation. We discuss at a high level the many opportunities for
further integration and unification of constraint technologies into more products and
tools, including the challenge of modeling and solving in highly interactive scenarios.
Most importantly, we open the discussion about various challenges faced by large
front-end centered companies, with many degrees of separation between organizations,
who must balance resources to focus over immediate and long-term deliverables.
 Industrial Success Stories of ASP and CP:
What’s Still Open?

Gerhard Friedrich

Institut für Angewandte Informatik

Alpen-Adria Universität Klagenfurt, Austria
e-mail: gerhard.friedrich@aau.at

Abstract More than 25 years ago together with Siemens we started to inves-
tigate the possibility of substituting the classical rule-based configuration
approach by model- based techniques. It turned out that in those days only
constrained programming (CP) had any real chance of meeting the application
demands. By exploiting CP we were able to significantly improve the produc-
tivity of highly trained employees (by more than 300 %) and to substantially
reduce software development and maintenance costs (by more than 80 %) [4].
Consequently, CP has been our method of choice for problem solving in
industrial projects since 1989 [3].

Some years ago, we started to investigate answer set programming (ASP) techniques
[2], mainly because of the possibility to apply a very expressive logical first-order
language for specifying problems. It emerged that, by using simply problem encoding,
we were able to solve difficult real world problem instances witnessing the enormous
improvements of logic programming over the last decades [1].
Although ASP and CP have proven their practical applicability, we will point out
challenges of large problems of the electronic and the semiconductor industry. In
particular, we will stress the power of problem-specific heuristics [5, 9] which turned
out to be the key in many applications of problem solvers.
Looking at the famous equation “algorithm = logic + control” [6] most of the current
work in the AI community assumes that control should be problem- independent and
only the logical specification depends on the problem to be solved, i.e. “algorithm =
logic(problem) + control”. It is not surprising that for the current problem solving
technology this is a practical approach up to a certain size of the problem instances, since
we deal with NP-hard problems in many cases. However, it is observed (and examples
are given [10, 7]) that problem-specific heuristics allow enormous run-time improve-
ments. This success is based on problem-specific control, i.e. “algorithm = logic
(problem) + control(problem)”. Unfortunately, the design of such problem-specific
heuristics is very time-consuming and redesigns are frequently required because of
recurrent changes of the problem. Interestingly, humans are very successful at devel-
oping such problem-specific heuristics. Therefore, we argue that the automation of
generating problem-specific heuristics with satisfying quality is still an important basic
AI research goal with high practical impact that should be achievable [8].
 Industrial Success Stories of ASP and CP XVII

References
1. Aschinger, M., Drescher, C., Friedrich, G., Gottlob, G., Jeavons, P., Ryabokon, A., Thor-
stensen, E.: Optimization methods for the partner units problem. In: Achterberg, T., Beck, J.C.
(eds.) CPAIOR 2011. LNCS, vol. 6697, pp. 4–19. Springer, Heidelberg (2011). http://dx.doi.
org/10.1007/978-3-642-21311-3
2. Brewka, G., Eiter, T., Truszczynski, M.: Answer set programming at a glance. Commun.
ACM 54(12), 92–103 (2011). http://doi.acm.org/10.1145/2043174.2043195
3. Falkner, A., Haselboeck, A., Schenner, G., Schreiner, H.: Benefits from three configurator
generations. In: Blecker, T., Edwards, K., Friedrich, G., Hvam, L., Salvodor, F. (eds.)
Innovative Processes and Products for Mass Customization, vol. 3, pp. 89–103 (2007)
4. Fleischanderl, G., Friedrich, G., Haselböck, A., Schreiner, H., Stumptner, M.: Configuring
large systems using generative constraint satisfaction. IEEE Intell. Syst. 13(4), 59–68 (1998)
5. Gebser, M., Kaufmann, B., Romero, J., Otero, R., Schaub, T., Wanko, P.: Domain-specific
heuristics in answer set programming. In: desJardins, M., Littman, M.L. (eds.) Proceedings
of the Twenty-Seventh AAAI Conference on Artificial Intelligence, July 14–18, 2013,
Bellevue, Washington, USA. AAAI Press (2013). http://www.aaai.org/ocs/index.php/AAAI/
AAAI13/paper/view/6278
6. Kowalski, R.A.: Algorithm = logic + control. Commun. ACM 22(7), 424–436 (1979). http://
doi.acm.org/10.1145/359131.359136
7. Mersheeva, V., Friedrich, G.: Multi-uav monitoring with priorities and limited energy
resources. In: Brafman, R.I., Domshlak, C., Haslum, P., Zilberstein, S. (eds.) Proceedings
of the Twenty-Fifth International Conference on Automated Planning and Scheduling,
ICAPS 2015, Jerusalem, Israel, June 7-11, 2015, ICAPS 2015. pp. 347–356. AAAI Press
(2015). http://www.aaai.org/ocs/index.php/ICAPS/ICAPS15/paper/view/10460
8. Pearl, J.: On the discovery and generation of certain heuristics. AI Mag. 4(1), 23–33 (1983).
http://www.aaai.org/ojs/index.php/aimagazine/article/view/385
9. Schrijvers, T., Tack, G., Wuille, P., Samulowitz, H., Stuckey, P.J.: Search combinators.
Constraints 18(2), 269–305 (2013). http://dx.doi.org/10.1007/s10601-012- 9137-8
10. Teppan, E.C., Friedrich, G., Falkner, A.A.: Quickpup: A heuristic backtracking algorithm for
the partner units configuration problem. In: Fromherz, M.P.J., Muñoz- Avila, H. (eds.)
Proceedings of the Twenty-Fourth Conference on Innovative Applications of Artificial
Intelligence, July 22–26, 2012, Toronto, Ontario, Canada. AAAI (2012). http://www.aaai.
org/ocs/index.php/IAAI/IAAI-12/paper/view/4793
 Synthesis of Constraint Solvers

Douglas R. Smith and Stephen J. Westfold

Kestrel Institute, Palo Alto, CA 94034 USA

{smith,westfold}@kestrel.edu

In [2], we present a mathematical framework for specifying and formally designing

high-performance constraint solving algorithms. The framework is based on concepts
from abstract interpretation which generalizes earlier work on a Galois
Connection-based model of Global Search algorithms. The main focus is on how to use
the framework to automate the calculations necessary to construct correct, efficient
problem-specific constraint solvers.
It is common practice in the constraint-solving community to solve a new problem
P by building a reduction to a well-studied problem Q that has a well-engineered
solver. One problem with this approach is that the reduction of P to Q often loses some
key structure which cannot then be exploited by the Q-solver. Our thesis is that a native
solver can always be generated for a constraint problem that outperforms a reduction to
an existing solver.
This talk focuses on three main results from [2]:
1. Algorithm theory – We develop and prove an algorithm theory for constraint solving
with propagation, conflict detection and analysis, backjumping, and learning that is
parametric on the constraint logic.
2. Design Method for Constraint Propagation – We prove that Arc Consistency is a
best-possible constraint propagation mechanism for arbitrary CSPs, and then showed
how to calculate optimal code for propagation. From Arc Consistency formula
schemes we calculate simple Definite Constraints that can be instantiated into the
optimal Definite Constraint Solver scheme [1].
3. Theory of Conflict Analysis – There are several mathematical formalisms for gen-
eralizing conflict analysis to arbitrary logics. We present a general pattern for cal-
culating resolution rules in a given logic, and prove how resolution can be iterated to
soundly infer a new constraint for backjumping and learning purposes.

References
1. Rehof, J., Mogenson, T.: Tractable constraints finite semilattices. Sci. Comput. Program. 35,
191–221 (1999)
2. Smith, D.R., Westfold, S.: Toward Synthesis Constraint Solvers. Tech. rep., Kestrel Institute
(2013). http://www.kestrel.edu/home/people/smith/pub/CW- report.pdf
 Contents

Technical Track

Encoding Linear Constraints with Implication Chains to CNF. . . . . . . . . . . . 3

Ignasi Abío, Valentin Mayer-Eichberger, and Peter J. Stuckey

Anytime Hybrid Best-First Search with Tree Decomposition

for Weighted CSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
David Allouche, Simon de Givry, George Katsirelos, Thomas Schiex,
and Matthias Zytnicki

Improved Constraint Propagation via Lagrangian Decomposition. . . . . . . . . . 30

David Bergman, Andre A. Cire, and Willem-Jan van Hoeve

Strengthening Convex Relaxations with Bound Tightening for Power

Network Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Carleton Coffrin, Hassan L. Hijazi, and Pascal Van Hentenryck

Broken Triangles Revisited. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

Martin C. Cooper, Aymeric Duchein, and Guillaume Escamocher

A Microstructure-Based Family of Tractable Classes for CSPs . . . . . . . . . . . 74

Martin C. Cooper, Philippe Jégou, and Cyril Terrioux

The Unary Resource with Transition Times . . . . . . . . . . . . . . . . . . . . . . . . 89

Cyrille Dejemeppe, Sascha Van Cauwelaert, and Pierre Schaus

A Global Constraint for a Tractable Class of Temporal

Optimization Problems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Alban Derrien, Jean-Guillaume Fages, Thierry Petit,
and Charles Prud’homme

Exploiting GPUs in Solving (Distributed) Constraint Optimization

Problems with Dynamic Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Ferdinando Fioretto, Tiep Le, Enrico Pontelli, William Yeoh,
and Tran Cao Son

Conflict Ordering Search for Scheduling Problems . . . . . . . . . . . . . . . . . . . 140

Steven Gay, Renaud Hartert, Christophe Lecoutre, and Pierre Schaus

Simple and Scalable Time-Table Filtering for the Cumulative Constraint . . . . 149
Steven Gay, Renaud Hartert, and Pierre Schaus
XX Contents

General Bounding Mechanism for Constraint Programs . . . . . . . . . . . . . . . . 158

Minh Hoàng Hà, Claude-Guy Quimper, and Louis-Martin Rousseau

Smallest MUS Extraction with Minimal Hitting Set Dualization . . . . . . . . . . 173

Alexey Ignatiev, Alessandro Previti, Mark Liffiton,
and Joao Marques-Silva

Upper and Lower Bounds on the Time Complexity

of Infinite-Domain CSPs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
Peter Jonsson and Victor Lagerkvist

Generalized Totalizer Encoding for Pseudo-Boolean Constraints . . . . . . . . . . 200

Saurabh Joshi, Ruben Martins, and Vasco Manquinho

Smaller Selection Networks for Cardinality Constraints Encoding . . . . . . . . . 210

Michał Karpiński and Marek Piotrów

PREFIX-PROJECTION Global Constraint for Sequential Pattern Mining . . . . 226

Amina Kemmar, Samir Loudni, Yahia Lebbah, Patrice Boizumault,
and Thierry Charnois

On Tree-Preserving Constraints. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244

Shufeng Kong, Sanjiang Li, Yongming Li, and Zhiguo Long

Modeling and Solving Project Scheduling with Calendars . . . . . . . . . . . . . . 262

Stefan Kreter, Andreas Schutt, and Peter J. Stuckey

Deterministic Estimation of the Expected Makespan of a POS Under

Duration Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
Michele Lombardi, Alessio Bonfietti, and Michela Milano

A Parallel, Backjumping Subgraph Isomorphism Algorithm

Using Supplemental Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
Ciaran McCreesh and Patrick Prosser

Automated Auxiliary Variable Elimination Through On-the-Fly

Propagator Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
Jean-Noël Monette, Pierre Flener, and Justin Pearson

Automatically Improving SAT Encoding of Constraint Problems

Through Common Subexpression Elimination in Savile Row . . . . . . . . . . . . 330
Peter Nightingale, Patrick Spracklen, and Ian Miguel

Exact Sampling for Regular and Markov Constraints

with Belief Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Alexandre Papadopoulos, François Pachet, Pierre Roy,
and Jason Sakellariou
 Contents XXI

Randomness as a Constraint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351

Steven D. Prestwich, Roberto Rossi, and S. Aramagan Tarim

Quasipolynomial Simulation of DNNF by a Non-determinstic

Read-Once Branching Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
Igor Razgon

MiniSearch: A Solver-Independent Meta-Search Language for MiniZinc . . . . 376

Andrea Rendl, Tias Guns, Peter J. Stuckey, and Guido Tack

Two Clause Learning Approaches for Disjunctive Scheduling. . . . . . . . . . . . 393

Mohamed Siala, Christian Artigues, and Emmanuel Hebrard

Bounding an Optimal Search Path with a Game of Cop and Robber

on Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
Frédéric Simard, Michael Morin, Claude-Guy Quimper,
François Laviolette, and Josée Desharnais

Restricted Path Consistency Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419

Kostas Stergiou

Machine Learning of Bayesian Networks Using Constraint Programming. . . . 429

Peter van Beek and Hella-Franziska Hoffmann

Hybridization of Interval CP and Evolutionary Algorithms

for Optimizing Difficult Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
Charlie Vanaret, Jean-Baptiste Gotteland, Nicolas Durand,
and Jean-Marc Alliot

A General Framework for Reordering Agents Asynchronously

in Distributed CSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
Mohamed Wahbi, Younes Mechqrane, Christian Bessiere,
and Kenneth N. Brown

Automatically Generating Streamlined Constraint Models with ESSENCE

and CONJURE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 480
James Wetter, Özgür Akgün, and Ian Miguel

Application Track

Constraint-Based Local Search for Finding Node-Disjoint Bounded-Paths

in Optical Access Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
Alejandro Arbelaez, Deepak Mehta, and Barry O’Sullivan

Open Packing for Facade-Layout Synthesis Under a General

Purpose Solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 508
Andrés Felipe Barco, Jean-Guillaume Fages, Elise Vareilles,
Michel Aldanondo, and Paul Gaborit
XXII Contents

Power Capping in High Performance Computing Systems . . . . . . . . . . . . . . 524

Andrea Borghesi, Francesca Collina, Michele Lombardi,
Michela Milano, and Luca Benini

A Constraint-Based Approach to the Differential Harvest Problem . . . . . . . . 541

Nicolas Briot, Christian Bessiere, and Philippe Vismara

Constrained Minimum Sum of Squares Clustering by Constraint

Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
Thi-Bich-Hanh Dao, Khanh-Chuong Duong, and Christel Vrain

A Constraint Programming Approach for Non-preemptive

Evacuation Scheduling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574
Caroline Even, Andreas Schutt, and Pascal Van Hentenryck

Solving Segment Routing Problems with Hybrid Constraint

Programming Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592
Renaud Hartert, Pierre Schaus, Stefano Vissicchio,
and Olivier Bonaventure

Modeling Universal Instruction Selection . . . . . . . . . . . . . . . . . . . . . . . . . . 609

Gabriel Hjort Blindell, Roberto Castañeda Lozano, Mats Carlsson,
and Christian Schulte

Optimizing the Cloud Service Experience Using Constraint Programming . . . 627

Serdar Kadioglu, Mike Colena, Steven Huberman, and Claire Bagley

Find Your Way Back: Mobility Profile Mining with Constraints . . . . . . . . . . 638
Lars Kotthoff, Mirco Nanni, Riccardo Guidotti, and Barry O’Sullivan

Joint Vehicle and Crew Routing and Scheduling . . . . . . . . . . . . . . . . . . . . . 654

Edward Lam, Pascal Van Hentenryck, and Philip Kilby

Constructing Sailing Match Race Schedules: Round-Robin Pairing Lists . . . . 671

Craig Macdonald, Ciaran McCreesh, Alice Miller, and Patrick Prosser

Design and Evaluation of a Constraint-Based Energy Saving and

Scheduling Recommender System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687
Seán Óg Murphy, Óscar Manzano, and Kenneth N. Brown

Scheduling Running Modes of Satellite Instruments

Using Constraint-Based Local Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704
Cédric Pralet, Solange Lemai-Chenevier, and Jean Jaubert
 Contents XXIII

Abstracts of Papers Fast Tracked to Constraints Journal

Using Finite Transducers for Describing and Synthesising Structural

Time-Series Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 723
Nicolas Beldiceanu, Mats Carlsson, Rémi Douence, and Helmut Simonis

Projection, Consistency, and George Boole. . . . . . . . . . . . . . . . . . . . . . . . . 724

J.N. Hooker

On computing Minimal Independent Support and its applications

to sampling and counting (Extended Abstract) . . . . . . . . . . . . . . . . . . . . . . 725
Alexander Ivrii, Sharad Malik, Kuldeep S. Meel, and Moshe Y. Vardi

General Game Playing with Stochastic CSP . . . . . . . . . . . . . . . . . . . . . . . . 726

Frédéric Koriche, Sylvain Lagrue, Éric Piette, and Sébastien Tabary

Visual Search Tree Profiling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 728

Maxim Shishmarev, Christopher Mears, Guido Tack,
and Maria Garcia de la Banda

Long-Haul Fleet Mix and Routing Optimisation with Constraint

Programming and Large Neighbourhood Search . . . . . . . . . . . . . . . . . . . . . 729
Philip Kilby and Tommaso Urli

Abstracts of Published Journal Track Papers

On the Reification of Global Constraints (Abstract) . . . . . . . . . . . . . . . . . . . 733

Nicolas Beldiceanu, Mats Carlsson, Pierre Flener, and Justin Pearson

MDD Propagation for Sequence Constraints . . . . . . . . . . . . . . . . . . . . . . . . 734

David Bergman, Andre A. Cire, and Willem-Jan van Hoeve

Discrete Optimization with Decision Diagrams . . . . . . . . . . . . . . . . . . . . . . 735

David Bergman, Andre A. Cire, Willem-Jan van Hoeve,
and John Hooker

A Hybrid Approach Combining Local Search and Constraint Programming

for a Large Scale Energy Management Problem . . . . . . . . . . . . . . . . . . . . . 736
Haris Gavranović and Mirsad Buljubašić

Representing and Solving Finite-Domain Constraint Problems

using Systems of Polynomials (Extended Abstract) . . . . . . . . . . . . . . . . . . . 737
Chris Jefferson, Peter Jeavons, Martin J. Green,
and M.R.C. van Dongen

A Quadratic Extended Edge-Finding Filtering Algorithm for Cumulative

Resource Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 738
Roger Kameugne, Laure Pauline Fotso, and Joseph Scott
XXIV Contents

Achieving Domain Consistency and Counting Solutions for Dispersion

Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 740
Gilles Pesant

meSAT: Multiple Encodings of CSP to SAT . . . . . . . . . . . . . . . . . . . . . . . 741

Mirko Stojadinović and Filip Maric

Constraint programming for LNG ship scheduling and inventory

management (Abstract) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 742
Willem-Jan van Hoeve

Revisiting the Limits of MAP Inference by MWSS on Perfect Graphs . . . . . . 743

Adrian Weller

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745

Technical Track
 Encoding Linear Constraints with Implication
Chains to CNF

Ignasi Abı́o1(B) , Valentin Mayer-Eichberger1,2 , and Peter J. Stuckey1,3

1
NICTA, Canberra, Australia
{ignasi.abio,valentin.mayer-eichberger,peterj.stuckey}@nicta.com.au
2
University of New South Wales, Sydney, Australia
3
University of Melbourne, Melbourne, Australia

Abstract. Linear constraints are the most common constraints occur-

ring in combinatorial problems. For some problems which combine lin-
ear constraints with highly combinatorial constraints, the best solving
method is translation to SAT. Translation of a single linear constraint to
SAT is a well studied problem, particularly for cardinality and pseudo-
Boolean constraints. In this paper we describe how we can improve
encodings of linear constraints by taking into account implication chains
in the problem. The resulting encodings are smaller and can propagate
more strongly than separate encodings. We illustrate benchmarks where
the encoding improves performance.

1 Introduction

In this paper we study linear integer constraints (LI constraints), that is, con-
straints of the form a1 x1 +· · ·+an xn # a0 , where the ai are integer given values,
the xi are finite-domain integer variables, and the relation operator # belongs
to {<, >, ≤, ≥, =, =}. We will assume w.l.o.g that # is ≤, the ai are positive and
all the domains of the variables are {0, 1..di }, since other cases can be reduced to
this one.1 Special case of linear constraints are: pseudo-Boolean (PB) constraints
where the domain of each variable is {0..1}, cardinality (CARD) constraints
where additionally ai = 1, 1 ≤ i ≤ n, and at-most-one (AMO) constraints where
additionally a0 = 1.
Linear integer constraints appear in many combinatorial problems such as
scheduling, planning or software verification, and, therefore, many different SMT
solvers [9,14] and encodings [4,6,11] have been suggested for handling them.
There are two main approaches to encoding linear constraints: cardinality con-
straints are encoded as some variation of a sorting network [3]; multi-decision
diagrams (MDDs) are used to encode more general linear constraints [5], which
in the special case of pseudo-Boolean constraints collapse to binary decision
diagrams (BDDs).

1
See [5] for details. Note that propagation is severely hampered by replacing equalities
by a pair of inequalities.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 3–11, 2015.
DOI: 10.1007/978-3-319-23219-5 1
4 I. Abı́o et al.

Any form of encoding linear constraints to SAT introduces many intermediate

Boolean variables, and breaks the constraint up into many parts. This gives us
the opportunity to improve the encoding if we can recognize other constraints in
the problem that help tighten the encoding of some part of the linear constraint.
Example 1. Consider a pseudo-Boolean constraint x1 + 2x2 + 2x3 + 4x4 + 5x5 +
6x6 + 8x7 ≤ 14 where we also have that x2 + x3 + x5 ≤ 1 we can rewrite the
constraints as x1 +2x235 +4x4 +3x5 +6x6 +8x7 ≤ 14 and x235 ≡ (x2 +x3 +x5 = 1),
where x235 is a new Boolean variable. Notice, that x235 can be used to encode
the at-most-one constraint. 
Example 2. Consider a pseudo-Boolean constraint 4x1 + 2x2 + 5x3 + 4x4 ≤ 9
and the implications x1 ← x2 (i.e. x1 ∨ ¬x2 ) and x2 ← x3 . Separately they do
not propagate, but considered together we can immediately propagate ¬x3 . 
In this paper we show how to encode pseudo-Boolean constraints taking into
account implication chains, as seen in Example 2. The resulting encodings are
no larger than the separate encoding, but result in strictly stronger propagation.
The approach also allows us to encode general linear integer constraints, and
is a strict generalization of the MDD encoding of linear integer constraints [5].
We show how these new combined encodings are effective in practice on a set of
hard real-life sports scheduling examples, and that the combination of pseudo-
Booleans with implication chains arises in a wide variety of models.

2 Preliminaries
2.1 SAT Solving
Let X = {x1 , x2 , . . .} be a fixed set of propositional variables. If x ∈ X then x
and ¬x are positive and negative literals, respectively. The negation of a literal
l, written ¬l, denotes ¬x if l is x, and x if l is ¬x. A clause is a disjunction of
literals l1 ∨ · · · ∨ ln . An implication x1 → x2 is notation for the clause ¬x1 ∨ x2 ,
similarly x1 ← x2 denotes x1 ∨ ¬x2 . A CNF formula is a conjunction of clauses.
A (partial) assignment A is a set of literals such that {x, ¬x} ⊆ A for any
x ∈ X , i.e., no contradictory literals appear. A literal l is true () in A if l ∈ A, is
false (⊥) in A if ¬l ∈ A, and is undefined in A otherwise. A clause C is true in A
if at least one of its literals is true in A. A formula F is true in A if all its clauses
are true in A. In that case, A is a model of F . Systems that decide whether a
formula F has any model are called SAT-solvers, and the main inference rule
they implement is unit propagation: given a CNF F and an assignment A, find
a clause in F such that all its literals are false in A except one, say l, which
is undefined, add l to A and repeat the process until reaching a fix-point. A
detailed explanation can be found in [7].
Let [l..u] where l and u are integers represent the set {l, . . . , u}. Let y be
an integer variable with domain [0..d]. The order encoding introduces Boolean
variables yi for 1 ≤ i ≤ d. A variable yi is true iff y < i. The encoding also
introduces the clauses yi → yi+1 for 1 ≤ i < d.
 Encoding Linear Constraints with Implication Chains to CNF 5

2.2 Multi Decision Diagrams

A directed acyclic graph is called an ordered Multi Decision Diagram (MDD) if

it satisfies the following properties:

– It has two terminal nodes, namely T (true) and F (false).

– Each non-terminal node is labeled by an array of Booleans [xi1 , . . . , xidi ]
representing the order encoding of integer variable yi where yi ranges from
[0..di ]. The variable yi is called the selector variable.
– Every node labeled by yi has di + 1 outgoing edges, labelled xi1 , ¬xi1 , . . . ,
¬xidi .
– Each edge goes from a node with selector yi to a node with selector variable
yj has i < j.

The MDD is quasi-reduced if no isomorphic subgraphs exist. It is reduced if,

moreover, no nodes with only one child exist. A long edge is an edge connecting
two nodes with selector variables yi and yj such that j > i + 1. In the following
we only consider quasi-reduced ordered MDDs without long edges, and we just
refer to them as MDDs for simplicity. A Binary Decision Diagram (BDD) is an
MDD where ∀i, di = 1.
An MDD represents a function f : {0, 1, . . . , d1 } × {0, 1, . . . , d2 } × · · · ×
{0, 1, . . . , dn } → {⊥, } in the obvious way. Moreover, given a fixed variable
ordering, there is only one MDD representing that function. We refer to [17] for
further details about MDDs.
A function f is anti-monotonic in argument yi if vi ≥ vi implies that
f (v1 , . . . , vi−1 , vi , vi+1 , . . . , vn ) =  ⇒ f (v1 , . . . , vi−1 , vi , vi+1 , . . . , vn ) =  for
all values v1 , . . . , vn , vi . An MDD is anti-monotonic if it encodes a function f
that is anti-monotonic in all arguments. We shall only consider anti-monotonic
MDDs in this paper.
Given an anti-monotonic MDD M, we can encode it into CNF by intro-
ducing a new Boolean variable bo to represent each node o in the MDD M;
unary clauses {bT , ¬bF , br } where r is the root node of the MDD; and clauses
{¬bo ∨ bo0 } ∪ {¬bo ∨ xij ∨ boj | j ∈ [1..di ]} for each node o of the form
mdd([xi1 , . . . , xidi ], [o0 , o1 , . . . , odi ]). See [5] for more details.
We can encode arbitrary MDDs to SAT using Tseitin transformation but the
encoding is substantially more complicated.

3 Pseudo Boolean Constraints and Chains

A chain x1 ⇐ x2 ⇐ · · · ⇐ xn is a constraint requiring x1 ← x2 , x2 ← x3 , . . . ,

xn−1 ← xn . A unary chain x1 is the trivial case that imposes no constraint. A
chain is compatible with an ordered list of Boolean variables x1 , . . . , xn if the
chain is of the form xl ⇐ xl+1 ⇐ · · · ⇐ xk , l ≤ k. Given an ordered list L
of Boolean variables x1 , . . . , xn a chain coverage S is a set of variable-disjoint
compatible chains such that each variable appears in exactly one chain. We
will sometimes treat a chain coverage S as a Boolean formula equivalent to the
6 I. Abı́o et al.

constraints of the chains appearing in S. Given a variable ordering and a set of

disjoint compatible chains we can always construct a chain coverage by adding
in unary chains.

Example 3. Given list L of variables x1 , . . . , x9 and chains x1 ⇐ x2 ⇐ x3 , x5 ⇐

x6 , x7 ⇐ x8 , then a chain coverage S of L is {x1 ⇐ x2 ⇐ x3 , x4 , x5 ⇐ x6 , x7 ⇐
x8 , x9 }. S represents the constraint x1 ← x2 ∧ x2 ← x3 ∧ x5 ← x6 ∧ x7 ← x8 . 

Given a Boolean variable ordering L, a PB constraint C and chain coverage

S of L we will demonstrate how to build an MDD to encode the constraint C
taking into account the chain constraints in S. First lets examine how chains
arise in models.

At-most-one constraints. Given PB C ≡ a1 x1 + . . . + an xn ≤ a0 and AMO

A ≡ x1 + x2 + . . . + xk ≤ 1. We can reformulate A using new variables xj
where x1 + . . . + xk = x1 + . . . + xk using the ladder encoding [12] which gives
xi → xi for i = 1 . . . k, xi → ¬xi+1 and xi+1 → xi for i = 1 . . . k − 1. If
[a1 , . . . , ak ] is the sorted array of coefficients [a1 , . . . , ak ] then C can be written
as a1 x1 + (a2 − a1 )x2 + · · · + (ak − ak−1 )xk + ak+1 xk+1 + · · · an xn ≤ a0 . The
chain comes from the auxiliary variables: x1 ⇐ x2 ⇐ · · · ⇐ xk .

General linear integer constraints. A general LI C ≡ a1 y1 + · · · am ym ≤ a0 can

be expressed as a PB with chains. We encode each integer yi with domain [0..di ]
by the order encoding [xi1 , . . . , xidi ] and then C can be rewritten as a1 (¬x11 ) +
· · · + a1 (¬x1d1 ) + · · · + am (¬xm1 ) + · · · + am (¬xmdm ) ≤ a0 with chains ¬xi1 ⇐
¬xi2 ⇐ · · · ⇐ ¬xidi .

Shared coefficients. Frequently, PB constraints contain a large number of coeffi-

cients that are the same. This structure can be exploited. A similar technique of
grouping shared coefficients is described in [5] which in our context is restated
using chains. Given a PB C ≡ ax1 + · · · + axk + ak+1 xk+1 + · · · + an xn ≤ a0
where the first k variables share the same coefficient. We introduce new vari-
ables x1 , . . . , xk to encode the sum x1 + · · · + xk so x1 + . . . + xk = x1 + . . . + xk
and encode this constraint (usually using some form of sorting network [3]).
This ensures that x1 ⇐ x2 ⇐ · · · ⇐ xk . Then C can be rewritten as
ax1 + · · · + axk + ak+1 xk+1 + · · · + an xn ≤ a0 . There are several advantages
with this rewritten version. The sorting network can be represented more com-
pactly than the same logic in an MDD (O(k · log2 k) vs O(k 2 )). Secondly, the
introduced variables xj are meaningful for the constraint and are likely to be
useful for branching and in conflict clause learning during the search. Moreover,
the sorted variables may be reusable for rewriting other constraints.

Binary implicants. Finally, a more general method is to automatically extract

chains from the global problem description. There are a number of methods
to detect binary implicants of CNF encodings [10,13]. Given a set of binary
implicants B and a PB constraint C we can search for a chain coverage S implied
 Encoding Linear Constraints with Implication Chains to CNF 7

by B, and an ordering L of the variables in C with which S is compatible, and

then encode the reordered constraint C making use of the chain coverage S.
In the experimental section of this paper we have only considered the first
three types of chains.

4 Translating Through MDDs with Chains

The main algorithm in this section generalizes the construction of an MDD in

[5]. We first restate definitions of the original algorithm and then show how
to take advantage of chains in the new construction. The CNF decomposition
has desirable properties, i.e. we show that the encoding is more compact and
propagates stronger.

4.1 Preliminaries for the Construction

Let M be the MDD of pseudo-Boolean C and let ν be a node of M with selector

variable xi . We define the interval of ν as the set of values α such that the MDD
rooted at ν represents the pseudo-Boolean constraint ai xi + · · · + an xn ≤ α. It is
easy to see that this definition corresponds in fact to an interval. The key point
in constructing the MDD is to label each node of the MDD with its interval
[β, γ].
In the following, for every i ∈ {1, 2, . . . , n + 1}, we use a set Li consisting of
pairs ([β, γ], M), where M is the MDD of the constraint ai xi + · · · + an xn ≤ a0
for every a0 ∈ [β, γ] (i.e., [β, γ] is the interval of M). All these sets are kept in
a tuple L = (L1 , L2 , . . . , Ln+1 ).
Note that by definition of the MDD’s intervals, if both ([β1 , γ1 ], M1 ) and
([β2 , γ2 ], M2 ) belong to Li then either [β1 , γ1 ] = [β2 , γ2 ] or [β1 , γ1 ] ∩ [β2 , γ2 ] = ∅.
Moreover, the first case holds if and only if M1 = M2 . Therefore, Li can be
represented with a binary search tree-like data structure, where insertions and
searches can be done in logarithmic time. The function search(K, Li ) searches
whether there exists a pair ([β, γ], M) ∈ Li with K ∈ [β, γ]. Such a tuple is
returned if it exists, otherwise an empty interval is returned in the first com-
ponent of the pair. Similarly, we also use function insert(([β, γ], M), Li ) for
insertions.

4.2 Algorithm and Properties of the Construction

In this section we show how to translate a PB C ≡ a1 x1 + . . . an xn ≤ a0

and a chain coverage S for variable order x1 , . . . , xn . Algorithm 1 describes
the construction of the MDD. The initial call is MDDChain(1, C, S). The call
MDDChain(i, C  , S) recursively builds an MDD for C  ∧ S by building the ith
level. If the chain including xi is xi ⇐ · · · ⇐ xk it builds an MDD node that
has child nodes with selector xk+1 . If the chain for xi is unary this is the usual
MDD (BDD) construction.
8 I. Abı́o et al.

Algorithm 1. Procedure MDDChain

Require: i ∈ {1, 2, . . . , n + 1} and pseudo-Boolean constraint C  : ai xi + . . . + an xn ≤
a0 and chain coverage S on [x1 , . . . , xn ]
Ensure: returns [β, γ] interval of C  and M its MDD
1: ([β, γ], M) ← search(a0 , Li ).
2: if [β, γ] = ∅ then
3: return ([β, γ], M).
4: else
5: δ0 ← 0
6: let {xi ⇐ xi+1 ⇐ · · · ⇐ xk } ∈ S % including unary chain xi
7: u←k−i+1
8: for all j such that 0 ≤ j ≤ u do
9: ([βj , γj ], Mj ) ← MDDChain(k + 1, ak+1 xk+1 + · · · + an xn ≤ a0 − δj , S).
10: δj+1 ← δj + ai+j
11: end for
12: M ← mdd([xi , . . . , xk ], M0 , . . . , Mu )
13: [β, γ] ← [β0 , γ0 ] ∩ [β1 + δ1 , γ1 + δ1 ] ∩ · · · ∩ [βu + δu , γu + δu ].
14: insert(([β, γ], M), Li ).
15: return ([β, γ], M).
16: end if

Example 4. The MDDs that result from MDDChain(1, C, S) where C ≡

4x1 + 2x2 + 5x3 + 4x4 ≤ 9 of Example 2 encoded with chain coverage (a)
S = {x1 , x2 , x3 , x4 } (no chains) and (b) S = {x1 ⇐ x2 ⇐ x3 , x4 } are shown
in Figure 1. The diagrams show [β, γ] for each node with the remainder of the
constraint at the left. Unit propagation of the CNF of (b) sets x3 = ⊥ immedi-
ately since 4x4 ≤ −1 is ⊥.

We can prove that the algorithm returns a correct MDD, that is no larger
than the MDD (BDD) encoding of C, and that the resulting CNF encoding is
domain consistent on the original variables x1 , . . . , xn . Proofs are available at [1].

Theorem 1. Given a pseudo-Boolean constraint C ≡ a1 x1 +· · ·+an xn ≤ a0 and

chain coverage S on [x1 , . . . , xn ] then MDDChain(1, C, S) returns an MDD M
representing function f such that constraint C ∧ S |= f . The running time of the
algorithm is O(n · a0 · log a0 ). 

Theorem 2. Given a pseudo-Boolean constraint C ≡ a1 x1 + · · · + an xn ≤ a0

and chain coverage S on [x1 , . . . , xn ] then the MDD MDDChain(1, C, S) has
no more nodes than MDDChain(1, C, {x1 , . . . , xn }), the BDD for C. 

Theorem 3. Given a pseudo-Boolean constraint C ≡ a1 x1 + · · · + an xn ≤ a0

and chain coverage S on [x1 , . . . , xn ] then unit propagation on the CNF encoding
of MDDChain(1, C, S) ∧ S enforces domain consistency of C ∧ S on variables
x1 , . . . , xn . 
 Encoding Linear Constraints with Implication Chains to CNF 9

4x1 + 2x2 + 5x3 + 4x4 ≤ [9, 9] [9, 9]

EE ¬x1  ,,
x1
ttt EE
y tt
t E"   ,,
,,
2x2 + 5x3 + 4x4 ≤ [5, 5]
JJJ¬x2
[9, 9]
 ,,
x2 yy x3  x
JJJ y  ,,x1 ,¬x1
|yy
x2 ¬x2 2
 %  ,,
5x3 + 4x4 ≤ [0, 3] [5, 8] [9, 9]   ,,
JJJ¬x3 EE ¬x3 
JJJx3 EE ¬x3 ,x3  ,,
x3
 %  E"   
4x4 ≤ [−∞, −1] [0, 3] [4, 9] [−∞, −1] [0, 3] [4, 9]
x4 ss FFF¬x4 ¬x ,x x4 ss F FF¬x4 ¬x ,x
s FF  4 4 s FF  4 4
 yssss  yssss
¬x4 ,x4 ¬x4 ,x4
# #
F T F T

(a) (b)

Fig. 1. The MDDs that result from 4x1 + 2x2 + 5x3 + 4x4 ≤ 9 encoded (a) without
and (b) with the chain x1 ⇐ x2 ⇐ x3 .

5 Experiments

To illustrate the advantage of combined compilation we consider a challenging

combinatorial optimization problem where both AMO and shared coefficients
chains arise.
Sports league scheduling is a challenging combinatorial optimization problem.
We consider scheduling a double round-robin sports league of N teams. All teams
meet each other once in the first N − 1 weeks and again in the second N − 1
weeks, with exactly one match per team each week. A given pair of teams must
play at the home of one team in one half, and at the home of the other in the
other half, and such matches must be spaced at least a certain minimal number
of weeks apart. Additional constraints include, e.g., that no team ever plays at
home (or away) three times in a row, other (public order, sportive, TV revenues)
constraints, blocking given matches on given days, etc.
Additionally, the different teams can propose a set of constraints with some
importance (low, medium or high). We aim not only to maximize the number of
these constraints satisfied, but also to assure that at least some of the constraints
of every team are satisfied. More information can be found in [2].
Low-importance constraints are given a weight of 1; medium-importance,
5, and high-importance, 10. For every constraint proposed by a team i, a new
Boolean variable xi,j is created. This variable is set to true  if the constraint
is violated. For every team, a pseudo-Booleanconstraint
 j wi,j xi,j ≤ Ki is
imposed. The objective function to minimize is i j wi,j xi,j . The data is based
on real-life instances.
Desired constraints typically refer to critical weeks in the schedule, e.g.
around Christmas, or other key dates, and preferences of different teams almost
always clash. Double round-robin tournaments contain a lot of AMO and EO
constraints (for instance, each week each team meets exactly one team). These
AMO constraints can be used to simplify the desired constraints.
10 I. Abı́o et al.

Table 1. Results for sports league scheduling, showing the number of runs that find a
solution of different quality after different time limits (seconds).

Quality Some solution cost ≤ 30 + best cost ≤ 20 + best cost ≤ 10 + best

Timelimit 300 900 3600 300 900 3600 300 900 3600 300 900 3600
MDD1 148 190 199 21 55 107 17 35 74 6 25 51
MDD2 151 194 199 27 59 115 19 38 81 12 25 43
MDD3 160 191 200 56 107 162 45 72 121 41 52 87
LCG 69 123 172 21 29 51 18 21 35 14 20 27
Gurobi 0 0 0 0 0 0 0 0 0 0 0 0

The benchmark consists of 10 instances and each method is run 20 times

with different seeds, for 200 total runs. Compared methods are: MDD1, the usual
MDD (in fact, BDD) method to encode PBs [4]; MDD2, the method of [5] using
sorting networks for the identical coefficients and then using an MDD; MDD3,
the method defined herein; LCG, using lazy clause generation [15]; and Gurobi,
using the MIP solver Gurobi. Barcelogic [8] SAT Solver was used in methods
MDD1, MDD2, MDD3 and LCG.
The results can be shown at Table 1. The number of times a solution has
been found within the time limit can be found at columns 2-4. Columns 5-7
present the number of times (within the timelimit) a method finds a solution
of cost at most best + 30, where best is the cost of the best solution found by
any method. Similarly, columns 8-10 and 11-13 contain the number of times a
solution of cost at most best + 20 and best + 10 has been found.
As we can see the new encoding substantially improves on previous encodings
of the problem. For these sports leagues scheduling problems it is well known
that other solving approaches do not compete with SAT encoding [2].

6 Conclusion and Future Work

We demonstrate a new method for encoding pseudo-Boolean constraints taking
into account implications chains. The improved encoding is beneficial on hard
benchmark problems. The approach is an extension on earlier work on encoding
linear constraints to SAT [5]. The approach is related to the propagator for
the increasing sum constraint y = a1 y1 + · · · + an yn ∧ y1 ≤ y2 ≤ · · · ≤ yn
described in [16], which combines a linear constraint with a “chain” of integer
inequalities. Interestingly, increasing sum is not directly encodable as an MDD
using the method herein, but it does suggest that the methods can be extended
to arbitrary sets of chains all compatible with a global variable order. Another
interesting direction for future work is to consider combining chains with the
sorting networks encodings of linear constraints (e.g. [11]).

Acknowledgments. NICTA is funded by the Australian Government as represented

by the Department of Broadband, Communications and the Digital Economy and the
Australian Research Council through the ICT Centre of Excellence program.
 Encoding Linear Constraints with Implication Chains to CNF 11

References
1. Abio, I., Mayer-Eichberge, V., Stuckey, P.: Encoding linear constraints with impli-
cation chains to CNF. Tech. rep., University of Melbourne (2015). http://www.
people.eng.unimelb.edu.au/pstuckey/papers/cp2015a.pdf
2. Abı́o, I.: Solving hard industrial combinatorial problems with SAT. Ph.D. thesis,
Technical University of Catalonia (UPC) (2013)
3. Abı́o, I., Nieuwenhuis, R., Oliveras, A., Rodrı́guez-Carbonell, E.: A parametric
approach for smaller and better encodings of cardinality constraints. In: Schulte,
C. (ed.) CP 2013. LNCS, vol. 8124, pp. 80–96. Springer, Heidelberg (2013)
4. Abı́o, I., Nieuwenhuis, R., Oliveras, A., Rodrı́guez-Carbonell, E., Mayer-
Eichberger, V.: A New Look at BDDs for Pseudo-Boolean Constraints. J. Artif.
Intell. Res. (JAIR) 45, 443–480 (2012)
5. Abı́o, I., Stuckey, P.J.: Encoding linear constraints into SAT. In: O’Sullivan, B.
(ed.) CP 2014. LNCS, vol. 8656, pp. 75–91. Springer, Heidelberg (2014)
6. Bailleux, O., Boufkhad, Y., Roussel, O.: New encodings of pseudo-boolean con-
straints into CNF. In: Kullmann, O. (ed.) SAT 2009. LNCS, vol. 5584, pp. 181–194.
Springer, Heidelberg (2009)
7. Biere, A., Heule, M., van Maaren, H., Walsh, T. (eds.): Handbook of Satisfiability,
Frontiers in Artificial Intelligence and Applications, vol. 185. IOS Press (2009)
8. Bofill, M., Nieuwenhuis, R., Oliveras, A., Rodrı́guez-Carbonell, E., Rubio, A.: The
barcelogic SMT solver. In: Gupta, A., Malik, S. (eds.) CAV 2008. LNCS, vol. 5123,
pp. 294–298. Springer, Heidelberg (2008)
9. Dutertre, B., de Moura, L.: The YICES SMT Solver. Tech. rep., Computer Science
Laboratory, SRI International (2006). http://yices.csl.sri.com
10. Eén, N., Biere, A.: Effective preprocessing in SAT through variable and clause elim-
ination. In: Bacchus, F., Walsh, T. (eds.) SAT 2005. LNCS, vol. 3569, pp. 61–75.
Springer, Heidelberg (2005)
11. Eén, N., Sörensson, N.: Translating Pseudo-Boolean Constraints into SAT. Journal
on Satisfiability, Boolean Modeling and Computation 2(1–4), 1–26 (2006)
12. Gent, I.P., Prosser, P., Smith, B.M.: A 0/1 encoding of the GACLex constraint for
pairs of vectors. In: ECAI 2002 workshop W9: Modelling and Solving Problems
with Constraints. University of Glasgow (2002)
13. Heule, M.J.H., Järvisalo, M., Biere, A.: Efficient CNF simplification based on
binary implication graphs. In: Sakallah, K.A., Simon, L. (eds.) SAT 2011. LNCS,
vol. 6695, pp. 201–215. Springer, Heidelberg (2011)
14. de Moura, L., Bjorner, N.: Z3: An Efficient SMT Solver. Tech. rep., Microsoft
Research, Redmond (2007). http://research.microsoft.com/projects/z3
15. Ohrimenko, O., Stuckey, P., Codish, M.: Propagation via lazy clause generation.
Constraints 14(3), 357–391 (2009)
16. Petit, T., Régin, J.-C., Beldiceanu, N.: A Θ(n) bound-consistency algorithm for
the increasing sum constraint. In: Lee, Jimmy (ed.) CP 2011. LNCS, vol. 6876, pp.
721–728. Springer, Heidelberg (2011)
17. Srinivasan, A., Ham, T., Malik, S., Brayton, R.: Algorithms for discrete function
manipulation. In: 1990 IEEE International Conference on Computer-Aided Design,
ICCAD 1990, pp. 92–95. Digest of Technical Papers (1990)
 Anytime Hybrid Best-First Search with Tree
Decomposition for Weighted CSP

David Allouche, Simon de Givry(B) , George Katsirelos,

Thomas Schiex, and Matthias Zytnicki

MIAT, UR-875, INRA, F-31320 Castanet Tolosan, France

{david.allouche,simon.degivry,george.katsirelos,
thomas.schiex,matthias.zytnicki}@toulouse.inra.fr

Abstract. We propose Hybrid Best-First Search (HBFS), a search strat-

egy for optimization problems that combines Best-First Search (BFS)
and Depth-First Search (DFS). Like BFS, HBFS provides an anytime
global lower bound on the optimum, while also providing anytime upper
bounds, like DFS. Hence, it provides feedback on the progress of search
and solution quality in the form of an optimality gap. In addition, it
exhibits highly dynamic behavior that allows it to perform on par with
methods like limited discrepancy search and frequent restarting in terms
of quickly finding good solutions.
We also use the lower bounds reported by HBFS in problems with
small treewidth, by integrating it into Backtracking with Tree Decompo-
sition (BTD). BTD-HBFS exploits the lower bounds reported by HBFS in
individual clusters to improve the anytime behavior and global pruning
lower bound of BTD.
In an extensive empirical evaluation on optimization problems from
a variety of application domains, we show that both HBFS and BTD-
HBFS improve both anytime and overall performance compared to their
counterparts.

Keywords: Combinatorial optimization · Anytime algorithm ·

Weighted constraint satisfaction problem · Cost function networks ·
Best-first search · Tree decomposition

1 Introduction

Branch and Bound search is a fundamental tool in exact combinatorial optimiza-

tion. For minimization, in order to prune the search tree, all variants of Branch
and Bound rely on a local lower bound on the cost of the best solution below a
given node.
Depth-First Search (DFS) always develops a deepest unexplored node. When
the gap between the local lower bound and a global upper bound on the cost of an
optimal solution – usually provided by the best known solution – becomes empty,
backtrack occurs. DFS is often used in Constraint Programming because it offers


c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 12–29, 2015.
DOI: 10.1007/978-3-319-23219-5 2
 Anytime Hybrid Best-First Search with Tree Decomposition 13

polyspace complexity, it takes advantage of the incrementality of local consisten-

cies and it has a reasonably good anytime behavior that can be further enhanced
by branching heuristics. This anytime behavior is however largely destroyed in
DFS variants targeted at solving problems with a reasonable treewidth such as
BTD [7] or AND/OR search [6].
Best-First Search (BFS) instead always develops the node with the lowest
lower bound first. It offers a running global lower bound and has been proved to
never develop more nodes than DFS for the same lower bound [22]. But it has a
worst-case exponential space complexity and the optimal solution is always the
only solution produced.
An ideal Branch and Bound algorithm would combine the best of all
approaches. It would have a bearable space complexity, benefit from the incre-
mentality of local consistencies and offer both updated global upper and lower
bounds as the problem is solved. It would also not loose all its anytime qualities
when used in the context of treewidth sensitive algorithms such as BTD.
With updated global lower and upper bounds, it becomes possible to com-
pute a current global optimality gap. This gap can serve as a meaningful indi-
cator of search progress, providing a direct feedback in terms of the criteria
being optimized. This gap also becomes of prime importance in the context of
tree-decomposition based Branch and Bound algorithms such as BTD [7] as
global bounds for each cluster can typically be used to enhance pruning in other
clusters.
In this paper, we introduce HBFS, an hybrid, easy to implement, anyspace
Branch and Bound algorithm combining the qualities of DFS and BFS. The only
limitation of HBFS is that it may require to compromise the anytime updating
of the global lower bound for space. This can be achieved dynamically dur-
ing search. HBFS can also be combined with a tree-decomposition to define the
more complex BTD-HBFS, a BTD variant offering anytime solutions and updated
global optimality gap.
On a set of more than 3,000 benchmark problems from various sources
(MaxCSP, WCSP, Markov Random Fields, Partial Weighted MaxSAT) includ-
ing resource allocation, bioinformatics, image processing and uncertain reasoning
problems, we observe that HBFS improves DFS in term of efficiency, while being
able to quickly provide good solutions – on par with LDS and Luby restarts –
and a global running optimality gap. Similarly, HBFS is able to improve the
efficiency and anytime capacities of BTD.

2 Background
Our presentation is restricted to binary problems for simplicity. Our implemen-
tation does not have such restriction. A binary Cost Function Network (CFN) is
a triplet (X, D, W ). X = {1, . . . , n} is a set of n variables. Each variable i ∈ X
has a finite domain Di ∈ D of values than can be assigned to it. The maximum
domain size is d. W is a set of cost functions. A binary cost function wij ∈ W is
a function wij : Di × Dj → [0, k] where k is a given maximum integer cost cor-
responding to a completely forbidden assignment (expressing hard constraints).
14 D. Allouche et al.

If they do not exist, we add to W one unary cost function for every variable such
that wi : Di → [0, k] and a zero arity constraint w∅ (a constant cost payed by
any assignment, defining a lower bound on the optimum). All these additional
cost functions will have initial value 0, leaving the semantics of the problem
unchanged.
The Weighted Constraint Satisfaction Problem (WCSP)  is to find a
 n
minimum
 cost complete assignment: min (a1 ,...,an )∈ i Di {w∅ + i=1 wi (ai ) +
w (a
wij ∈W ij i j , a )}, an optimization problem with an associated NP-complete
decision problem.
The WCSP can be solved exactly using Branch and Bound maintaining some
lower bound: at each node ν of a tree, we use the local non naive lower bound
ν.lb = w∅ provided by a given soft arc consistency [5]. Each node corresponds to
a sequence of decisions ν.δ. The root node has an empty decision sequence. When
a node is explored, an unassigned variable is chosen and a branching decision
to either assign the variable to a chosen value (left branch, positive decision)
or remove the value from the domain (right branch, negative decision) is taken.
The number of decisions taken to reach a given node ν is the depth of the node,
ν.depth. A node of the search tree that corresponds to a complete assignment is
called a leaf. At this point, ν.lb is assumed to be equal to the node cost (which
is guaranteed by all soft arc consistencies).
The graph G = (X, E) of a CFN has
one vertex for each variable and one edge
(i, j) for every binary cost function wij ∈
W . A tree decomposition of this graph is
defined by a tree (C, T ). The set of nodes
of the tree is C = {C1 , . . . , Cm } where Ce
is a set of variables (Ce ⊆ X) called a clus-
ter. T is a set of edges connecting clusters
and forming a tree (a connected acyclic
graph). The set ofclusters C must cover
all the variables ( Ce ∈C Ce = X) and all
the cost functions (∀{i, j} ∈ E, ∃Ce ∈
C s.t. i, j ∈ Ce ). Furthermore, if a vari-
able i appears in two clusters Ce and Cg ,
i must also appear in all the clusters Cf on
the unique path from Ce to Cg in (C, T ). Fig. 1. A tree-decomposition of the
If the cardinality of the largest cluster in a CELAR06 radio frequency assignment
tree decomposition is ω + 1 then the width problem, rooted in C1 with subproblem
of the decomposition is ω. The treewidth P5 highlighted.
of a graph is the minimum width among
all its decompositions [24].

3 Hybrid Best-First Search

Classical BFS explores the search tree by keeping a list open of open nodes
representing unexplored subproblems. Initially, this list is reduced to the root
 Anytime Hybrid Best-First Search with Tree Decomposition 15

node at depth 0. Iteratively, a best node is explored: the node is removed and
replaced by its two left and right children with updated decisions, lower bound
and depth. In this paper we always choose as best node a node with the smallest
ν.lb, breaking ties by selecting a node with maximum ν.depth. The first leaf of the
tree explored is then guaranteed to be an optimal solution [14,22]. The list open
may reach a size in O(dn ) and, if incrementality in the lower bound computation
is sought, each node should hold the minimum data-structures required for soft
arc consistency enforcing (in O(ed) per node).
The pseudocode for Hybrid BFS is described as Algorithm 1. HBFS starts
with the empty root node in the list of open nodes. It then iteratively picks a
best node ν from the open list as above, replays all the decisions in ν.δ leading to
an assignment Aν , while maintaining consistency. It then performs a depth-first
search probe starting from that node for a limited number Z of backtracks. The
DFS algorithm is a standard DFS algorithm except for the fact that, when the
bound on the number of backtracks is reached, it places all the nodes corre-
sponding to open right branches of its current search state in the open list (see
Figure 2).
At the price of increased memory usage, this
1
hybrid maintains the advantages of depth-first
search. Since it spends a significant amount of its
time in a DFS subroutine, it can exploit the incre- 2
mentality of arc consistency filtering during DFS
search without any extra space cost: nodes in the 3 6
open list will just contain decisions δ and lower
bound lb, avoiding the extra O(ed) space required 4 5 7
for incrementality during BFS. However, each time
a node is picked up in open, the set of ν.depth 8
decisions must be “replayed” and local consistency
reinforced from the root node state, leading to Fig. 2. A tree that is par-
redundant propagation. This cost can be mitigated tially explored by DFS with
to some degree by merging all decisions into one. backtrack limit = 3. Nodes
Hence, a single fixpoint has to be computed rather with a bold border are
than ν.depth. Additionally, the cost can be fur- leaves, nodes with no border
ther reduced using other techniques employed by are placed in the open list
copying solvers [26]. Regardless of these mitigation after the backtrack bound is
techniques, some redundancy is unavoidable, hence exceeded. Nodes are num-
the number of backtracks performed at each DFS bered in the order they are
visited.
probe should be large enough to avoid excessive
redundancy.
Second, as it is allowed to perform Z backtracks in a depth-first manner
before it picks a new node, it may find new and better incumbent solutions, thus
it is anytime. The number of backtracks of each DFS probe should be sufficiently
small to offer quick diversification: by exploring a new best node, we are offered
the opportunity to reconsider early choices, similarly to what LDS [8] and Luby
16 D. Allouche et al.

randomized restarts [18] may offer. Additionally, with early upper bounds, we
can also prune the open node list and remove all nodes such that ν.lb ≥ ub.
To balance the conflicting objectives of reducing repeated propagation and
diversification, we dynamically adjust the amount of backtracks Z that can be
performed during one DFS probe by trying to keep the observed rate of redun-
dantly propagated decisions between reasonable bounds (α and β). In all the
algorithms here, we assume that the number of nodes (Nodes) and backtracks
(Backtracks) are implicitly maintained during search.

Function HBFS(clb,cub) : pair(integer,integer)

open := ν(δ = ∅, lb = clb) ;
while (open = ∅ and clb < cub) do
ν :=pop(open) /* Choose a node with minimum lower bound and maximum
depth */;
Restore state ν.δ, leading to assignment Aν , maintaining local consistency ;
NodesRecompute := NodesRecompute + ν.depth ;
cub :=DFS(Aν ,cub,Z)/* puts all right open branches in open */ ;
clb := max(clb, lb(open)) ;
if (NodesRecompute > 0) then
if (NodesRecompute/Nodes > β and Z ≤ N ) then Z := 2 × Z;
else if (NodesRecompute/Nodes < α and Z ≥ 2) then Z := Z/2;
return (clb, cub);

Algorithm 1. Hybrid Best-First Search. Initial call: HBFS(w∅ ,k) with Z = 1.

This hybrid does not preserve the polyspace complexity of DFS. However, it
can easily be made anyspace. If memory is exhausted (or a memory upper bound
is reached, with the same effect), the algorithm can switch from bounded DFS
to complete DFS. This means that for every node it picks from the open list, it
explores the entire subtree under that node. Hence, it will not generate any new
open nodes. It can continue in this mode of operation until memory pressure is
relieved.
Finally, this method computes stronger global lower bounds than DFS, as the
cost of a best node in the open list defines a global lower bound, as in BFS. DFS
instead cannot improve on the global lower bound computed at the root until it
finally visits the first right branch. In the context of a single instance this is only
important in the sense that it provides a better estimation of the optimality gap.
However, we will see that this can improve performance in decomposition-based
methods.

3.1 Related Work

Alternate search space exploration schemes have been proposed in the field of
heuristic search, as variations of A* search. These schemes can be applied to dis-
 Anytime Hybrid Best-First Search with Tree Decomposition 17

crete optimization, yielding other variants of best-first search. However, depth-

first search is not effective or even feasible in domains where A* search is used:
for example, it is possible in planning to have exponentially long sequences of
actions when short plans exist. Hence, methods like BRFSL(k) [27] can only do
bounded-depth DFS probes. Also, in contrast to HBFS, they do not insert the
open nodes of the DFS probes into the open list of BFS. Other methods like
Weighted best-first search [23], ARA* [16] and ANA* [2] weigh future assign-
ments more heavily in order to bias the search towards solutions. We do not
need to modify the branching heuristic in any way in HBFS.
Stratification [1], which solves a weighted MaxSAT instance by iteratively
considering larger subsets of its clauses, starting with those that have the high-
est weight, provides similar benefits to HBFS, as provides solutions quickly and
produces lower bounds. This techniques, however, can be viewed as a wrapper
over an optimization method and is therefore orthogonal to HBFS.
Alternate heuristics for choosing the next node to explore may yield different
algorithms. When we can identify a preferred value to assign at each choice point,
the discrepancy of a node ν is the number of right branches in the path from the
root to ν. If we always open the node with the smallest discrepancy, set Z = 1
and disable the adaptive heuristic, HBFS is identical to Limited Discrepancy
Search (LDS)1 [8].
In ILP, a closely related approach is so-called BFS with diving heuristics [3].
Such heuristics perform a single depth-first probe trying to find a feasible solu-
tion. Although the idea is quite close to that of HBFS, it is typically restricted
to a single branch, the open nodes it leaves are not added to the open node
file and is treated separately from the rest of the search process. This is in part
motivated by the fact that DFS is considered impractical in ILP [17] and by the
fact that the lower bounding method (LP) used is not as lightweight as those
used in WCSP.

4 Hybrid Best-First Search and Tree Decompositions

When the graph of a CFN has bounded treewidth, the O(dn ) worst-case com-
plexity of DFS can be improved using a tree decomposition of the CFN graph.
We can trivially observe that the tree decomposition can be rooted by selecting
a root cluster denoted C1 . The separator of a non root cluster Ce is Ce ∩ pa(Ce ),
where pa(Ce ) is the parent of Ce in T . Local consistency can be enforced on
e
the problem and provide a cluster-localized lower-bound w∅ for each cluster Ce .
The sum of these cluster-wise lower bounds is a lower bound for the complete
problem. Beyond this trivial observation, Terrioux and Jégou [28] and de Givry
et al. [7] have extended BTD [9] (which we call BTD-DFS here for clarity) from
pure satisfaction problems to the case of optimization (WCSP), in a way simi-
lar to AND/OR search [19]. Next, we briefly describe BTD-DFS, as given by de
Givry et al, as we base our own algorithm on this.
1
In WCSP optimization, we always have a non naive value heuristic that selects a
value (i, a) with minimum unary marginal cost wi (a) or better, the EAC support [13].
18 D. Allouche et al.

In BTD-DFS, by always assigning the variables of a cluster before the variables

of its descendant clusters, it is possible to exploit the fact that assigning a cluster
Ce separates all its child clusters children(Ce ). Each child cluster Cf is the root
of a subproblem Pf defined by the subtree rooted in Cf which becomes indepen-
dent of others. So, each subproblem Pf conditioned by the current assignment
Af of its separator, can be independently and recursively solved to optimality.
If we memoize the optimum cost of every solved conditioned subproblem Pe |Ae
in a cache, then Pe |Ae will never be solved again and an overall O(ndω+1 ) time
complexity can be guaranteed.
Although this simple strategy offers an attractive worst case theoretical
bound, it may behave poorly in practice. Indeed, each conditioned subprob-
lem Pe |Ae is always solved from scratch to optimality. This ignores additional
information that can be extracted from already solved clusters. Imagine Ce has
been assigned and that we have an upper bound ub (a solution) for the problem
Pe |Ae . Assume that Ce has two children Cf and Cf  and that we have solved the
e
first subproblem Pf |Af to optimality. By subtracting the lower bound w∅ and
the optimum of Pf |Af from ub, we obtain the maximum cost that a solution of
Pf  |Af  may have in order to be able to improve over ub. Instead of solving it
from scratch, we can solve Pf  |Af  with this initial upper bound and either find
an optimal solution – which can be cached – or fail. If we fail, we have proved
a global lower bound on the cost of an optimal solution of Pf  |Af  . This lower
bound can be cached and prevent repeated search if Pf  |Af  is revisited with
the same or a lower initial upper bound. Otherwise, the problem will be solved
again and again either solved to optimality or fail and provide an improved global
lower bound. This has been shown to improve search in practice while offering
a theoretical bound on time complexity in O(kn.dω+1 ) (each time a subproblem
Pf |Af is solved again, the global lower bound increases at least by 1).
In practice, we therefore cache two values, LB Pe |Ae and UB Pe |Ae , for every
visited assignment Ae of the separator of every cluster Ce . We always assume
caching is done implicitly: LB Pe |Ae is updated every time a stronger lower bound
is proved for Pe |Ae and UB Pe |Ae when an updated upper bound is found. When
an optimal solution is found and proved to be optimal, we will therefore have
LB Pe |Ae = UB Pe |Ae . Thanks to these cached bounds and to the cluster-wise local
e
lower bounds w∅ , an improved local lower bound lb(Pe |Ae ) for the subproblem
Pe |Ae can be computed by recursively summing the maximum of the cached and
local bound (see [7]).
We show pseudocode for the resulting algorithm combining BTD and DFS in
Algorithm 2. Grayed lines in this code are not needed for the DFS variant and
should be ignored. The algorithm is called on root cluster C1 , with an assignment
A1 = ∅, a set of unassigned variables V = C1 and initial lower and upper bound
clb and cub set respectively to lb(P1 |∅) and k (the maximum cost). The last
argument, RecCall is a functional argument that denotes which function will be
used to recurse inside BTD-DFS. Here, RecCall will be initially equal to BTD-
DFS itself. The algorithm always returns two identical values equal to the current
 Anytime Hybrid Best-First Search with Tree Decomposition 19

Function BTD-DFS(A,Ce ,V ,clb,cub,RecCall ) : pair(integer,integer)

if (V = ∅) then
i :=pop(V ) /* Choose an unassigned variable in Ce */ ;
a :=pop(Di ) /* Choose a value */ ;
Assign a to i, maintaining local consistency on subproblem lb(Pe |A ∪ {(i = a)}) ;
clb  := max(clb, lb(Pe |A ∪ {(i = a)})) ;
if (clb < cub) then
(cub, cub) := BTD-DFS(A ∪ {(i = a)}, Ce , V − {i}, clb , cub,RecCall );
Ce .backtracks := Ce .backtracks + 1;
if (max(clb, lb(Pe |A)) < cub) then
Remove a from i, maintaining local consistency on subproblem
lb(Pe |A ∪ {(i = a)}) ;
clb  := max(clb, lb(Pe |A ∪ {(i = a)})) ;
if (clb < cub) then
if (Ce .backtracks < Ce .limit and Backtracks < Pe .limit) then
(cub, cub) := BTD-DFS(A ∪ {(i = a)}, Ce , V , clb , cub,RecCall );
else /* Stop depth-first search */
Push current search node in open list of Pe |A at position clb ;
else
S := Children(Ce ) ;
/* Solve all clusters with non-zero optimality gap and unchanged lb or ub */ ;
while (S = ∅ and lb(Pe |A) < cub) do
Cf :=pop(S) /* Choose a child cluster */ ;
if (LB Pf |A < UB Pf |A ) then
cub  := min(UB Pf |A , cub − [lb(Pe |A) − lb(Pf |Af )]) ;
(clb  , cub  ) := RecCall (A, Cf , Cf , lb(Pf |Af ), cub ,RecCall );
Update LB Pf |A and UB Pf |A using clb  and cub  ;
e 
cub := min(cub, w∅ + Cf ∈Children(Ce ) UB Pf |A );
if max(clb, lb(Pe |A)) < cub then
Push current search node in open list of Pe |A at position max(clb, lb(Pe |A)) ;
Ce .limit := Ce .backtracks /* Stop depth-first search */ ;
return (cub, cub)

Algorithm 2. BTD using depth-first search

upper bound.2 Caches are initially empty and return naive values LB Pe /A = 0
and UB Pe /A = k for all clusters and separator assignments.

4.1 Using HBFS in BTD

BTD-DFS has two main disadvantages: first, it has very poor anytime behavior,
as it cannot produce a solution in a decomposition with k leaves until k − 1 leaf
clusters have been completely solved. This affects the strength of pruning, as
2
This is clearly redundant for BTD-DFS, but allows a more uniform presentation with
BTD-HBFS.
20 D. Allouche et al.

values are only pruned if the current lower bound added to the marginal cost of
the value exceeds the upper bound. Second, because child clusters are examined
in order, only the lower bounds of siblings earlier than Cf in that order can
contribute to pruning in Cf . For example, consider a cluster Ce with 3 child
e
clusters Cf1 , Cf2 , Cf3 . Assume that ub = 31 and under an assignment A, w∅ has
known cost 10 while Pf1 |Af1 , Pf2 |Af2 and Pf3 |Af3 all have optimal cost 10, and
lb(Pf1 |Af1 ) = lb(Pf2 |Af2 ) = lb(Pf3 |Af3 ) = 0. Clearly the subproblem under Ce
cannot improve on the upper bound, but when we solve Cf1 and Cf2 , BTD-DFS
does not reduce the effective upper bound at all. However, it may be relatively
easy to prove a lower bound of 7 for each of the child clusters. If we had this
information, we could backtrack.
HBFS has the ability to quickly provide good lower and upper bounds and inter-
rupts itself as soon as the limit number of backtracks is reached. Using HBFS instead
of DFS in BTD should allow to quickly probe each subproblem to obtain interme-
diate upper and lower bounds for each of them. The upper bounds can be used to
quickly build a global solution, giving anytime behavior to BTD. The lower bounds
of all subproblems can be used to improve pruning in all other clusters.
The pseudocode of BTD-HBFS is described as Algorithm 3. It takes the same
arguments as BTD-DFS but ignores the last one (used only to pass informa-
tion from BTD-HBFS to BTD-DFS). BTD-HBFS relies on BTD-DFS pseudocode,
assuming that all grayed lines of BTD-DFS are active. These reactivated lines
in Algorithm 2 impose per-cluster and per-subproblem backtrack limits. Every
cluster Ce has a counter Ce .backtracks for number of backtracks performed inside
the cluster and an associated limit Ce .limit. Every subproblem Pe has a limit
Pe .limit on the number of backtracks N performed inside the subproblem. Ini-
tially, Ce .limit = Pe .limit = ∞ for all Ce ∈ C.
Every subproblem Pe |Ae has its own list of open nodes Pe |Ae .open for each
upper bound which it is given. The value of the upper bound participates in
the definition of the actual search space that needs to be explored. If the same
subproblem is revisited later with a lower upper bound, then the search space
shrinks and we can just copy the open list associated with the higher bound and
prune all nodes ν such that ν.lb ≥ ub. But if the upper bound is more relaxed
than any previous upper bound then we need to create a new open list starting
with the root node.
Finally, the loop in line 1 is interrupted as soon as the optimality gap reduces
or the number of backtracks reaches the subproblem limit, making the search
more dynamic. If subproblems quickly update their bound, the remaining back-
tracks can be used in a higher cluster. However, the subproblem under each
child cluster is guaranteed to get at least Z backtracks. The result is that we
spend most of the time in leaf clusters. When one cluster exhausts its budget,
the search quickly returns to the root cluster.

Example 1. Consider the example in figure 3. We have a CFN with the tree
decomposition given in the box labeled (C, T ) and the search in each cluster is
shown in a box labeled by that cluster’s name. Let N = 2 and Z = 1 in this
example. The search visits nodes as they are numbered in the figure. When it
 Anytime Hybrid Best-First Search with Tree Decomposition 21

reaches node 4, cluster C1 is completely instantiated and hence it descends into

C2 and after node 7 it descends into C4 . After node 10, we have performed a
backtrack in this cluster, and since Z = 1 we end this DFS probe and return
control to BTD-HBFS. The limit on the number of backtracks in P4 is still not
exceeded, so we choose a new node from the open list, node 11, a conflict. Again
we return control to BTD-HBFS and, having exceeded the backtrack limit on
P4 , exit this cluster, but with an improved lower bound. Since C4 exceeded its
backtrack limit before improving its lower bound, no more search is allowed in
parent clusters. The search is allowed, however, to visit sibling clusters, hence it
explores C5 (nodes 12–14), which it exits with an improved upper bound before
exceeding its backtrack limit, and C3 (nodes 15–18). Once it returns to node 7
after cluster C5 , that node is not closed, because one of the child clusters is not
closed. It is instead put back on the open list. Similarly node 4 is put back on
the open list of C1 . At that point, best-first search picks another node from the
open list of C1 , node 19, and continues from there. 


Function BTD-HBFS(A,Ce ,V ,clb,cub, ) : pair(integer,integer)

open := open list of Pe |A(cub) ;
if (open = ∅) then
if exists minimum cub s.t. cub > cub and open(Pe |A(cub )) = ∅ then
open = {ν ∈ open(Pe |A(cub )) | ν.lb < cub}
else
open = {∅} /* Contains only the root node at position clb */
Pe .limit := Backtracks + N /* Set a global backtrack limit for the subproblem */ ;
clb  := max(clb, lb(open)) ;
cub  := cub ;
1 while (open = ∅ and clb  < cub  and (Ce = C1 or (clb  = clb and cub  =
cub and Backtracks < Pe .limit))) do
ν :=pop(open) /* Choose a node with minimum lower bound and maximum
depth */ ;
Restore state ν.δ, leading to assignment Aν , maintaining local consistency ;
NodesRecompute := NodesRecompute + ν.depth ;
Ce .limit := Ce .backtracks + Z /* Set a depth-first search backtrack limit */ ;
(cub  , cub  ) :=BTD-DFS(Aν ,Ce ,Vν ,max(clb  , lb(ν), lb(Pe |Aν )),cub  ,BTD-HBFS) ;
clb  := max(clb  , lb(open)) ;
if (NodesRecompute > 0) then
if (NodesRecompute/Nodes > β and Z ≤ N ) then Z := 2 × Z;
else if (NodesRecompute/Nodes < α and Z ≥ 2) then Z := Z/2;
return (clb  , cub  ) /* invariant clb  ≥ clb and cub  ≤ cub */ ;

Algorithm 3. Hybrid Best-First Search with Tree Decomposition.

BTD-HBFS addresses both the issues of BTD that we identified above. First, it is
anytime, because as soon as UB Pe |Ae < k for all subproblems, we can combine the
assignments that gave these upper bounds to produce a global solution. Second, it
constantly updates lower bounds for all active subproblems, so the search effort in
each subproblem immediately exploits all other lower bounds.
22 D. Allouche et al.

(C, T ) C1
C1 1

C2 C3 2 19

C4 C5 3

C2 /A1 C3 /A1
5 15

6 16 17

7 18

C4 /A2 C5 /A2
8 12

9 11 13

10 14

Fig. 3. An example of a run of BTD-HBFS. White nodes are open, nodes with a black
border are conflicts (square) or solutions (circle). Grey nodes with a white border are
explored but put back in the open list. Grey nodes with no border are closed.

Like HBFS, BTD-HBFS can be made anyspace, i.e., its memory usage can be
limited to any amount beyond what is needed for BTD-DFS, including zero. The
cache of bounds can also be limited, at the expense of additional subproblem
recomputations, leading to worst-case complexity exponential in the tree decom-
position height.

Theorem 1. Given a CFN P with treewidth ω, BTD-HBFS computes the opti-

mum in time O(kndω+1 ) and space O(knd2ω ).

Proof (Sketch). For correctness, observe that BTD-DFS solves independent sub-
problems separately using DFS, hence using HBFS or any other solution method
does not affect correctness. Each leaf node in an internal cluster is closed only
 Anytime Hybrid Best-First Search with Tree Decomposition 23

when all child clusters are solved, hence all bounds for each subproblem and
open node list are correct. Finally, exploration at the root cluster continues until
the optimality gap is closed. Complexity stems from the complexity of BTD and
the additional overhead of storing open lists for each separator assignment and
upper bound. 


We implemented a simpler version of this algorithm with better space com-

plexity: each time BTD-HBFS is called on Pe |A with a higher upper bound than
previously stored, we wipe the open list and replace it with the root node of Ce .
This removes theoretical guarantees on the performance of the algorithm, but
does not hurt practical performance, as we will see.

4.2 Related Work

AND/OR Branch and Bound search has already been combined with BFS [20].
The resulting AOBF algorithm, has good worst-case time complexity similar to
BTD-DFS, but otherwise has the space-intensive non anytime behavior of BFS.
The poor anytime ability of BTD has been addressed by breadth-rotating
AND/OR search (BRAO) [21]. BRAO interleaves DFS on all components, so
it can combine the incumbents of all components to produce a global solution.
However, as it performs DFS on each component, it does not produce better
lower bounds.
OR-decomposition [12] is an anytime method that exploits lower bounds
produced by other clusters by performing DFS in which it interleaves variable
choices from all components, and uses caching to achieve the same effect as BTD.
However, the global lower bound it computes depends on the partial assignments
of all components. Thus it may revisit the same partial assignment of one com-
ponent many times. This may also inhibit its anytime behavior, as a high cost
partial assignment in one component will prevent other components from reach-
ing good solutions. Moreover, the local lower bound for each component is only
updated by visiting the right branch at its root.
Russian Doll Search [25], uses DFS to solve each cluster of a rooted tree
decomposition in topological order. This method is not anytime, as it cannot
produce a solution until it starts solving the root cluster. Moreover, it computes
lower bounds that are independent of the separator assignment, hence can be
lower than their true value.

5 Experimental Results

We used benchmark instances including stochastic graphical models from the

UAI evaluation in 2008 and 2010, the Probabilistic Inference Challenge 2011,
the Weighted Partial Max-SAT Evaluation 2013, the MiniZinc Challenge 2012
and 2013, Computer Vision and Pattern Recognition problems from OpenGM23

3
http://hci.iwr.uni-heidelberg.de/opengm2/
24 D. Allouche et al.

and additional instances from the CostFunctionLib4 . This is a total of more than
3,000 instances that can be encoded as Cost Function Networks, available at
http://genoweb.toulouse.inra.fr/∼degivry/evalgm, with domain sizes that range
from d = 2 to 503, n = 2 to 903, 884 variables, and e = 3 to 2, 912, 880 cost
functions of arity from r = 2 to 580. We used toulbar2 version 0.9.8.0-dev5 on
a cluster of 48-core Opteron 6176 nodes at 2.3 GHz with 378 GB RAM, with a
limit of 24 simultaneous jobs per node.
In all cases, the local lower bound is provided by maintaining EDAC [13].
The variable ordering includes both weighted-degree [4] and last-conflict [15]
heuristics. The value ordering is to select the EAC support value first [13]. All
executions used a min-fill variable ordering for DAC preprocessing. For HBFS, we
set the node recomputation parameters to [α, β] = [5%, 10%] and the backtrack
limit N to 10, 000.
The methods based on BTD use a different value ordering heuristic: if a
solution is known for a cluster, it keeps the same value if possible and if not
uses EAC support values as the previous methods. A min-fill ordering is used
for building a tree decomposition. Children of a cluster are statically sorted by
minimum separator size first and smallest number of subproblem variables next.
Our aim is to determine whether HBFS is able to improve over DFS both in
terms of number of problems solved (including the optimality proof) and in its
anytime behavior. Similarly, we compare BTD-HBFS to BTD-DFS. We include in
our comparison two methods that are known to significantly improve the upper
bound anytime behavior of DFS: Limited Discrepancy Search [8] and DFS with
Luby restarts [18].
We also include results from BRAO [21] using the daoopt solver6 with static
mini-bucket lower bounds of different strength (i-bound set to 15 and 35) and
without local search nor iterative min-fill preprocessing. daoopt is restricted to
cost functions expressed by complete tables, hence we exclude most MaxSAT
families (except MIPLib and MaxClique) in tests where we use it.

5.1 Proving Optimality

We show in figure 4 a cactus plot comparing all the algorithms that do not use
tree decompositions, but also include BTD-HBFSrk for reference (see below). We
see that HBFS is the best performing decomposition-unaware algorithm. It out-
performs DFS and DFS with Luby restarts significantly, and slightly outperforms
LDS.
Although our benchmark set includes very large instances, and our HBFS
implementation does not include automatic control of space usage, no instance
required more than 32 GB. The median memory usage was 36.8 MB for DFS
and 38.2 MB for hybrid BFS. The worst-case largest ratio between HBFS and

4
https://mulcyber.toulouse.inra.fr/projects/costfunctionlib
5
Available in the git repository at https://mulcyber.toulouse.inra.fr/projects/
toulbar2/ in the bfs branch.
6
https://github.com/lotten/daoopt
 Anytime Hybrid Best-First Search with Tree Decomposition 25

Fig. 4. Number of solved instances within a given time. Methods in the legend are
sorted at time=20min.

B
12.1M B = 31.4 on MRF Grid instance grid20x20.f15 (unsolved in
DFS was 379.8M
one hour by both methods).
In figure 5, we compare algorithms exploiting a tree decomposition (BTD-
like) or a pseudo-tree (BRAO).We see that BTD-HBFS slightly outperforms BTD-
DFS, both outperforming BRAO. However, many of these instances have large
treewidth and BTD-like methods are not ideal for these. Even for instances with
small treewidth, the decomposition is often a deep tree in which each cluster
shares all but one variables with its parent. In these cases, following the tree
decomposition imposes a static variable ordering on BTD, while HBFS degrades
to DFS. Finding good tree decompositions is not straightforward [10,11]. A sim-
ple way to improve one is to merge clusters until no separator has size greater
than k, even if this increases width. We call the algorithms that apply this
BTD-DFSrk and BTD-HBFSrk . Figure 5 includes results for BTD-DFSr4 and BTD-
HBFSr4 . BTD-DFSr4 is significantly better than BTD-DFS and BTD-HBFSr4 out-
performs BTD-DFSr4 by an even greater margin. BTD-HBFSr4 is also the overall
best performer as shown in figure 4.

5.2 Anytime Behavior

To analyze the algorithms’ anytime behavior, we first show in figure 6 the evo-
lution of the lower and upper bounds for two instances: the SPOT5 404 (left)
and the RLFAP CELAR06 instances (right). We solve both instances using DFS,
LDS, DFS with Luby restarts, HBFS, BTD-DFS and BTD-HBFS. In both instances,
we see that HBFS and BTD-HBFS improve significantly on the upper bound any-
time ability of DFS and BTD-DFS, respectively. Moreover, the lower bound that
they report increases quickly in the beginning and keeps increasing with time.
For all other algorithms, the lower bound increases by small amounts and infre-
quently, when the left branch of the root node is closed. The HBFS variants are
as fast as the base algorithms in proving optimality.
26 D. Allouche et al.

Fig. 5. Number of solved instances as time passes on a restricted benchmark set

(without MaxSAT). Methods in the legend are sorted at time=20min.
130
HBFS
DFS-Luby
120 DFS-LDS
DFS
110

100

90

80

70

60
0.1 1 10 100 1000

Fig. 6. Evolution of the lower and upper bounds (Y axis, in cost units) as time (X axis,
in seconds) passes for HBFS, Luby restart, LDS, and DFS on SPOT5 404 instance (left)
and also BTD, and BTD-HBFS for the RLFAP CELAR06 instance (right). Methods are
sorted in increasing time to find the optimum. For each curve, the first point represents
the time where the optimum is found and the second point the time (if any) of proof
of optimality.

In figure 7, we summarize the evolution of lower and upper bounds for each
algorithm over all instances that required more than 5 sec to be solved by DFS.
Specifically, for each instance I we normalize all costs as follows: the initial
lower bound produced by EDAC (which is common to all algorithms) is 0; the
best – but potentially suboptimal – solution found by any algorithm is 1; the
worst solution is 2. This normalization is invariant to translation and scaling.
Additionally, we normalize time from 0 to 1 for each pair of algorithm A and
instance I, so that preprocessing ends at time 0 and each run finishes at time
1. This time normalization is different for different instances and for different
algorithms on the same instance. A point x, y on the lower bound line for
algorithm A in figure 7 means that after normalized runtime x, algorithm A has
proved on average over all instances a normalized lower bound of y and similarly
for the upper bound. We show both the upper and lower bound curves for all
 Anytime Hybrid Best-First Search with Tree Decomposition 27

Fig. 7. Average evolution of normalized upper and lower bounds for each algorithm.

algorithms evaluated here. In order for the last point of each curve to be visible,
we extend all curves horizontally after 1.0.
This figure mostly ignores absolute performance in order to illustrate the
evolution of upper and lower bounds with each algorithm, hence cannot be inter-
preted without the additional information provided by the cactus plots in figures
4 and 5. It confirms that HBFS improves on DFS in terms of both upper and
lower bound anytime behavior and similarly for BTD-HBFSr4 over BTD-DFSr4
and BRAO, with the latter being especially dramatic. The two HBFS variants
are, as expected, significantly better than all other algorithms in terms of the
lower bounds they produce. HBFS and BTD-HBFSr4 produce solutions of the
same quality as LDS, while DFS-Luby is slightly better than this group on this
restricted benchmark set (without MaxSAT).
Despite the fact that time to solve an instance is normalized away in figure
7, it does give some information that is absent from the cactus plots and that
is the average normalized lower and upper bounds at time 1. Figure 7 tells us
that DFS-Luby finds the best solution most often, as its upper bound curve is
the lowest at time 1. It is followed closely by the HBFS variants and LDS, while
DFS and BTD-DFSr4 are significantly worse. On the other hand, DFS-Luby is
significantly worse than the HBFS variants in the cactus plot. HBFS and BTD-
HBFSr4 give better lower bounds in those instances that they failed to solve, so
their lower bound curves are higher at point 1.

6 Conclusions
Hybrid BFS is an easily implemented variant of the Branch and Bound algo-
rithm combining advantages of BFS and DFS. While being a generic strategy,
applicable to essentially any combinatorial optimization framework, we used it
to improve Depth-First Branch and Bound maintaining soft arc consistency and
28 D. Allouche et al.

tested it on a large benchmark set of problems from various formalisms, includ-

ing Cost Function Networks, Markov Random Field, Partial Weighted MaxSAT
and CP instances representing a variety of application domains in bioinformat-
ics, planning, resource allocation, image processing and more. We showed that
HBFS improves on DFS or DFS equipped with LDS or restarts in terms of num-
ber of problems solved within a deadline but also in terms of anytime quality
and optimality gap information.
HBFS is also able to improve Tree Decomposition aware variants of DFS such
as BTD, being able to solve more problems than the previous DFS based BTD
on the same set of benchmarks. BTD is targeted at problems with relatively
low treewidth and has been instrumental in solving difficult radio-link frequency
assignment problems. On such problems, BTD-HBFS provides to BTD the same
improvements as to DFS.
Its ability to provide feedback on the remaining search effort, to describe the
current remaining search space in a list of open nodes and to decompose search
in self-interrupted DFS probes makes it a very dynamic search method, very
attractive for implementing multi-core search.

Acknowledgments. We are grateful to the Genotoul (Toulouse) Bioinformatic plat-

form for providing us computational support for this work.

References
1. Ansótegui, C., Bonet, M.L., Gabàs, J., Levy, J.: Improving sat-based weighted
maxsat solvers. In: Proc. of CP 2012, Québec City, Canada, pp. 86–101 (2012)
2. van den Berg, J., Shah, R., Huang, A., Goldberg, K.: ANA*: Anytime nonparamet-
ric A*. In: Proceedings of Twenty-Fifth AAAI Conference on Artificial Intelligence
(AAAI 2011) (2011)
3. Berthold, T.: Primal heuristics for mixed integer programs. Master’s thesis, Tech-
nischen Universität Berlin (2006). urn:nbn:de:0297-zib-10293
4. Boussemart, F., Hemery, F., Lecoutre, C., Sais, L.: Boosting systematic search by
weighting constraints. In: ECAI, vol. 16, p. 146 (2004)
5. Cooper, M., de Givry, S., Sanchez, M., Schiex, T., Zytnicki, M., Werner, T.: Soft
arc consistency revisited. Artificial Intelligence 174(7), 449–478 (2010)
6. Dechter, R., Mateescu, R.: And/or search spaces for graphical models. Artificial
Intelligence 171(2), 73–106 (2007)
7. de Givry, S., Schiex, T., Verfaillie, G.: Exploiting tree decomposition and soft local
consistency in weighted CSP. In: Proc. of the National Conference on Artificial
Intelligence, AAAI 2006, pp. 22–27 (2006)
8. Harvey, W.D., Ginsberg, M.L.: Limited discrepency search. In: Proc. of the 14th
IJCAI, Montréal, Canada (1995)
9. Jégou, P., Terrioux, C.: Hybrid backtracking bounded by tree-decomposition of
constraint networks. Artif. Intell. 146(1), 43–75 (2003)
10. Jégou, P., Terrioux, C.: Combining restarts, nogoods and decompositions for solv-
ing csps. In: Proc. of ECAI 2014, Prague, Czech Republic, pp. 465–470 (2014)
11. Jégou, P., Terrioux, C.: Tree-decompositions with connected clusters for solving
constraint networks. In: Proc. of CP 2014, Lyon, France, pp. 407–423 (2014)
 Anytime Hybrid Best-First Search with Tree Decomposition 29

12. Kitching, M., Bacchus, F.: Exploiting decomposition in constraint optimization

problems. In: Stuckey, P.J. (ed.) CP 2008. LNCS, vol. 5202, pp. 478–492. Springer,
Heidelberg (2008)
13. Larrosa, J., de Givry, S., Heras, F., Zytnicki, M.: Existential arc consistency: get-
ting closer to full arc consistency in weighted CSPs. In: Proc. of the 19th IJCAI,
pp. 84–89, Edinburgh, Scotland (August 2005)
14. Lawler, E., Wood, D.: Branch-and-bound methods: A survey. Operations Research
14(4), 699–719 (1966)
15. Lecoutre, C., Saı̈s, L., Tabary, S., Vidal, V.: Reasoning from last conflict(s) in
constraint programming. Artificial Intelligence 173, 1592–1614 (2009)
16. Likhachev, M., Gordon, G.J., Thrun, S.: ARA*: Anytime A* with provable bounds
on sub-optimality. In: Advances in Neural Information Processing Systems, p. None
(2003)
17. Linderoth, J.T., Savelsbergh, M.W.: A computational study of search strategies
for mixed integer programming. INFORMS Journal on Computing 11(2), 173–187
(1999)
18. Luby, M., Sinclair, A., Zuckerman, D.: Optimal speedup of las vegas algorithms. In:
Proceedings of the 2nd Israel Symposium on the Theory and Computing Systems,
pp. 128–133. IEEE (1993)
19. Marinescu, R., Dechter, R.: AND/OR branch-and-bound for graphical models. In:
Proc. of IJCAI 2005, Edinburgh, Scotland, UK, pp. 224–229 (2005)
20. Marinescu, R., Dechter, R.: Best-first AND/OR search for graphical models. In:
Proceedings of the National Conference on Artificial Intelligence, pp. 1171–1176.
AAAI Press, MIT Press, Menlo Park, Cambridge (1999, 2007)
21. Otten, L., Dechter, R.: Anytime and/or depth-first search for combinatorial opti-
mization. AI Communications 25(3), 211–227 (2012)
22. Pearl, J.: Heuristics – Intelligent Search Strategies for Computer Problem Solving.
Addison-Wesley Publishing Comp. (1985)
23. Pohl, I.: Heuristic search viewed as path finding in a graph. Artificial Intelligence
1(3), 193–204 (1970)
24. Robertson, N., Seymour, P.D.: Graph minors. II. Algorithmic aspects of tree-width.
Journal of Algorithms 7(3), 309–322 (1986)
25. Sanchez, M., Allouche, D., de Givry, S., Schiex, T.: Russian doll search with tree
decomposition. In: IJCAI, pp. 603–608 (2009)
26. Schulte, C.: Comparing trailing and copying for constraint programming. In: Logic
Programming, Las Cruces, New Mexico, USA, pp. 275–289 (1999)
27. Stern, R., Kulberis, T., Felner, A., Holte, R.: Using lookaheads with optimal best-
first search. In: AAAI (2010)
28. Terrioux, C., Jégou, P.: Bounded backtracking for the valued constraint satisfaction
problems. In: Rossi, F. (ed.) CP 2003. LNCS, vol. 2833, pp. 709–723. Springer,
Heidelberg (2003)
 Improved Constraint Propagation via
Lagrangian Decomposition

David Bergman1 , Andre A. Cire2 , and Willem-Jan van Hoeve3(B)

1
School of Business, University of Connecticut, Mansfield, USA
david.bergman@business.uconn.edu
2
University of Toronto Scarborough, Toronto, Canada
acire@utsc.utoronto.ca
3
Tepper School of Business, Carnegie Mellon University, Pittsburgh, USA
vanhoeve@andrew.cmu.edu

Abstract. Constraint propagation is inherently restricted to the local

information that is available to each propagator. We propose to improve
the communication between constraints by introducing Lagrangian
penalty costs between pairs of constraints, based on the Lagrangian
decomposition scheme. The role of these penalties is to force variable
assignments in each of the constraints to correspond to one another. We
apply this approach to constraints that can be represented by decision
diagrams, and show that propagating Lagrangian cost information can
help improve the overall bound computation as well as the solution time.

1 Introduction
Modern finite-domain constraint programming (CP) solvers employ a constraint
propagation process in which domain changes for the variables are propagated
between constraints. To allow for more communication and knowledge sharing
between constraints, several techniques have been proposed. One possibility is to
propagate more structural information than variable domains, such as (relaxed)
decision diagrams [1,10]. Another option, in the context of optimization prob-
lems, is to combine constraints with the objective function, and utilize mathe-
matical programming relaxations for stronger cost-based filtering [7,15]. These
approaches, however, have in common that consistency checks are done sepa-
rately and independently for each constraint. Higher-order consistencies, such as
pairwise consistency [12] can consider multiple constraints simultaneously, but
may suffer from a relatively high computational cost.
We propose an alternative, and generic, approach to improve the propaga-
tion between constraints based on Lagrangian decomposition [9]. In Lagrangian
decomposition, the constraint set of a given problem is partitioned into struc-
tured subproblems, each of which is defined on a duplicate copy of the original
variables. To link the subproblems, constraints are added to ensure that each
of the duplicates is equal to the original variable. These latter constraints are
then relaxed with an associated Lagrangian multiplier, and moved to the objec-
tive. This results in independent subproblems that can be separately optimized.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 30–38, 2015.
DOI: 10.1007/978-3-319-23219-5 3
 Improved Constraint Propagation via Lagrangian Decomposition 31

Intuitively, the idea is to force the variables to take the same value in each con-
straint, via the Lagrangian penalties, which are iteratively updated. This will
somehow synchronize the consistency checks for each of the constraints; instead
of allowing each constraint to check its consistency w.r.t. an arbitrary tuple, we
iteratively arrive at tuples with minimal disagreement.
Since constraint programming has been designed to work with (global) con-
straints that capture a specific structure of the problem, the application of
Lagrangian decomposition in this context seems natural and promising. Indeed,
we show that the Lagrangian decomposition is not only useful to improve the
bound on the objective, but can also be applied for cost-based domain filtering.
The structure of the paper is as follows. We first provide an overview of the
most relevant related work in Section 2. In Section 3 we recall the Lagrangian
decomposition scheme. We apply this to constraint programming models in
Section 4. Experimental results on instances with multiple alldiff constraints
are given in Section 5, while Section 6 provides results on set covering problems.
We conclude in Section 7.

2 Related Work
Lagrangian relaxations have been widely applied in operations research as well
as constraint programming. One of the first applications in CP is the work by
Benoist et al. [2] on the Traveling Tournament Problem. A formal treatment
was provided by Sellmann [16] who showed that optimal Lagrangian multipliers
may not result in the most effective domain filtering. Recently, [8] introduced
a framework for automated Lagrangian relaxation in a constraint programming
context. That work explicitly generalizes Lagrangian relaxations to CP problems
using measures of constraint violations, or degrees of satisfiability.
Adapting weights for improving propagation has also been applied in the
context of Valued Constraint Satisfaction Problems [6]. In that work, a lin-
ear programming model is proposed for computing Optimal Soft Arc Consis-
tency, but Lagrangian relaxations are not used. Khemmoudj et al. [13] combine
arc consistency with Lagrangian relaxation for filtering constraint satisfaction
problems (CSPs). They consider binary CSPs (i.e., each constraint has at most
two variables in its scope) in extensional form. Lastly, Bergman et al. [3] intro-
duce Lagrangian relaxations in the context of propagating (relaxed) decision
diagrams.

3 Lagrangian Decomposition
Lagrangian decomposition has been introduced to strengthen Lagrangian bounds
for integer linear optimization problems [9]. Consider an integer linear program
of the form:
(P ) max{f x | Ax ≤ b, Cx ≤ d, x ∈ X},
for some feasible set X, where x ∈ Rn is a vector of variables, f ∈ Rn represents
a ‘weight’ vector, A and C represent constraint coefficient matrices, and b and
32 D. Bergman et al.

c are constant right-hand size vectors. This is equivalent to the reformulated

program
max{f x | Ay ≤ b, Cx ≤ d, x = y, x ∈ X, y ∈ Y },
for any set Y containing X.
The Lagrangian decomposition of P consists in dualizing the equality con-
straints x = y with Lagrangian multipliers λ ∈ Rn :

LP (λ) := max{f x + λ(y − x) | Cx ≤ d, x ∈ X, Ay ≤ b, y ∈ Y }

= max{(f − λ)x | Cx ≤ d, x ∈ X} + max{λy | Ay ≤ by ∈ Y }

The Lagrangian dual is to find those Lagrangian multipliers λ that provide the
best bound:
min LP (λ).
λ

Guignard and Kim [9] show that the optimal bound obtained from this
Lagrangian decomposition is at least as strong as the standard Lagrangian
bounds from dualizing either Ax ≤ b or Cx ≤ d. Lagrangian decomposition may
be particularly useful when the problem is composed of several well-structured
subproblems, such as those defined by (global) constraints in CP models.

4 Application to Constraint Programming

We apply Lagrangian decomposition to constraint optimization problems

(COPs), which include constraint satisfaction problems (CSPs) as special case.
It is important to note that this approach will transform each of the original
constraints into an ‘optimization constraint’; instead of representing a witness
for feasibility the constraint now has to represent a witness for optimality, even
if the constraint is not directly linked to the objective function, or in case of
feasibility problems.
When the variables have numeric domains, the method from Section 3 can
be directly applied. In general, however, domains need not be numeric, and we
will therefore focus our discussion on this more general case. Consider a COP
with variables x1 , . . . , xn that have given finite domains xi ∈ Di :

max f (x1 , . . . , xn )
s.t. Cj (x1 , . . . , xn ) j ∈ {1, . . . , m} (1)
xi ∈ Di i ∈ {1, . . . , n}

For simplicity we assume here that all variables appear in all constraints, but
we can allow a different subset of variables for each constraint. Also, Cj may
represent any substructure, for example a global constraint, a table constraint, or
a collection of constraints. We introduce for each variable xi and each constraint
j = 1, . . . , m a duplicate variable yij with domain Di . We let the set yi1 represent
 Improved Constraint Propagation via Lagrangian Decomposition 33

our ‘base’ variables, to which we will compare the variables yij for j = 2, . . . , m.
The reformulated COP is as follows:
max f (y11 , . . . , yn1 )
s.t. Cj (y1j , . . . , ynj ) j ∈ {1, . . . , m}
yi1 = yij i ∈ {1, . . . , n}, j ∈ {2, . . . , m}
yij ∈ Di i ∈ {1, . . . , n}, j ∈ {1, . . . , m}

To establish the Lagrangian decomposition, we relax the constraints yi1 = yij

and move these into the objective as yi1 = yij with associated Lagrangian multi-
pliers. To measure its violation, we propose to represent yi1 = yij with the set of
constraints ((yij = v) − (yi1 = v)) for all v ∈ Di , where (yij = v) is interpreted as
a binary value representing the truth value of the expression. Lastly, we define
a Lagrangian multiplier for each i, j (j ≥ 2) and each v ∈ D(xi ) as a vector
j
λi := λji [v].

The Lagrangian objective function can then be written as:

m 
 n 
max f (y11 , . . . , yn1 ) + λji [v]((yij = v) − (yi1 = v))
j=2 i=1 v∈D(xi )
n 
m 
 
= f (y11 , . . . , yn1 ) + λji [yij ] − λji [yi1 ]
j=2 i=1

j
This leads to the following decomposition (for any given set of multipliers λi ):
⎧ ⎫
⎨ m  n ⎬
max f (y11 , . . . , yn1 ) − λji [yi1 ] | C1 (y11 , . . . , yn1 )
⎩ ⎭
m
  n j=2 i=1 
  j j
+ max λi [yi ] | Cj (y1j , . . . , ynj )
j=2 i=1

which are m independent subproblems.

mLet zj be the optimal objective value for
subproblem j ∈ {1, . . . , m}. Then j=1 zj is a valid bound on f (x1 , . . . , xn ).
Note that the duplicate variables have only been introduced for the formal
description of the method. In practice, all constraints Cj use the original vari-
ables.

Design Choices. The Lagrangian decomposition scheme can be adapted by

allocating parts of original objective to different subproblems. Moreover, we can
introduce equality constraints between any pair of subproblems. We will illus-
trate the latter in the following example.
34 D. Bergman et al.

Example 1. Consider the following CSP:

C1 : alldiff(x1 , x2 , x3 ) C2 : alldiff(x2 , x4 , x5 ) C3 : alldiff(x3 , x5 )

x1 ∈ {a, b}, x2 ∈ {b, c}, x3 ∈ {a, c}, x4 ∈ {a, b}, x5 ∈ {a, b, c}

This CSP is domain consistent as well as pairwise consistent, and has one solution
(x1 , x2 , x3 , x4 , x5 ) = (b, c, a, a, b).
We construct a Lagrangian decomposition based on the constraints
C1 , C2 , C3 . To link these, we only need to introduce the constraints y22 = y21 ,
y33 = y31 , y53 = y52 , and their associated multipliers. This yields the following
three subproblems, with respective objective values z1 , z2 , z3 :
 2 3

z1 = max −λ2 [y21 ] − λ3 [y31 ] | alldiff(y11 , y21 , y31 )
 2 3

z2 = max λ2 [y22 ] − λ5 [y52 ] | alldiff(y22 , y42 , y52 )
 3 3

z3 = max λ3 [y33 ] + λ5 [y53 ] | alldiff(y33 , y53 )

This CSP can be considered as a COP with a zero-valued objective function

so that the value z1 + z2 + z3 is an upper bound on the satisfiability of this
problem, for any Lagrangian multipliers; if the bound is below zero, the problem
is unsatisfiable. And so, the optimal Lagrangian decomposition bound is 0.  

Cost-Based Domain Filtering. In addition to pruning the search based on

the overall bound LP (λ) and a given lower bound B, we can apply cost-based
domain filtering. The difference with existing cost-based filtering methods is that
the bounds from the different subproblems can all be conditioned on a specific
variable/value pair. To this end, let zj |xi =v be the optimal objective value for
subproblem j ∈ {1, . . . , m} in which yij = v. We have the following result:

Proposition 1. If j zj |xi =v < B then v can be removed from Di .
This result may be particularly effective when there is no single subproblem that
collects all variables. We continue our example to give an illustration.

Example 2. Continuing Example 1, consider the following Lagrangian multi-

pliers (all others are zero): λ22 [b] = 0.5, λ33 [a] = 0.5, λ35 [a] = 0.5. This yields
z1 = −0.5, z2 = 0.5, z3 = 0.5, and a total bound of 0.5. Even though this is not
optimal, when we condition x2 = b or x5 = c, the bound becomes −0.5 in both
cases, and by Proposition 1 we can remove those values from their respective
domains. We can similarly remove values a from D1 , b from D2 , and c from D3
using the multipliers λ22 [c] = −0.5, λ33 [c] = 0.5, λ35 [c] = 0.5. 


Example 2 implies the following result:

Proposition 2. Cost-based filtering based on Lagrangian decomposition can be
stronger than pairwise consistency.
 Improved Constraint Propagation via Lagrangian Decomposition 35

Implementation Issues. To apply Lagrangian propagation efficiently, it is

important that each constraint is optimized efficiently. For many constraints
optimization versions are already available [11], to which the Lagrangian costs
can be immediately added. For example, in our experiments we represent con-
straints by decision diagrams, which permit to find the optimal solution quickly
via a shortest (or longest) path calculation. Also, cost-based propagators are
available that filter sub-optimal arcs from the decision diagram. Second, the
search for optimal Lagrangian multipliers can be done with different methods
[14]. Regardless, any set of multipliers results in a valid relaxation, and we do
not necessarily need to solve the Lagrangian dual to optimality. In our imple-
mentation, we compute the multipliers once at the root node and reuse them
during the CP search process.

5 Application: Multiple Alldifferent Constraints

As first application, we consider systems of multiple overlapping alldiff con-
straints, as in [1]. These are defined on a set X = {x1 , . . . , xn } of variables with
domain {1, . . . , n}. Each alldiff constraint is defined on a subset of variables
Sj ⊂ X, for j = 1, . . . , k. We then consider the following COP:
 n

max wi xi | alldiff(Sj ) ∀j ∈ {1, . . . , k}
i=1

We generated instances with n = 10, 11, 12, 13, 14, k = 4, and |Sj | = n − 2 for
all j = 1, . . . , 5. For the Lagrangian decomposition, we partition the alldiff
constraints into two arbitrary subsets of size two, and define one multi-valued
decision diagram (MDD) for each subset. In other words, we apply MDD prop-
agation to these subsets of alldiff constraints. The two MDDs thus formed
are the basis for the Lagrangian decomposition, which follows the description in
Section 4 (where the j-th MDD represents constraint set Cj ).
We implemented the MDD propagation as well as the Lagrangian decom-
position as a global constraint in IBM ILOG CPO 12.6, similar to [5]. The
(near-)optimal Lagrangian multipliers were computed using the Kelly-Cheney-
Goldstein method [14], using IBM ILOG CPLEX 12.6 as the linear programming
solver. We fix the CP search to be lexicographic in the order of the variables, to
ensure the search tree is the same across all instances. We compare the perfor-
mance with and without Lagrangian multipliers.
In Figure 1.a we show the root node percent gap for the 25 instances (where
the optimal value was obtained by formulating an integer linear program and
solving the instances using CPLEX). The reduction in the gap can be sub-
stantial, in some case several orders of magnitude. This reduction in the opti-
mality gap and additional cost-based filtering due to the Lagrangian multipliers
enables more instances to be solved in shorter computational time, as depicted in
Figure 1.b. Depending on the configuration of the alldiff systems the improve-
ment can be marginal and in some cases negligible.
36 D. Bergman et al.

a. Root node gap comparison b. Performance plot

Fig. 1. Evaluating the impact of the Lagrangian decomposition on systems of mul-

tiple alldiff constraints. (a) compares the root node gap obtained with Lagrangian
decomposition (Mdd Lag Gap) and without (Mdd Gap) and (b) depicts a performance
profile comparing the number of instances solved (N Solved) within a given time limit
(horizontal axis) with Lagrangian decomposition (With Lag) and without (No Lag).

6 Application: Set Covering

The set covering problem is defined on a universe of n elements U = {1, . . . , n}.
Given a collection of subsets C1 , . . . , Cm ⊆ U and weights wi (i = 1, . . . , n), the
problem is to find a set of elements S ⊂ U of minimum total weight such that all
S ∩ Cj is not empty for all j = 1, . . . , m. Using a binary variable xi to represent
whether element i is in S, the problem can be formulated as the following COP:
⎧ ⎫
⎨ n  ⎬
min wi xi | xi ≥ 1 ∀j ∈ {1, . . . , m}, xi ∈ {0, 1} ∀i ∈ {1, . . . , n}
⎩ ⎭
i=1 i∈Cj

Instead of defining a subproblem for each separate constraint, we create exact

binary decision diagram (BDD) representations for collections of them. That
is, using the construction method described in [4], we create a BDD by adding
constraints one at the time, until the exact width exceeds a given limit (in our
case 100 nodes on any layer). We then create the next BDD, and so forth. This
forms a partition of the constraint set, each of which is represented by an exact
BDD. For the instances we considered, we construct 10 or 11 BDDs per instance.
We also slightly modify the Lagrangian decomposition method by represent-
ing the original objective function in each of the BDDs, and dualizing constraints

xji = xji for every pair (j, j  ). Hence, the Lagrangian bound is no longer the sum
of the bounds of the respective BDDs, rather the average over the objectives.
In previous work [4], it was shown that the bounds from BDDs were most
effective when the constraint matrix has a relatively small bandwidth. We there-
fore used the same benchmark generator to evaluate the impact of the Lagrangian
decomposition for increasing bandwidths. We generated instances with n = 150
variables, randomly generated costs, and uniform-randomly selected subsets Cj
from within a given bandwidth of size 55 to 75 (five instances for each band-
width). To generate the costs, we let c(i) represent the number of subsets Cj that
 Improved Constraint Propagation via Lagrangian Decomposition 37

a. Impact of Lagrangian decomposition b. Comparison with single relaxed BDD

Fig. 2. Evaluating the bound from Lagrangian decomposition for set covering problems
of varying bandwidth.

contain element i. Then the cost for variable xi is taken uniform randomly in
[0.75∗c(i), 1.25∗c(i)]. The results are shown in Figure 2, showing the average over
the five instances per bandwidth. Figure 2.a depicts four lines: the optimal solu-
tion (found by CPLEX), the average bound without using Lagrangian decom-
position, the maximum bound without using Lagrangian decomposition, and
lastly the average bound when using the Lagrangian decomposition. Lagrangian
decomposition generates bounds of much better quality than the independent
BDDs. For example, for bandwidth 65 the average bound of 27.53 is improved
to 77.80 using the Lagrangian decomposition, on average.
We also compare the Lagrangian decomposition to the original BDD relax-
ation from [4] that represents all constraints in a single BDD respecting a given
maximum width. A larger width leads to a stronger relaxation and better bounds.
Figure 2.b compares the percent gap (between the lower bound and the optimal
solution) of the BDD relaxation for maximum widths 2,000 and 20,000 with
that of the Lagrangian decomposition. We note that the BDD relaxation with
maximum width 2,000 has about the same memory requirements as the separate
BDDs for the Lagrangian decomposition. As the bandwidth increases, the qual-
ity of BDD relaxation rapidly declines, while the Lagrangian decomposition is
much more stable and outperforms the BDD relaxation (decreasing the gap from
117.6% to 33% for bandwidth 75 and maximum width 2,000). This demonstrates
that Lagrangian decompositions can be used to improve BDD-based optimiza-
tion when a single BDD relaxation can no longer provide sufficient power to
represent the entire problem. We do note, however, that the Lagrangian decom-
position takes more time to compute (on average 60s) compared to the single
BDD relaxation (on average 1.4s for width 2,000 and 17s for width 20,000).

7 Conclusion
We have introduced Lagrangian decomposition in the context of constraint pro-
gramming as a generic approach to improve the constraint propagation process.
The key idea is that we penalize variables in different constraints to take differ-
ent assignments. We have shown how this approach can be utilized for stronger
38 D. Bergman et al.

cost-based domain filtering, and that it leads to improved bounds for systems of
alldiff constraints and set covering problems.

References
1. Andersen, H.R., Hadzic, T., Hooker, J.N., Tiedemann, P.: A constraint store based
on multivalued decision diagrams. In: Bessière, C. (ed.) CP 2007. LNCS, vol. 4741,
pp. 118–132. Springer, Heidelberg (2007)
2. Benoist, T., Laburthe, F., Rottembourg, B.: Lagrange relaxation and con-
straint programming collaborative schemes for traveling tournament problems. In:
Proceedings of the International Workshop on Integration of Artificial Intelligence
and Operations Research Techniques in Constraint Programming for Combinato-
rial Optimization Problems (CPAIOR 2001) (2001)
3. Bergman, D., Cire, A.A., van Hoeve, W.J.: Lagrangian Bounds from Decision Dia-
grams. Constraints 20(3), 346–361 (2015)
4. Bergman, D., van Hoeve, W.-J., Hooker, J.N.: Manipulating MDD relaxations for
combinatorial optimization. In: Achterberg, T., Beck, J.C. (eds.) CPAIOR 2011.
LNCS, vol. 6697, pp. 20–35. Springer, Heidelberg (2011)
5. Cire, A.A., van Hoeve, W.J.: Multivalued Decision Diagrams for Sequencing Prob-
lems. Operations Research 61(6), 1411–1428 (2013)
6. Cooper, M.C., de Givry, S., Sanchez, M., Schiex, T., Zytnicki, M., Werner, T.: Soft
arc consistency revisited. Artificial Intelligence 174(7–8), 449–478 (2010)
7. Focacci, F., Lodi, A., Milano, M.: Cost-based domain filtering. In: Jaffar, J. (ed.)
CP 1999. LNCS, vol. 1713, pp. 189–203. Springer, Heidelberg (1999)
8. Fontaine, D., Michel, L., Van Hentenryck, P.: Constraint-based lagrangian relax-
ation. In: O’Sullivan, B. (ed.) CP 2014. LNCS, vol. 8656, pp. 324–339. Springer,
Heidelberg (2014)
9. Guignard, M., Kim, S.: Lagrangian Decomposition: A Model Yielding Stronger
Lagrangian Bounds. Mathematical Programming 39, 215–228 (1987)
10. Hoda, S., van Hoeve, W.-J., Hooker, J.N.: A systematic approach to MDD-
based constraint programming. In: Cohen, D. (ed.) CP 2010. LNCS, vol. 6308,
pp. 266–280. Springer, Heidelberg (2010)
11. van Hoeve, W.J., Katriel, I.: Global constraints. In: Handbook of Constraint Pro-
gramming, pp. 169–208. Elsevier (2006)
12. Janssen, P., Jégou, P., Nouguier, B., Vilarem, M.C.: A filtering process for general
constraint-satisfaction problems: achieving pairwise-consistency using an associ-
ated binary representation. In: IEEE International Workshop on Tools for Arti-
ficial Intelligence, Architectures, Languages and Algorithms, pp. 420–427. IEEE
(1989)
13. Khemmoudj, M.O.I., Bennaceur, H., Nagih, A.: Combining arc-consistency and
dual lagrangean relaxation for filtering CSPS. In: Barták, R., Milano, M. (eds.)
CPAIOR 2005. LNCS, vol. 3524, pp. 258–272. Springer, Heidelberg (2005)
14. Lemaréchal, C.: Lagrangian relaxation. In: Jünger, M., Naddef, D. (eds.) Com-
putational Combinatorial Optimization. LNCS, vol. 2241, pp. 112–156. Springer,
Heidelberg (2001)
15. Régin, J.-C.: Arc consistency for global cardinality constraints with costs. In:
Jaffar, J. (ed.) CP 1999. LNCS, vol. 1713, pp. 390–404. Springer, Heidelberg (1999)
16. Sellmann, M.: Theoretical foundations of cp-based lagrangian relaxation. In:
Wallace, M. (ed.) CP 2004. LNCS, vol. 3258, pp. 634–647. Springer, Heidelberg
(2004)
 Strengthening Convex Relaxations with Bound
Tightening for Power Network Optimization

Carleton Coffrin1,2(B) , Hassan L. Hijazi1,2 , and Pascal Van Hentenryck1,2

1
NICTA - Optimisation Research Group, Canberra, Australia
2
College of Engineering and Computer Science, Australian National University,
Canberra, Australia
{carleton.coffrin,hassan.hijazi,pvh}@nicta.com.au

Abstract. Convexification is a fundamental technique in (mixed-

integer) nonlinear optimization and many convex relaxations are
parametrized by variable bounds, i.e., the tighter the bounds, the
stronger the relaxations. This paper studies how bound tightening can
improve convex relaxations for power network optimization. It adapts
traditional constraint-programming concepts (e.g., minimal network and
bound consistency) to a relaxation framework and shows how bound
tightening can dramatically improve power network optimization. In par-
ticular, the paper shows that the Quadratic Convex relaxation of power
flows, enhanced by bound tightening, almost always outperforms the
state-of-the-art Semi-Definite Programming relaxation on the optimal
power flow problem.

Keywords: Continuous constraint networks · Minimal network · Bound

consistency · Convex relaxation · AC power flow · QC relaxation · AC
optimal power flow

1 Introduction
In (mixed-integer) nonlinear optimization, convexification is used to obtain dual
bounds, complementing primal heuristics. In many cases, these convex relax-
ations are parametrized by variable bounds and the tighter the bounds are, the
stronger the relaxations. There is thus a strong potential for synergies between
convex optimization and constraint programming. This paper explores these
synergies in the context of power system optimization.
The power industry has been undergoing a fundamental transformation in
recent years. Deregulation, the emergence of power markets, pressure for reduced
capital investment, and the need to secure a clean sustainable energy supply all
stress the importance of efficiency and reliability in the design and operation of
power networks. As a result, optimization has become a critical component of
NICTA—NICTA is funded by the Australian Government through the Department
of Communications and the Australian Research Council through the ICT Centre
of Excellence Program.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 39–57, 2015.
DOI: 10.1007/978-3-319-23219-5 4
40 C. Coffrin et al.

the emerging smart-grid [28] and has resulted in millions of dollars in annual
savings [32].
Power network applications range from long-term network design and invest-
ment tasks [7,12,21] to minute-by-minute operation tasks [14,16,17,19,23]. All
of these optimization problems share a common core, the Alternating Current
(AC) power flow equations, which model the steady-state physics of power flows.
These equations form a system of continuous non-convex nonlinear equations
that prove to be a significant challenge for existing general-purpose optimiza-
tion tools. It is thus not surprising that, in the last decade, significant attention
has been devoted to developing computationally efficient convex relaxations.
The main contribution of this paper is to show that constraint program-
ming can substantially improve the quality of convex relaxations for power flow
applications. To obtain this result, the paper defines the concept of constraint
relaxation networks and generalizes traditional consistency notions to these net-
works, including minimal network and bound consistency. These concepts, and
the associated algorithms, are then applied to optimal power flow applications
with and without load uncertainty. The experimental results demonstrate the
significant value of bound tightening for power flow applications. In particular,

1. Bound tightening reduces the domains of the variables by as much as 90%

in many cases.
2. In over 90% of the test cases considered, propagation over the convex relax-
ation was sufficient to close the optimality gap within 1%. Only 4 of the test
cases considered remain open.
3. The network consistency algorithm improves the quality of the Quadratic
Convex (QC) relaxation [18] considerably. The QC relaxation now outper-
forms, in the vast majority of the cases, the established state-of-the-art Semi-
Definite Programming (SDP) relaxation on the optimal power flow problem.
4. Parallelization can significantly reduce the runtime requirements of bound
tightening, making the proposed algorithms highly practical.

The rest of the paper is organized as follows. Section 2 reviews the AC power flow
feasibility problem and introduces the notations. Section 3 reviews the state-of-
the-art QC power flow relaxation, which is essential for building efficient consis-
tency algorithms. Section 4 formalizes the idea of constraint relaxation networks
and Section 5 applies this formalism to AC power flows. Section 6 studies the
quality of bound tightening in this application domain and Section 7 evalu-
ates the proposed methods on the ubiquitous AC Optimal Power Flow problem.
Section 8 illustrates the potential of the proposed methods on power flow appli-
cations incorporating uncertainty and Section 9 concludes the paper.

2 AC Power Flow
A power network is composed of a variety of components such as buses, lines,
generators, and loads. The network can be interpreted as a graph (N, E) where
the set of buses N represent the nodes and the set of lines E represent the edges.
 Strengthening Convex Relaxations with Bound Tightening 41

Note that E is a set of directed arcs and E R will be used to indicate those arcs in
the reverse direction. To break numerical symmetries in the model and to allow
easy comparison of solutions, a reference node r ∈ N is also specified.
Every node i ∈ N in the network has three properties, voltage Vi = vi ∠θi ,
power generation Sig = pgi + iqig , and power consumption Sid = pdi + iqid , all of
which are complex numbers due to the oscillating nature of AC power. Each line
(i, j) ∈ E has an admittance Yij = gij + ibij , also a complex number. These
network values are connected by two fundamental physical laws, Kirchhoff’s
Current Law (KCL),

Sig − Sid = Sij ∀ i ∈ N (1)
(i,j)∈E∪E R

and Ohm’s Law,

Sij = Yij∗ (Vi Vi∗ − Vi Vj∗ ) ∀ (i, j) ∈ E ∪ E R . (2)

Note that bold values indicate parameters that are constant in the classic AC
power flow problem and non-bold values are the decision variables.
In addition to these physical laws, the following operational constraints are
required in AC power flows. Generator output limitations on S g ,

Sigl ≤ Sig ≤ Sigu ∀i ∈ N . (3)

Line thermal limits on Sij ,

R
|Sij | ≤ su
ij ∀(i, j) ∈ E ∪ E . (4)

Bus voltage limits on Vi ,

vil ≤ |Vi | ≤ viu ∀i ∈ N (5)

and line phase angle difference limits on Vi Vj∗ ,

 
θij
Δl
≤ ∠ Vi Vj∗ ≤ θij
Δu
∀(i, j) ∈ E (6)

Note that power networks are designed and operated so that −π/3 ≤ θ Δl ≤
θ Δu ≤ π/3 [22] and values as low as π/18 are common in practice [33]. Addi-
tionally the values of v l , v u , su must be positive as they are bounds on the
magnitudes of complex numbers.
Combining all of these constraints and expanding them into their real-number
representation yields the AC Power Flow Feasibility Problem (AC-PF) presented
in Model 1. The input data is indicated by bold values and a description of
the decision variables is given in the model. Constraint (7a) sets the reference
angle. Constraints (7b)–(7c) capture KCL and constraints (7d)–(7e) capture
Ohm’s Law. Constraints (7f) link the phase angle differences on the lines to the
bus variables and constraints (7g) enforce the thermal limit on the lines. This
particular formulation of AC-PF is advantageous as the auxiliary variables θΔ , p,
42 C. Coffrin et al.

Model 1. The AC Power Flow Feasibility Problem (AC-PF)

variables:
pgi ∈ (pgl
i , pi ) ∀i ∈ N - active power generation
gu

qig ∈ (qigl , qigu ) ∀i ∈ N - reactive power generation

vi ∈ (vil , viu ) ∀i ∈ N - bus voltage magnitude
θi ∈ (−∞, ∞) ∀i ∈ N - bus voltage angle
Δ
θij Δl
∈ (θij Δu
, θij ) ∀(i, j) ∈ E - angle difference on a line (aux.)
R
pij ∈ (−su
ij , sij ) ∀(i, j) ∈ E ∪ E
u
- active power flow on a line (aux.)
R
qij ∈ (−su u
ij , sij ) ∀(i, j) ∈ E ∪ E - reactive power flow on a line (aux.)
subject to:
θr = 0 (7a)

pgi − pdi = pij ∀i ∈ N (7b)
(i,j)∈E∪E R

qig − qid = qij ∀i ∈ N (7c)
(i,j)∈E∪E R

pij = gij vi2 − gij vi vj cos(θij

Δ Δ
) − bij vi vj sin(θij ) (i, j) ∈ E ∪ E R (7d)
qij = −bij vi2 + Δ
bij vi vj cos(θij ) − Δ
gij vi vj sin(θij ) (i, j) ∈ E ∪ E R
(7e)
Δ
θij = θi − θj ∀(i, j) ∈ E (7f)
p2ij + 2
qij ≤ (su
ij )
2
∀(i, j) ∈ E ∪ E R
(7g)

and q isolate the problem’s non-convexities in constraints (7d)–(7e) and enable

all but one of the operational constraints to be captured by the variable bounds.
This continuous constraint satisfaction problem is NP-Hard in general [24,40]
and forms a core sub-problem that underpins a wide variety of power network
optimization tasks.
To address the computational difficulties of AC-PF, convex relaxations (i.e.
polynomial time) have attracted significant interest in recent years. Such relax-
ations include the Semi-Definite Programming (SDP) [2], Second-Order Cone
(SOC) [20], Convex-DistFlow (CDF) [13], and the recent Quadratic Convex (QC)
[18] relaxations. To further improve these relaxations, this paper proposes con-
sistency notions and associated propagation algorithms for AC power flows. A
detailed evaluation on 57 AC transmission system test cases demonstrates that
the propagation algorithms enable reliable and efficient methods for improving
these relaxations on a wide variety of power network optimization tasks via
industrial-strength convex optimization solvers (e.g., Gurobi, Cplex, Mosek).
The next section reviews the QC relaxation in detail, which forms the core of
the proposed propagation algorithms.
 Strengthening Convex Relaxations with Bound Tightening 43

3 The Quadratic Convex (QC) Relaxation

The QC relaxation [18] was introduced to utilize the bounds on the voltage vari-
ables v and θΔ , which are ignored by the other relaxations. The key idea is to use
the variable bounds to derive convex envelopes around the non-convex aspects of
the AC-PF problem. The derivation begins by lifting the voltage product terms
in to the higher dimensional W-space using the following equalities:

wi = vi2 i ∈ N (8a)
R Δ
wij = vi vj cos(θij ) ∀(i, j) ∈ E (8b)
I Δ
wij = vi vj sin(θij ) ∀(i, j) ∈ E (8c)

When Model 1 is lifted into this W-space, all of the remaining constraints are
convex. On its own, this lifted model is a weak relaxation but the QC relaxation
strengthens it by developing convex relaxations of the nonlinear equations (8a)–
(8c) for the operational bounds on variables v and θΔ . The convex envelopes for
the square and bilinear functions are well-known [27], i.e.,

2 T
x̌ ≥ x2
x  ≡ (T-CONV)
x̌ ≤ (xu + xl )x − xu xl
⎧
⎪
⎪ ˇ ≥ xl y + y l x − xl y l
xy
⎪
⎪
⎨xy
ˇ ≥ xu y + y u x − x u y u
xyM ≡ (M-CONV)
⎪
⎪ ˇ ≤ xl y + y u x − x l y u
xy
⎪
⎪
⎩
ˇ ≤ xu y + y l x − x u y l
xy
Under the assumption that the phase angle difference bound is within −π/2 ≤
θ Δl ≤ θ Δu ≤ π/2, relaxations for sine and cosine are given by

ˇ ≤ cos(x)
cx
cos(x)C ≡ cos(xl )−cos(xu )
(C-CONV)
ˇ ≥
cx (xl −xu )
(x − xl ) + cos(xl )
⎧
m

m
⎪
⎪sxˇ ≤ cos x
x − xm
+ sin x
if xl < 0 ∧ xu > 0
⎪
⎪ 2 2 2
⎪
⎪
m

m
⎪
⎪ m
⎪
⎪ ˇ ≥ cos 2
sx x
x + 2 − sin 2
x x
if xl < 0 ∧ xu > 0
⎪
⎪
⎪
⎨sxˇ ≤ sin(x) if xl ≥ 0
sin(x)S ≡ (S-CONV)
⎪
⎪
l
)−sin(xu )
⎪
⎪ ˇ ≥ sin(x
sx (x − xl ) + sin(xl ) if xl ≥ 0
⎪
⎪
(xl −xu )
⎪
⎪
⎪
l
)−sin(xu )
⎪
⎪ ˇ ≤ sin(x
sx l −xu ) (x − xl ) + sin(xl ) if xu ≤ 0
⎪
⎪
(x
⎩
ˇ ≥ sin(x)
sx if xu ≤ 0
where xm = max(−xl , xu ). These are a generalization of the relaxations pro-
posed in [18] to support asymmetrical bounds on x. Utilizing these building
44 C. Coffrin et al.

Model 2. The QC Power Flow Feasibility Problem (QC-PF)

variables: Variables of Model 1
stij ∈ (−1, 1) ∀(i, j) ∈ E - relaxation of the sine (aux.)
ctij ∈ (0, 1) ∀(i, j) ∈ E - relaxation of the cosine (aux.)
vvij ∈ (vil vjl , viu vju ) ∀(i, j) ∈ E - relaxation of the voltage product (aux.)


wi ∈ (vil )2 , (viu )2 ∀i ∈ N - relaxation of the voltage square (aux.)
R
wij ∈ (0, ∞) ∀(i, j) ∈ E - relaxation of the voltage and cosine product (aux.)
I
wij ∈ (−∞, ∞) ∀(i, j) ∈ E - relaxation of the voltage and sine product (aux.)
subject to: (7a)–(7c),(7f)–(7g)
CONV(wi = vi2 ∈ (vil , viu )) ∀i ∈ N (10a)
Δ
CONV(ctij = cos(θij ) ∈ Δl
(θij , θij
Δu
)) ∀(i, j) ∈ E (10b)
Δ
CONV(stij = sin(θij ) ∈ Δl
(θij , θij
Δu
)) ∀(i, j) ∈ E (10c)
CONV(vvij = vi vj ∈ (vil , viu ) × (vjl , vju )) ∀(i, j) ∈ E (10d)
R
CONV(wij = vvij ctij ∈ l
(vvij u
, vvij ) × (ctlij , ctu
ij ) ∀(i, j) ∈ E (10e)
I
CONV(wij = vvij stij ∈ l
(vvij u
, vvij ) × (stlij , stu
ij ) ∀(i, j) ∈ E (10f)
R 2 I 2
(wij ) + (wij ) ≤ wi wj ∀(i, j) ∈ E (10g)
R I
pij = gij wi − gij wij − bij wij ∀(i, j) ∈ E (10h)
R I
qij = −bij wi + bij wij − gij wij ∀(i, j) ∈ E (10i)
R I
pji = gij wj − gij wij + bij wij ∀(i, j) ∈ E (10j)
R I
qji = −bij wj + bij wij + gij wij ∀(i, j) ∈ E (10k)

blocks, convex relaxations for equations (8a)–(8c) can be obtained by composing

R
relaxations of the subexpressions, for example, wij ≡ vi vj M cos(θi − θj )C M .
Lastly, the QC relaxation proposes to strengthen these convex relaxations with
a valid second-order cone constraint [11,18,20],
R 2 I 2
(wij ) + (wij ) ≤ wi wj ∀(i, j) ∈ E (9)

The complete QC relaxation of the AC-PF problem is presented in Model

2 (QC-PF), which incorporates many of the components of Model 1. In the
model, the constraint CONV(y = f (x) ∈ D) is used to indicate that y lies in a
convex relaxation of function f within the domain D. Constraints (10a)–(10f)
implement the convex relaxations and constraints (10g) further strengthen these
relaxations. Constraints (10h)–(10k) capture the line power flow in terms of the
W-space variables.

The Impact of Tight Bounds in the QC Relaxation: Constraints (10a)–(10f)

in Model 2 highlight the critical role that the bounds on v and θΔ play in
 Strengthening Convex Relaxations with Bound Tightening 45

1.0

1.0

0.8
0.5

0.5

0.4
Cosine
0.0

0.0

0.0
0.8
−0.5

−0.5
Sine

0.4
Variable Bound
Constraint
−1.0

−1.0
Feasible Region Sine Cosine

0.0
−1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0

Domain (radians) Domain (radians) Domain (radians)

Fig. 1. The Impact of Variable Bounds on the Convex Relaxations.

the strength of the QC relaxation. Figure 1 illustrates this point by showing

the convex relaxations for sine and cosine over the domains θΔ ∈ (−π/3, π/3)
and θΔ ∈ (−π/3, 0). This figure indicates two key points: (1) Although the
reduction in the size of the bound is 50% in this case, the area inside of the
convex relaxations have been reduced even more significantly; (2) Both the sine
and cosine functions are monotonic when the sign of θΔ is known, which produces
tight convex relaxations.

4 Consistency of Constraint Relaxation Networks

As discussed in Section 2, the core of many applications in power network opti-
mization is a continuous constraint network. Moreover, the QC relaxation of
this continuous constraint network depends on the bounds of the variables. As
a result, constraint propagation now has two benefits: On one hand, it reduces
the domains of the variables while, on the other hand, it strengthens the relax-
ation. These two processes reinforce each other, since tighter constraints generate
tighter bounds creating a virtuous cycle. This section generalizes traditional con-
sistency notions to this new context through the concepts of constraint schemes
and constraint relaxations. Since solutions to continuous constraint networks
are real numbers and computer implementations typically rely on floating-point
numbers, some care must be exercised in formalizing these notions. The formal-
ization also assumes that only bound reasoning is of interest, since these are
continuous constraint networks. However, the concepts generalize naturally to
domain reasoning.

Continuous Constraint Networks: Constraint networks are defined in terms of

a set of variables X = {x1 , . . . , xn } ranging over intervals I = {I1 , . . . , In }
and a set of constraints. An interval I = [l, u] denotes the set of real numbers
{r ∈  | l ≤ r ≤ u}. This paper only considers floating-point intervals, i.e.,
intervals whose bounds are floating-point numbers. If r is a real number, [r]
denotes the smallest floating-point interval containing r, [r]− the largest floating-
point number no greater than r, and [r]+ the smallest floating-point number no
smaller than r. A variable assignment assigns to each variable xi a value from its
46 C. Coffrin et al.

interval Ii . A constraint is a function (X → ) → Bool which, given a variable

assignment, returns a truth value denoting whether the assignment satisfies the
constraint.

Definition 1 (Continuous Constraint Network (CCN)). A continuous

constraint network is a triple (X, I, C) where X = (x1 , . . . , xn ) is a collection of
variables ranging over I = (I1 , . . . , In ) and C is a set of constraints.

Definition 2 (Solution to a CCN). A solution to a CCN (X, I, C), where

X = (x1 , . . . , xn ) and I = (I1 , . . . , In ), is an assignment σ = {x1 ← v1 ; . . . ; xn ←
vn } such that vi ∈ Ii and for all c ∈ C: c(σ) holds. The set of solutions to a
CCN P is denoted by Σ(P).

In the following we use max(x, Σ) to denote the maximum value of variable x

in the assignments Σ, i.e., max(x, Σ) = maxσ∈Σ σ(x), where σ(x) denotes the
value of variable x in assignment σ. The value min(x, Σ) is defined similarly.
The following definition adapts the traditional concept of minimal constraint
network [31] to continuous constraint networks.

Definition 3 (Minimal CCN). A CCN P = (X, I, C), where X =

(x1 , . . . , xn ) and I = (I1 , . . . , In ), is minimal if, for each variable xi , the interval
Ii = [li , ui ] satisfies li = [min(xi , Σ(P))]− ∧ ui = [max(xi , Σ(P))]+ .

Note that the bounds are not necessarily solutions themselves but are as tight
as the floating-point accuracy allows for. Given a CCN P = (X, I, C), its largest
minimal network P  = (X , I , C) (Im ⊆ I) always exists and is unique since
there are only finitely many floating-point intervals.
The concept of bound consistency [39] captures a relaxation of the mini-
mal network: It only requires the variable bounds to be tight locally for each
constraint.

Definition 4 (Bound Consistency for CCNs). A CCN P = (X, I, C),

where X = (x1 , . . . , xn ) and I = (I1 , . . . , In ), is bound-consistent if each con-
straint c is bound-consistent with respect to I. A constraint c is bound-consistent
with respect to I if the continuous constraint network (X, I, {c}) is minimal.

Once again, given a CCN P = (X, I, C), its largest bound-consistent network
P = (X, Im , C) (Im ⊆ I) always exists and is unique. In the following, we use
minCCN (X, I, C) and bcCCN (X, I, C) to denote these networks, i.e.,

minCCN(X, I, C) = max{Im ⊆ I | (X, Im , C) is minimal},

bcCCN(X, I, C) = max{Im ⊆ I | (X, Im , C) is bound-consistent}.

Constraint Relaxation Networks: The convex relaxations used in the QC relax-

ation depend on the variable bounds, i.e., the stronger the bounds the stronger
the relaxations. Since the relaxations change over time, it is necessary to intro-
duce new consistency notions: constraint schemes and constraint relaxations.
 Strengthening Convex Relaxations with Bound Tightening 47

Definition 5 (Continuous Constraint Scheme). A constraint scheme r is a

function I → (X → ) → Bool which, given a collection of intervals, returns a
constraint. Moreover, the scheme r satisfies the following monotonicity property:
I ⊆ I  ⇒ (r(I  )(σ) ⇒ r(I)(σ))
for all collections of intervals I and I  , and variable assignment σ.
The monotonicity property ensures that tighter bounds produce tighter con-
straints. Traditional constraints are constraint schemes that just ignore the ini-
tial bounds. A constraint relaxation is a constraint scheme that preserves the
solutions to the original constraint.
Definition 6 (Constraint Relaxation). A constraint scheme r is a relaxation
of constraint c if, for all assignment σ and bounds I = ([σ(x1 )], . . . , [σ(xn )]), we
have r(I)(σ) ⇒ c(σ).1
Example 1. Consider the constraint c(x, y, z) which holds if z = xy. Given
bounds [xl , xu ] and [y l , y u ] for variables x and y, the McCormick relaxation
[27] is a constraint scheme specified by the collection of constraints in M-CONV.
Note that this envelope ignores the bound on variable z. Additionally this con-
straint scheme is also a constraint relaxation of c because it is known to be the
convex envelope of z = xy for any bounds on x and y [27].
Definition 7 (Continuous Constraint Relaxation Network (CCRN)).
A constraint relaxation network is a triple (X, I, R) where X is a collection of
variables ranging over I and R is a set of constraint relaxations.
In the following, we use R(I) to denote {r(I) | r ∈ R} if R is a set of relaxations.

Consistency of Constraint Relaxation Networks: We now generalize the con-

cepts of minimal and bound-consistent networks to CCRNs. The definitions cap-
ture the fact that no additional bound tightening is possible for the relaxations
induced by the bounds.
Definition 8 (Minimal CCRN). A CCRN P = (X, I, R), X = (x1 , . . . , xn )
and I = (I1 , . . . , In ), is minimal if the CCN network (X, I, R(I)) is.
Definition 9 (Bound-Consistent CCRN). A CCRN P = (X, I, R), where
X = (x1 , . . . , xn ) and I = (I1 , . . . , In ), is bound-consistent if the CCN network
(X, I, R(I)) is.
Once again, the largest minimal or bound-consistent network of a CCRN exists
and is unique by monotonicity of constraint relaxations. In the following, we use
minCCRN (X, I, C) and bcCCRN (X, I, C) to denote these networks, i.e.,
minCCRN(X, I, R) = max{Im ⊆ I | (X, Im , R) is minimal},
bcCCRN(X, I, R) = max{Im ⊆ I | (X, Im , R) is bound-consistent}.
1
Note that some of the convex relaxations used in the QC relaxation are only valid
within some bounds. This is easily captured by assuming that the constraint itself
imposes these bounds.
48 C. Coffrin et al.

minCCRN(X, I, R) bcCCN(X, I, C)
I n := I; I n := I;
repeat repeat
I o := I n ; I o := I n ;
I n := minCCN(X, I o , R(I o )); for all c ∈ C
until I o = I n Icn 
:= minCCN(X, I o , {c});
return I n ; I n := c∈C Icn ;
until I o = I n
return I n ;
Fig. 2. Computing the Minimal Fig. 3. Computing the Largest
Continuous Constraint Relaxation Bound-Consistent Constraint Net-
Networks work.

The following property establishes the soundness of bound tightenings in

CCRNs.

Proposition 1. Let (X, I, C) be a CCN and let (X, I, R) be a CCRN such that
R = {r | c ∈ C ∧ r is a relaxation of c}. Then,

minCCN(X, I, C) ⊆ minCCRN(X, I, R),

bcCCN(X, I, C) ⊆ bcCCRN(X, I, R).

The minimal and bound-consistent relaxation networks can be computed by

a simple fixpoint algorithm that iterates the consistency algorithm over the
increasingly tighter relaxation networks. Figure 2 depicts the algorithm for
computing a minimal network. The algorithm is similar for bound consistency.
Observe that the bound-consistency algorithm has a fixpoint algorithm embed-
ded inside the top-level fixpoint.

4.1 Relation to Concepts in Global Optimization

The idea of bounds propagation for global optimization goes as far back as
[6]: It was subsequently implemented in the Numerica system which also per-
forms bound propagation on a linearization of the nonlinear constraints [37,38].
The notion of using bound reductions for improving convex relaxations of non-
convex programs was first widely recognized in the Branch-and-Reduce (BnR)
algorithm [34]. BnR is a natural extension of Branch-and-Bound over continuous
domains, which includes additional steps to reduce the domains of the variables
at each search node. This line of work has developed into two core bound reduc-
tion ideas: (1) Feasibility-Based Range Reduction (FBRR), which is concerned
with pruning techniques based on feasibility information and (2) Optimality
Based Range Reduction (OBRR), which develops bound reductions based on
Lagrangian-duality arguments [35]. A variety of methods have been developed
for FBRR and OBRR with various pruning strength and computational time
tradeoffs [5,25,34]. However, all these methods are non-global bound reduction
techniques and may be iterated until a desired level of consistency is achieved.
 Strengthening Convex Relaxations with Bound Tightening 49

CCRNs and the associated consistency notions (minCCRN, bcCCRN) devel-

oped herein are examples of FBRR methods. The idea of computing minCCRN
is discussed informally in [4] for the special case where the relaxation is a system
of linear or convex equations (note that the algorithm in Figure 2 applies for any
kind of CSP). It is often noted in the FBRR literature that just one iteration of
the minCCRN is too costly to compute [4,35], let alone the full fixpoint. The pre-
ferred approach is to perform some bound propagation (not always to the fixpoint)
on linear relaxations of the non-convex problem [5,25,35]. In fact, specialized algo-
rithms have been proposed for computing bound consistency on purely linear sys-
tems for this purpose [4]. The linear bound-consistency computations discussed in
[4,5] are weaker forms of the bcCCRN notion considered here since it does not
explicitly mention re-linearizing the relaxation after bound propagation is com-
plete and re-computing bounds consistency. It is important to note that the algo-
rithm in Figure 2 seamlessly hybridizes the FBRR ideas from global optimization to
CP systems, which include arbitrary global constraints that are outside the scope
of purely mathematical programs. This advantage is utilized in the next section.

5 Constraint Relaxation Networks for Power Flows

This section discusses how to compute the largest minimal and bound-consistent
networks for the relaxation model (X, I, R) defined by Model 2. Observe first
that the convex relaxations used in Model 2 are all monotonic.
Proposition 2. The convex relaxations T-CONV, M-CONV, C-CONV, and S-
CONV are monotonic.

Minimal Network: The largest minimal network is computed by Algorithm QC-N

which applies the fixpoint algorithm minCCRN shown in Figure 2 to Model 2.
The underlying minCCN networks are computed by optimizing each variable
independently, i.e.,
Ixn := [ min σ(x), max σ(x)];
σ:R(σ) σ:R(σ)

Observe that this computation is inherently parallel, since all the optimizations
are independent.

Bound-Consistent Network: The largest bound-consistent network is computed

by Algorithm QC-B which applies the bound-consistency counterpart of algo-
rithm minCCRN to Model 2. The bcCCN networks needed in this fixpoint
algorithm are computed by the algorithm shown in Figure 3. Algorithm bcCCN
computes the intervals Icn that are bound-consistent for each constraint c ∈ C
before taking the intersection of these intervals. The process is iterated until a
fixpoint is obtained. This algorithm was selected because the bound-consistency
computations can be performed in parallel.
Observe also that the QC-B algorithm is applied to a version of Model 2
using a global constraint
line power qc(pij , qij , vi , θi , pji , qji , vj , θj )
50 C. Coffrin et al.

Fig. 4. QC Consistency Algorithms – Quality Analysis.

Fig. 5. QC Consistency Algorithms – Runtime Analysis.

that captures constraints (7f)–(7g), (10a)–(10k) for each line (i, j) ∈ E. The
use of this global constraint means that QC-B computes a stronger form of
bounds consistency than one based purely on Model 2. This stronger level of
consistency is necessary to obtain reasonable bound tightenings. Note that all the
optimizations in algorithms QC-N and QC-B are convex optimization problems
which can be solved in polynomial time.

6 Strength and Performance of the Bound Tightening

This section evaluates the benefits of QC-N and QC-B on the general feasibility
problem in Model 1. Algorithms QC-N and QC-B were implemented in AMPL
[15] using IPOPT 3.12 [41] to solve the convex nonlinear programs. The prop-
agation algorithms were executed on Dell PowerEdge R415 servers with Dual
2.8GHz AMD 6-Core Opteron 4184 CPUs and 64GB of memory with a conver-
gence tolerance of  = 0.001. Their performance is evaluated on 57 transmission
system test cases from the NESTA 0.3.0 archive [10] ranging from 3 to 300 buses.
Figure 4 summarizes the results of QC-B and QC-N on three key metrics: the
phase angle difference domains (θΔ ), the voltage domains (v), and the number of
lines where the sign of θΔ is determined. Each plot summarizes the distribution
of 57 values as a standard box-and-whisker plot, where the width of the box
reflects the first and third quartiles, the black line inside the box is the median,
and the whiskers reflect min and max values up to 1.5 IQR with the remaining
data points plotted as outliers. In these plots values to the left are preferable.
The domain reduction of the QC-N approach is substantial, typically pruning
the domain θΔ by 90% and the domain of v by 30% and determining the sign
of θΔ for about half of the lines. Across all of the metrics, it is clear that QC-N
has significant benefits over QC-B.
Figure 5 summarizes the runtime performance of QC-B and QC-N on three
key metrics: Total CPU time (T1 ), fully parallel CPU wall-clock time (T∞ ), and
 Strengthening Convex Relaxations with Bound Tightening 51

the number of fixpoint iterations. The total runtimes of the QC-B and QC-N
algorithms vary widely based on the size of the network under consideration
and can range from seconds to hours. Fortunately, regardless of the size of the
network, the number of iterations in the fixpoint computation is small (often
less than 10). As a result, the parallel runtime of the algorithms scale well with
the size of the network and rarely exceeds 1 minute, which is well within the
runtime requirements of the majority of network optimization applications.2

7 Application to AC Optimal Power Flow

This section assesses the benefits of QC-B and QC-N on the ubiquitous AC
Optimal Power Flow problem (AC-OPF) [29,30]. The goal of the AC-OPF is
to find the cheapest way to satisfy the loads given the network flow constraints
and generator costs functions, which are typically quadratic. If c2i , c1i , c0i are
the cost coefficients for generating power at bus i ∈ N , the AC-OPF objective
function is given by,

minimize: c2i (pgi )2 + c1i (pgi ) + c0i (11)
i∈N

The complete non-convex AC-OPF problem is Model 1 with objective (11) and
the QC relaxation of this problem is Model 2 with objective (11).
The goal is to compare five AC-OPF relaxations for bounding primal AC-
OPF solutions produced by IPOPT, which only guarantees local optimality. The
five relaxations under consideration are as follows:
1. QC - as defined in Model 2.
2. QC-B - bcCCRN for Model 2.
3. QC-N - minCCRN for Model 2.
4. SDP - a state-of-the-art relaxation based on semi-definite programming [26].
5. SDP-N - the SDP relaxation strengthened with bounds from QC-N.
There is no need to consider other existing relaxations as the QC and SDP
dominate them [11]. The computational environment and test cases are those of
Section 6. SDPT3 4.0 [36] was used to solve the SDP models.
Table 1 presents the detailed performance and runtime results on all 57 test
cases. They can be summarized as follows: (1) The optimality gaps of the QC
relaxation are significantly reduced by both QC-N and QC-B; (2) QC-N closes
the AC-OPF optimality gap to below 1% in 90% of the cases considered and
closes 10 open test cases; (3) QC-N almost always outperforms the SDP relax-
ation in quality with comparable parallel runtimes; (4) For the test cases with
significant optimality gaps, QC-N outperforms the SDP relaxation most often,
even when the SDP relaxation is strengthened with QC-N bounds (i.e., SDP-N).
2
Dedicated high performance computational resources are commonplace in power
system operation centers. The T∞ runtime is realistic in these settings where high-
level of reliability is critical.
52 C. Coffrin et al.

Table 1. Quality and Runtime Results of Convex Relaxations on the AC-OPF Problem
(bold - best in row (runtime used to break ties in quality), — - solving error)

$/h Optimality Gap (%) T∞ Runtime (sec.)

Test Case AC SDP-N SDP QC-N QC-B QC AC SDP-N SDP QC-N QC-B QC
case3 lmbd 5812 0.1 0.4 0.1 1.0 1.2 0.2 6.8 4.7 0.5 0.4 0.1
case4 gs 156 0.0 0.0 0.0 0.0 0.0 0.2 7.2 4.8 0.4 0.8 0.1
case5 pjm 17551 5.2 5.2 9.3 14.5 14.5 0.1 6.4 5.1 0.9 0.3 0.2
case6 c 23 0.0 0.0 0.0 0.3 0.3 0.0 6.9 5.4 1.3 0.4 0.1
case6 ww 3143 0.0 0.0 0.0 0.1 0.6 0.3 5.4 5.4 0.8 2.7 0.1
case9 wscc 5296 0.0 0.0 0.0 0.0 0.0 0.2 6.2 4.9 1.5 0.7 0.1
case14 ieee 244 0.0 0.0 0.0 0.1 0.1 0.1 4.8 5.2 2.0 0.4 0.1
case24 ieee rts 63352 0.0 0.0 0.0 0.0 0.0 0.2 8.5 6.0 3.2 0.5 0.2
case29 edin 29895 0.0 0.0 0.0 0.1 0.1 0.4 8.2 7.8 15.8 1.4 1.1
case30 as 803 0.0 0.0 0.0 0.1 0.1 0.3 6.9 5.4 2.3 0.5 0.1
case30 fsr 575 0.0 0.0 0.1 0.3 0.4 0.2 5.5 6.1 2.2 1.0 0.2
case30 ieee 205 0.0 0.0 0.0 5.3 15.4 0.4 7.8 6.3 0.7 0.7 0.3
case39 epri 96505 0.0 0.0 0.0 0.0 0.0 0.2 6.5 7.1 2.1 0.6 0.2
case57 ieee 1143 0.0 0.0 0.0 0.1 0.1 0.1 11.4 9.1 5.1 0.9 0.4
case73 ieee rts 189764 0.0 0.0 0.0 0.0 0.0 0.5 12.5 8.5 4.7 0.7 0.5
case118 ieee 3720 0.1 0.1 0.4 1.0 1.7 0.3 18.2 12.0 21.1 6.0 0.8
case162 ieee dtc 4237 1.0 1.1 0.7 3.8 4.2 0.7 57.6 34.2 25.9 7.0 1.5
case189 edin 849 — 0.1 0.1 — 0.2 0.9 12.3 13.3 6.5 59.9 1.6
case300 ieee 16894 0.1 0.1 0.1 1.0 1.2 0.9 40.8 25.5 48.2 14.4 2.4
case3 lmbd api 367 0.0 1.3 0.0 0.5 1.8 0.2 4.0 4.0 0.5 1.5 0.1
case4 gs api 767 0.0 0.0 0.0 0.2 0.7 0.9 6.9 3.9 0.8 0.3 0.1
case5 pjm api 2994 0.0 0.0 0.0 0.4 0.4 0.0 6.9 7.0 0.2 0.4 0.1
case6 c api 807 0.0 0.0 0.0 0.5 0.5 0.6 5.3 5.4 0.3 0.4 0.1
case6 ww api 273 — 0.0 0.0 2.1 13.1 0.2 4.5 15.0 0.4 0.4 0.1
case9 wscc api 656 0.0 0.0 0.0 0.0 0.0 0.4 5.4 6.4 0.8 0.9 0.1
case14 ieee api 323 0.0 0.0 0.2 1.3 1.3 0.1 6.4 4.7 0.5 0.4 0.1
case24 ieee rts api 6421 0.7 1.4 0.3 3.3 13.8 0.2 8.6 7.2 1.4 1.6 0.2
case29 edin api 295764 — — 0.1 0.4 0.4 0.3 12.6 7.8 28.4 1.1 3.2
case30 as api 571 0.0 0.0 0.0 2.4 4.8 0.4 7.6 6.0 3.7 0.8 0.2
case30 fsr api 372 3.6 11.1 2.7 42.8 46.0 0.2 7.9 6.7 1.4 0.4 0.2
case30 ieee api 411 0.0 0.0 0.0 0.9 1.0 0.3 8.9 6.5 1.0 0.5 0.2
case39 epri api 7466 0.0 0.0 0.0 0.8 3.0 0.1 9.2 6.5 4.9 1.9 0.2
case57 ieee api 1430 0.0 0.1 0.0 0.2 0.2 0.4 8.8 8.1 3.2 0.6 0.4
case73 ieee rts api 20123 0.9 4.3 0.1 3.6 12.0 0.6 15.4 9.5 11.4 2.0 0.6
case118 ieee api 10258 16.7 31.5 11.8 38.9 44.0 0.6 14.2 14.6 11.3 4.7 0.8
case162 ieee dtc api 6095 0.6 1.0 0.1 1.4 1.5 0.4 51.9 32.8 25.5 2.1 1.5
case189 edin api 1971 — 0.1 0.0 — 5.6 0.3 14.3 13.5 8.3 67.1 1.1
case300 ieee api 22825 0.0 0.0 0.2 0.6 0.8 0.9 47.5 28.5 71.1 3.6 2.6
case3 lmbd sad 5992 0.1 2.1 0.0 0.2 1.2 0.1 5.1 4.2 0.2 0.9 0.1
case4 gs sad 324 0.0 0.0 0.0 0.5 0.8 0.1 4.4 3.9 0.1 1.3 0.1
case5 pjm sad 26423 0.0 0.0 0.0 0.7 1.1 0.1 5.7 5.3 0.2 0.4 0.1
case6 c sad 24 0.0 0.0 0.0 0.4 0.4 0.1 6.7 4.6 0.2 0.3 0.1
case6 ww sad 3149 0.0 0.0 0.0 0.1 0.3 0.1 5.9 5.4 0.2 0.2 0.1
case9 wscc sad 5590 0.0 0.0 0.0 0.2 0.4 0.3 5.5 4.4 0.1 0.5 0.1
case14 ieee sad 244 0.0 0.0 0.0 0.1 0.1 0.1 7.5 4.6 0.5 0.3 0.1
case24 ieee rts sad 79804 1.4 6.1 0.1 3.4 3.9 0.3 9.3 5.7 0.6 0.4 0.3
case29 edin sad 46933 5.8 28.4 0.9 20.0 20.6 0.5 7.1 8.5 15.5 0.3 1.6
case30 as sad 914 0.1 0.5 0.0 2.9 3.1 0.1 6.4 6.8 2.3 0.3 0.2
case30 fsr sad 577 0.1 0.1 0.1 0.5 0.6 0.1 6.2 6.8 1.9 0.3 0.2
case30 ieee sad 205 0.0 0.0 0.0 2.0 4.0 0.3 7.0 6.0 0.6 0.6 0.1
case39 epri sad 97219 0.0 0.1 0.0 0.0 0.0 0.1 7.1 6.0 1.0 0.9 0.2
case57 ieee sad 1143 0.0 0.0 0.0 0.1 0.1 0.4 8.8 7.6 1.9 0.8 0.3
case73 ieee rts sad 235241 2.4 4.1 0.1 3.1 3.5 0.3 9.7 8.4 3.6 0.6 0.8
case118 ieee sad 4323 4.0 7.6 1.4 7.6 8.3 0.4 15.4 13.8 5.9 0.6 1.0
case162 ieee dtc sad 4368 1.7 3.6 0.4 5.9 6.9 0.9 46.8 37.7 27.3 2.1 1.4
case189 edin sad 914 — 1.2 0.5 — 2.2 0.6 11.4 17.4 12.2 49.6 1.1
case300 ieee sad 16912 0.1 0.1 0.1 0.8 1.2 0.9 25.2 30.8 45.6 5.5 2.4
 Strengthening Convex Relaxations with Bound Tightening 53

Overall, these results clearly establish QC-N is the new state-of-the-art con-
vex relaxation of the AC-OPF. General purpose global optimization solvers (e.g.,
Couenne 0.4 [3] and SCIP 3.1.1 [1,8]) were also considered for comparison. Pre-
liminary results indicated that these general purpose solvers are much slower
than the dedicated power flow relaxations considered here and cannot produce
competitive lower bounds on these networks with in 10 hours of computation.

8 Propagation with Load Uncertainty

Loads in power systems are highly predictable. In transmission systems, it is
commonplace for minute-by-minute load forecasts to be within 5% of the true
values [9]. This high degree of predictability can be utilized by the bound tight-
ening algorithms proposed here. Indeed, if the feasible set of Model 1 is increased
to include a range of possible load values, determined by the forecast, then the
algorithms compute a description of all possible future power flows. This section
studies the power of bound propagation in this setting.
Model 3 presents an extension of Model 1 to incorporate load uncertainty.
New decision variables for the possible load values are introduced (i.e., pd , q d )
and their bounds come from the extreme values of the load forecasting model.
The lower bounds on active power generation (pg ) are also increased to include
0, as generators may become inactive at some point in the future (e.g., due to
scheduled maintenance or market operations). Constraints (12a)–(12b) incorpo-
rate the load variables into KCL. The other constraints remain the same as in
Model 1. Because only the KCL constraints are modified in this formulation,
the QC relaxation of Model 3 (QC-U) is similar to Model 1, as described in
Section 3. For the experimental evaluation, the 57 deterministic test cases were
extended into uncertain load cases by adopting a forecast model of ±5% of the
deterministic load value.
Figure 6 compares the quality of minCCRN on the QC-U model (QC-U-N)
to minCCRN in the deterministic case (QC-N) in order to illustrate the pruning

Model 3. The AC-PF Program with Load Uncertainty (AC-PF-U)

variables: Variables of Model 1

pdi ∈ (pdl du
i , pi ) ∀i ∈ N - active power load interval

qid ∈ (qidl , qidu ) ∀i ∈ N - reactive power load interval

pgi ∈ (0, pdu
i ) ∀i ∈ N - active power generation interval

subject to: (7a), (7d)–(7g)


pgi − pdi = pij ∀i ∈ N (12a)
(i,j)∈E∪E R

qig − qid = qij ∀i ∈ N (12b)
(i,j)∈E∪E R
54 C. Coffrin et al.

QC−N

QC−U−N

0 20 40 60 80 100
Angle Diff. Sign Unknown (% of lines)

Fig. 6. QC Consistency Algorithms with Load Uncertainty – Quality Analysis.

Fig. 7. Comparison of AC-OPF Bound Improvements of QC Variants.

loss due to uncertainty. The results indicate that, even when load uncertainty is
incorporated, minCCRN still prunes the variable domains significantly, typically
reducing the voltage angle domains by 80% and the voltage magnitude domains
by 10%, and determining the sign of θΔ for about 30% of the lines. The domain
reduction on θΔ in QC-U-N is particularly significant.
Figure 7 considers the AC-OPF and summarizes the optimality gaps pro-
duced under load certainty and uncertainty. QC-U-N produces significant
improvement in optimality gaps, moving from < 5% (QC) to less than < 1%.
Obviously, load certainty (QC-N) closes the remaining 1%.

9 Conclusion
This paper studied how bound tightening can improve convex relaxations by
adapting traditional constraint-programming concepts (e.g., minimal network
and bound consistency) to a relaxation framework. It showed that, on power
flow applications, bound tightening over the QC relaxation can dramatically
reduce variable domains. Moreover, on the ubiquitous AC-OPF problem, the QC
relaxation, enhanced by bound tightening, almost always outperforms the state-
of-the-art SDP relaxation on the optimal power flow problem. The paper also
showed that bound tightening yields significant benefits under load uncertainty,
demonstrating a breadth of applicability. These results highlight the significant
potential synergies between constraint programming and convex optimization
for complex engineering problems.

References
1. Achterberg, T.: Scip: solving constraint integer programs. Mathematical Program-
ming Computation 1(1), 1–41 (2009). http://dx.doi.org/10.1007/s12532-008-0001-1
 Strengthening Convex Relaxations with Bound Tightening 55

2. Bai, X., Wei, H., Fujisawa, K., Wang, Y.: Semidefinite programming for optimal
power flow problems. International Journal of Electrical Power & Energy Systems
30(67), 383–392 (2008)
3. Belotti, P.: Couenne: User manual (2009). https://projects.coin-or.org/Couenne/
(accessed April 10, 2015)
4. Belotti, P., Cafieri, S., Lee, J., Liberti, L.: On feasibility based bounds tightening
(2012). http://www.optimization-online.org/DB HTML/2012/01/3325.html
5. Belotti, P., Lee, J., Liberti, L., Margot, F., Wachter, A.: Branching and bounds
tightening techniques for non-convex minlp. Optimization Methods Software 24(4–
5), 597–634 (2009)
6. Benhamou, F., McAllester, D., Van Hentenryck, P.: Clp (intervals) revisited.Tech.
rep., Brown University, Providence, RI, USA (1994)
7. Bent, R., Coffrin, C., Gumucio, R., Van Hentenryck, P.: Transmission network
expansion planning: Bridging the gap between ac heuristics and dc approximations.
In: Proceedings of the 18th Power Systems Computation Conference (PSCC 2014),
Wroclaw, Poland (2014)
8. Berthold, T., Heinz, S., Vigerske, S.: Extending a cip framework to solvemiqcps.
In: Lee, J., Leyffer, S. (eds.) Mixed Integer Nonlinear Programming. The IMA
Volumes in Mathematics and its Applications, vol. 154, pp. 427–444. Springer
New York (2012)
9. Chen, Y., Luh, P., Guan, C., Zhao, Y., Michel, L., Coolbeth, M., Friedland, P.,
Rourke, S.: Short-term load forecasting: Similar day-based wavelet neural networks.
IEEE Transactions on Power Systems 25(1), 322–330 (2010)
10. Coffrin, C., Gordon, D., Scott, P.: NESTA, The Nicta Energy System Test Case
Archive. CoRR abs/1411.0359 (2014). http://arxiv.org/abs/1411.0359
11. Coffrin, C., Hijazi, H., Van Hentenryck, P.: The QC Relaxation: Theoretical and
Computational Results on Optimal Power Flow. CoRR abs/1502.07847 (2015).
http://arxiv.org/abs/1502.07847
12. Coffrin, C., Van Hentenryck, P.: Transmission system restoration: Co-optimization
of repairs, load pickups, and generation dispatch. International Journal of Electrical
Power & Energy Systems (2015) (forthcoming)
13. Farivar, M., Clarke, C., Low, S., Chandy, K.: Inverter var control for distribution
systems with renewables. In: 2011 IEEE International Conference on Smart Grid
Communications (SmartGridComm), pp. 457–462, October 2011
14. Fisher, E., O’Neill, R., Ferris, M.: Optimal transmission switching. IEEE Transac-
tions on Power Systems 23(3), 1346–1355 (2008)
15. Fourer, R., Gay, D.M., Kernighan, B.: AMPL: a mathematical programming lan-
guage. In: Wallace, S.W. (ed.) Algorithms and Model Formulations in Mathemat-
ical Programming, pp. 150–151. Springer-Verlag New York Inc., New York (1989)
16. Fu, Y., Shahidehpour, M., Li, Z.: Security-constrained unit commitment with ac
constraints*. IEEE Transactions on Power Systems 20(3), 1538–1550 (2005)
17. Hedman, K., Ferris, M., O’Neill, R., Fisher, E., Oren, S.: Co-optimization of gen-
eration unit commitment and transmission switching with n-1 reliability. In: 2010
IEEE Power and Energy Society General Meeting, pp. 1–1, July 2010
18. Hijazi, H., Coffrin, C., Van Hentenryck, P.: Convex quadratic relaxations
of mixed-integer nonlinear programs in power systems (2013). http://www.
optimization-online.org/DB HTML/2013/09/4057.html
56 C. Coffrin et al.

19. Hijazi, H., Thiebaux, S.: Optimal ac distribution systems reconfiguration. In:
Proceedings of the 18th Power Systems Computation Conference (PSCC 2014),
Wroclaw, Poland (2014)
20. Jabr, R.: Radial distribution load flow using conic programming. IEEE Transac-
tions on Power Systems 21(3), 1458–1459 (2006)
21. Jabr, R.: Optimization of ac transmission system planning. IEEE Transactions on
Power Systems 28(3), 2779–2787 (2013)
22. Kundur, P.: Power System Stability and Control. McGraw-Hill Professional (1994)
23. Lavaei, J., Low, S.: Zero duality gap in optimal power flow problem. IEEE Trans-
actions on Power Systems 27(1), 92–107 (2012)
24. Lehmann, K., Grastien, A., Van Hentenryck, P.: AC-Feasibility on Tree Networks
is NP-Hard. IEEE Transactions on Power Systems (2015) (to appear)
25. Liberti, L.: Writing global optimization software. In: Liberti, L., Maculan, N.
(eds.) Global Optimization, Nonconvex Optimization and Its Applications, vol. 84,
pp. 211–262. Springer, US (2006). http://dx.doi.org/10.1007/0-387-30528-9 8
26. Madani, R., Ashraphijuo, M., Lavaei, J.: Promises of conic relaxation for
contingency-constrained optimal power flow problem (2014). http://www.ee.
columbia.edu/lavaei/SCOPF 2014.pdf (accessed February 22, 2015)
27. McCormick, G.: Computability of global solutions to factorable nonconvex pro-
grams: Part i convex underestimating problems. Mathematical Programming 10,
146–175 (1976)
28. Miller, J.: Power system optimization smart grid, demand dispatch, and micro-
grids, September 2011. http://www.netl.doe.gov/smartgrid/referenceshelf/
presentations/SE%20Dist%20Apparatus%20School Final 082911 rev2.pdf
(accessed April 22, 2012)
29. Momoh, J., Adapa, R., El-Hawary, M.: A review of selected optimal power flow lit-
erature to 1993. i. nonlinear and quadratic programming approaches. IEEE Trans-
actions on Power Systems 14(1), 96–104 (1999)
30. Momoh, J., El-Hawary, M., Adapa, R.: A review of selected optimal power flow
literature to 1993. ii. newton, linear programming and interior point methods.
IEEE Transactions on Power Systems 14(1), 105–111 (1999)
31. Montanari, U.: Networks of Constraints : Fundamental Properties and Applications
to Picture Processing. Information Science 7(2), 95–132 (1974)
32. Ott, A.: Unit commitment in the pjm day-ahead and real-time mar-
kets, June 2010. http://www.ferc.gov/eventcalendar/Files/20100601131610-Ott,
%20PJM.pdf (accessed April 22, 2012)
33. Purchala, K., Meeus, L., Van Dommelen, D., Belmans, R.: Usefulness of DC power
flow for active power flow analysis. In: Power Engineering Society General Meeting,
pp. 454–459 (2005)
34. Ryoo, H., Sahinidis, N.: A branch-and-reduce approach to global optimization.
Journal of Global Optimization 8(2), 107–138 (1996)
35. Sahinidis, N.: Global optimization and constraint satisfaction: the branch-and-
reduce approach. In: Bliek, C., Jermann, C., Neumaier, A. (eds.) Global Optimiza-
tion and Constraint Satisfaction. LNCS, vol. 2861, pp. 1–16. Springer, Heidelberg
(2003)
36. Toh, K.C., Todd, M., Ttnc, R.H.: Sdpt3 - a matlab software package for semidefi-
nite programming. Optimization Methods and Software 11, 545–581 (1999)
37. Van Hentenryck, P., McAllister, D., Kapur, D.: Solving Polynomial Systems Using
a Branch and Prune Approach. SIAM Journal on Numerical Analysis 34(2) (1997)
 Strengthening Convex Relaxations with Bound Tightening 57

38. Van Hentenryck, P., Michel, L., Deville, Y.: Numerica: a Modeling Language for
Global Optimization. The MIT Press, Cambridge (1997)
39. Van Hentenryck, P., Saraswat, V., Deville, Y.: The design, implementation, and
evaluation of the constraint language cc(FD). In: Podelski, A. (ed.) Constraint Pro-
gramming: Basics and Trends. LNCS, vol. 910, pp. 293–316. Springer, Heidelberg
(1995)
40. Verma, A.: Power grid security analysis: An optimization approach. Ph.D. thesis,
Columbia University (2009)
41. Wächter, A., Biegler, L.T.: On the implementation of a primal-dual interior point
filter line search algorithm for large-scale nonlinear programming. Mathematical
Programming 106(1), 25–57 (2006)
 Broken Triangles Revisited

Martin C. Cooper1(B) , Aymeric Duchein1 , and Guillaume Escamocher2

1
IRIT, University of Toulouse III, 31062 Toulouse, France
{cooper,Aymeric.Duchein}@irit.fr
2
INSIGHT Centre for Data Analytics, University College Cork, Cork, Ireland
guillaume.escamocher@insight-centre.org

Abstract. A broken triangle is a pattern of (in)compatibilities between

assignments in a binary CSP (constraint satisfaction problem). In
the absence of certain broken triangles, satisfiability-preserving domain
reductions are possible via merging of domain values. We investigate the
possibility of maximising the number of domain reduction operations by
the choice of the order in which they are applied, as well as their inter-
action with arc consistency operations. It turns out that it is NP-hard
to choose the best order.

1 Introduction
The notion of broken triangle has generated a certain amount of interest in the
constraints community: it has led to the definition of novel tractable classes [3,7],
variable elimination rules [1] and domain reduction rules [4,5]. The merging of
pairs of values in the same variable domain which do not belong to a broken
triangle has been shown to lead to considerable reduction of search space size
for certain benchmark instances of binary CSP [4]. The corresponding reduction
operation, known as BTP-merging, is satisfiability-preserving and is therefore
worthy of a deeper theoretical analysis as a potentially useful preprocessing
operation. An obvious question is whether the order in which BTP-merging
operations, and other domain-reduction operations such as arc consistency, are
performed has an effect on the number of possible merges.
Definition 1. A binary CSP instance I consists of
– a set X of n variables,
– a domain D(x) of possible values for each variable x ∈ X, with d the maxi-
mum domain size,
– a relation Rxy ⊆ D(x) × D(y), for each pair of distinct variables x, y ∈ X,
which consists of the set of compatible pairs of values (a, b) for variables
(x, y).
A partial solution to I on Y = {y1 , . . . , yr } ⊆ X is a set {y1 , a1 , . . . , yr , ar }
such that ∀i, j ∈ [1, r], (ai , aj ) ∈ Ryi yj . A solution to I is a partial solution on
X.
M.C. Cooper—supported by ANR Project ANR-10-BLAN-0210 and EPSRC grant
EP/L021226/1.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 58–73, 2015.
DOI: 10.1007/978-3-319-23219-5 5
 Broken Triangles Revisited 59

For simplicity of presentation, Definition 1 assumes that there is exactly one

constraint relation for each pair of variables. An instance I is arc consistent if
for each pair of distinct variables x, y ∈ X, for each value a ∈ D(x), there is a
value b ∈ D(y) such that (a, b) ∈ Rxy .



• a
d •hhhh 
A h
h• b

y A   
A  x
e A•

z

Fig. 1. A broken triangle on two values a, b ∈ D(x).

We now formally define the value-merging operation based on absence of

certain broken triangles. A broken triangle on values a, b is shown in Figure 1.
In all figures in this paper, the pairs of values joined by a solid line are exactly
those belonging to the corresponding constraint relation.
Definition 2. A broken triangle on the pair of variable-value assignments a, b ∈
D(x) consists of a pair of assignments d ∈ D(y), e ∈ D(z) to distinct variables
y, z ∈ X \ {x} such that (a, d) ∈ / Rxy , (b, d) ∈ Rxy , (a, e) ∈ Rxz , (b, e) ∈
/ Rxz
and (d, e) ∈ Ryz . The pair of values a, b ∈ D(x) is BT-free if there is no broken
triangle on a, b.
BTP-merging values a, b ∈ D(x) in a binary CSP consists in replacing a, b
in D(x) by a new value c which is compatible with all variable-value assignments
compatible with at least one of the assignments x, a or x, b (i.e. ∀y ∈ X \ {x},
∀d ∈ D(y), (c, d) ∈ Rxy iff (a, d) ∈ Rxy or (b, d) ∈ Rxy )
When a, b ∈ D(x) are BT-free in a binary CSP instance I, the instance I 
obtained from I by merging a, b ∈ D(x) is satisfiable if and only if I is satisfiable.
Furthermore, given a solution to the instance resulting from the merging of two
values, we can find a solution to the original instance in linear time [4].
The paper is structured as follows. In Section 2 we investigate the interaction
between arc consistency and BTP-merging. In Section 3 we show that finding the
best order in which to apply BTP-mergings is NP-hard, even for arc-consistent
instances. In Section 4 we prove that this remains true even if we only perform
merges at a single variable. In Section 5 we take this line of work one step further
by showing that it is also NP-hard to find the best sequence of merges by a weaker
property combing virtual interchangeability and neighbourhood substitutability.

2 Mixing Arc Consistency and BTP-merging

BTP-merging can be seen as a generalisation of neighbourhood substitutabil-
ity [6], since if a ∈ D(x) is neighbourhood substitutable for b ∈ D(x) then a, b
60 M.C. Cooper et al.

can be BTP-merged. The possible interactions between arc consistency (AC) and
neighbourhood substitution (NS) are relatively simple and can be summarised
as follows:
1. The fact that a ∈ D(x) is AC-supported or not at variable y remains invariant
after the elimination of any other value b (in D(x) \ {a} or in the domain
D(z) of any variable z = x) by neighbourhood substitution.
2. An arc-consistent value a ∈ D(x) that is neighbourhood substitutable
remains neighbourhood substitutable after the elimination of any other value
by arc consistency.
3. On the other hand, a value a ∈ D(x) may become neighbourhood substi-
tutable after the elimination of a value c ∈ D(y) (y = x) by arc consistency.
Indeed, it has been shown that the maximum cumulated number of elimina-
tions by arc consistency and neighbourhood substitution can be achieved by
first establishing arc consistency and then applying any convergent sequence
of NS eliminations (i.e. any valid sequence of eliminations by neighbourhood
substitution until no more NS eliminations are possible) [2].

   
• • • •
 

 
a • c •
b • @
•l • @
 @ l  @ @
@ll @ @
@ l @ @
@ l  @ @ 
(a) @• l• (b) @• @•
 

Fig. 2. (a) An instance in which applying AC leads to the elimination of all values
(starting with the values a and b), but applying BTP merging leads to just one elimi-
nation, namely the merging of a with b (with the resulting instance shown in (b)).

The interaction between arc consistency and BTP-merging is not so simple

and can be summarised as follows:
1. The fact that a ∈ D(x) is AC-supported or not at variable y remains invariant
after the BTP-merging of any other pair of other values b, c (in D(x)\{a} or in
the domain D(z) of any variable z = x). However, after the BTP-merging of
two arc-inconsistent values the resulting merged value may be arc consistent.
An example is given in Figure 2(a). In this 3-variable instance, the two values
a, b ∈ D(x) can be eliminated by arc consistency (which in turn leads to
 Broken Triangles Revisited 61

the elimination of all values), or alternatively they can be BTP-merged (to

produce the value c) resulting in the instance shown in Figure 2(b) in which
no more eliminations are possible by AC or BTP-merging.
2. A single elimination by AC may prevent a sequence of several BTP-mergings.
An example is given in Figure 3(a). In this 4-variable instance, if the value b
is eliminated by AC, then no other eliminations are possible by AC or BTP-
merging in the resulting instance (shown in Figure 3(b)), whereas if a and b
are BTP-merged into a new value d (as shown in Figure 3(c)) this destroys
a broken triangle thus allowing c to be BTP-merged with d (as shown in
Figure 3(d)).
3. On the other hand, two values in the domain of a variable x may become
BTP-mergeable after an elimination of a value c ∈ D(y) (y = x) by arc
consistency.

3 The Order of BTP-mergings

It is known that BTP-merging can both create and destroy broken triangles [4].
This implies that the choice of the order in which BTP-mergings are applied
may affect the total number of merges that can be performed. Unfortunately,
maximising the total number of merges in a binary CSP instance turns out to be
NP-hard, even when bounding the maximum size of the domains d by a constant
as small as 3. For simplicity of presentation, we first prove this for the case in
which the instance is not necessarily arc consistent. We will then prove a tighter
version, namely NP-hardness of maximising the total number of merges even in
arc-consistent instances.
Theorem 1. The problem of determining if it is possible to perform k BTP-
mergings in a boolean binary CSP instance is NP-complete.
Proof. For a given sequence of k BTP-mergings, verifying if this sequence is
correct can be performed in O(kn2 d2 ) time because looking for broken triangles
for a given couple of values takes O(n2 d2 ) [4]. As we can verify a solution in
polynomial time, the problem of determining if it is possible to perform k BTP-
mergings in a binary CSP instance is in NP. So to complete the proof of NP-
completeness it suffices to give a polynomial-time reduction from the well-known
3-SAT problem. Let I3SAT be an instance of 3-SAT (SAT in which each clause
contains exactly 3 literals) with variables X1 , . . . , XN and clauses C1 , . . . , CM .
We will create a boolean binary CSP instance ICSP which has a sequence of
k = 3 × M mergings if and only if I3SAT is satisfiable.
For each variable Xi of I3SAT , we add a new variable zi to ICSP . For each
occurrence of Xi in the clause Cj of I3SAT , we add two more variables xij and
yij to ICSP . Each D(zi ) contains only one value ci and each D(xij ) (resp. D(yij ))
contains only two values ai and bi (resp. ai and bi ). The roles of variables xij
and yij are the following:
Xi = true ⇔ ∀j, ai , bi can be merged in D(xij ) (1)
Xi = f alse ⇔ ∀j, ai , bi can be merged in D(yij ) (2)
62 M.C. Cooper et al.

   
• • • •

%@ @ @ @
% @ @ @ @
% @ @ @ @
 % @ @   @ @ 
a •h% h @•
@ ( a •hh @•
@ (
@ hh( h( h(((( hhh
h( (((
b •% ( h( h h @ @
( (h(
( h h @
l (( hhh (( hhh
(@
c •e l@
@• c •(@ @•
 @e   @ @ 
l
@e @l @ @
@e @ l @ @
@e@l  @ @ 
(a) l•
e• @
@ (b) @• @•
 

   
• • • •
 @ @ %@,@

@ @ % ,@ @
@ @ %,, @ @
 @ @  %, @ @ 
d •hhh @ (@• % 
, @ @•
hh( ((((
@ ( h(
h h @ •,
% @
((( hhh l
c •(@ h
@•  e
l
@•
 @ @  e 
@ @ ell
@ @ e l
@ @  el 
(c) @• @• (d) e• l•
 

Fig. 3. (a) An instance in which applying AC leads to one elimination (the value b)
(as shown in (b)), but applying BTP merging leads to two eliminations, namely a with
b (shown in (c)) and then d with c (shown in (d)).

In order to prevent the possibility of merging both (ai , bi ) and (ai , bi ), we
define the following constraints for zi , xij and yij : ∀j Rxij zi = {(bi , ci )} and
Ryij zi = {(bi , ci )}; ∀j ∀ k Rxij yik = {(ai , ai )}. These constraints are shown
in Figure 4(a) for a single j (where a pair of points not joined by a solid line
are incompatible). By this gadget, we create a broken triangle on each yij when
merging values in the xij and vice versa.
Then for each clause Ci = (Xj , Xk , Xl ), we add the following constraints in
order to have at least one of the literals Xj , Xk , Xl true: Ryji yki = {(aj , bk )},
Ryki yli = {(ak , bl )} and Ryli yji = {(al , bj )}. This construction, shown in
Figure 4(b), is such that it allows two mergings on the variables yji , yki , yli
before a broken triangle is created. For example, merging aj , bj and then ak , bk
creates a broken triangle on ai , bi . So a third merging is not possible.
 Broken Triangles Revisited 63

   
ai • • ai aj • • bk

bi • • bi bj • • ak

 @  S    

xij @  yij yji S  yki
@ S • bl
@• ci S
 S• al
zi 
(a) (b) yli

Fig. 4. (a) Representation of the variable Xi and its negation (by the possibility of
performing a merge in D(xij ) or D(yij ), respectively, according to rules (1),(2)). (b)
Representation of the clause (Xj ∨ Xk ∨ Xl ). Pairs of points joined by a solid line are
compatible and incompatible otherwise.

If the clause Ci contains a negated literal Xj instead of Xj , it suffices to

replace yji by xji . Indeed, Figure 5 shows the construction for the clause (Xj ∨
Xk ∨ Xl ) together with the gadgets for each variable. The maximum number
of mergings that can be performed are one per occurrence of each variable in
a clause, which is exactly 3 × M . Given a sequence of 3 × M mergings in the
CSP instance, there is a corresponding solution to I3SAT given by (1) and (2).
The above reduction allows us to code I3SAT as the problem of testing the
existence of a sequence of k = 3 × M mergings in the corresponding instance
ICSP . This reduction being polynomial, we have proved the NP-completeness
of the problem of determining whether k BTP merges are possible in a boolean
binary CSP instance.

zj zk
 
cj • ck •
@  @
  @   @ 

bj • @• bj 
bk • @• bk
T  
•Taj
aj • T ak • • ak
  T 
 
yji
 xji
T  yki xki
T 
al •

• al

bl • • bl
 e 
%
yli e   %x
e % li
e%
cl •

zl

Fig. 5. Gadget representing the clause (Xj ∨ Xk ∨ Xl ).

64 M.C. Cooper et al.

The reduction given in the proof of Theorem 1 supposes that no arc consis-
tency operations are used. We will now show that it is possible to modify the
reduction so as to prevent the elimination of any values in the instance ICSP
by arc consistency, even when the maximum size of the domains d is bounded
by a constant as small as 3. Recall that an arc-consistent instance remains arc-
consistent after any number of BTP-mergings.

Theorem 2. The problem of determining if it is possible to perform k BTP-

mergings in an arc-consistent binary CSP instance is NP-complete, even when
only considering binary CSP instances where the size of the domains is bounded
by 3.

Proof. In order to ensure arc consistency of the instance ICSP , we add a new
value di to the domain of each of the variables xij , yij , zi . However, we can-
not simply make di compatible with all values in all other domains, because
this would allow all values to be merged with di , destroying in the process the
semantics of the reduction.
In the three binary constraints concerning the triple of variables xij , yij , zi ,
we make di compatible with all values in the other two domains except di . In
other words, we add the following tuples to constraint relations, as illustrated in
Figure 6:

– ∀i∀j, (ai , di ), (bi , di ), (di , ci ) ∈ Rxij zi

– ∀i∀j, (ai , di ), (bi , di ), (di , ci ) ∈ Ryij zi
– ∀i∀j, (ai , di ), (bi , di ), (di , ai ), (di , bi ) ∈ Rxij yij

This ensures arc consistency, without creating new broken triangles on ai , bi or

ai , bi , while at the same time preventing BTP-merging with the new value di .
It is important to note that even after BTP-merging of one of the pairs ai , bi or
ai , bi , no BTP-merging is possible with di in D(xij ), D(yij ) or D(zi ) due to the
presence of broken triangles on this triple of variables. For example, the pair of
values ai , di ∈ D(xij ) belongs to a broken triangle on ci ∈ D(zi ) and di ∈ D(yij ),
and this broken triangle still exists if the values ai , bi ∈ D(yij ) are merged.
We can then simply make di compatible with all values in the domain of all
variables outside this triple of variables. This ensures arc consistency, and does
i z
 
• ci

AS

• AS

 QA Sdi
 

 @Q S 
@AQ



bi •P @
AQ
S• bi
P 
ai •h
hP 
P @
A
( • ai

(h
(h(P h(
(P
hP
(
di • hA• di
 
xij yij

Fig. 6. Ensuring arc consistency between the variables zi , yij , xij by addition of new
values di .
 Broken Triangles Revisited 65

not introduce any broken triangles on ai , bi or ai , bi . With these constraints we
ensure arc consistency without changing any of the properties of ICSP used in
the reduction from 3-SAT described in the proof of Theorem 1. For each pair of
values ai , bi ∈ D(xij ) and ai , bi ∈ D(yij ), no new broken triangle is created since
these two values always have the same compatibility with all the new values dk .
As we have seen, the constraints shown in Figure 6 prevent any merging of the
new values dk .

Corollary 1. The problem of determining if it is possible to perform k value

eliminations by arc consistency and BTP-merging in a binary CSP instance is
NP-complete, even when only considering binary CSP instances where the size
of the domains is bounded by 3.

From a practical point of view, an obvious question concerns the existence of

a non-optimal but nevertheless useful heuristic for choosing the order of BTP-
merges. Imagine that a heuristic concerning the order in which to apply BTP-
merges involves first finding all possible merges before choosing between them.
If this is the case, then once a merge has been chosen and performed, the list of
possible merges has to be recalculated. This process is thus already an order of
magnitude slower than the simple technique (applied in the experiments in [4])
consisting of performing a merge as soon as it is detected.

4 Optimal Sequence of BTP-mergings at a Single

Variable
We now show that even when only considering a single domain, finding the
optimal order of BTP-mergings is NP-Complete. For simplicity of presentation,
we first prove this for the case in which the instance is not necessarily arc-
consistent. We then prove a tighter version for arc-consistent instances.

Theorem 3. The problem of determining if it is possible to perform k BTP-

mergings within a same domain in a binary CSP instance is NP-Complete.

Proof. For a given sequence of k BTP-mergings within a domain of a binary CSP

instance I, verifying if this sequence is correct can be performed in O(kN 2 d2 )
time, with N being the number of variables in I and d being the size of the largest
domain in I, because looking for broken triangles for a given couple of values
takes O(N 2 d2 ). As we can verify a solution in polynomial time, the problem of
determining if it is possible to perform k BTP-mergings within a single domain in
a binary CSP instance is in NP. So to complete the proof of NP-Completeness, it
suffices to give a polynomial-time reduction from the well-known SAT problem.
Let ISAT be a SAT instance with n variables X1 , X2 , . . . , Xn and m clauses
C1 , C2 , . . . , Cm . We reduce ISAT to a binary CSP instance ICSP containing a
variable v0 such that ISAT is satisfiable if and only if we can make k = n + m
BTP-mergings within D(v0 ). ICSP is defined in the following way:
1. ICSP contains 1+n(2+4m+9(n−1)) variables v0 , v1 , v2 , . . . , vn(2+4m+9(n−1)) .
66 M.C. Cooper et al.

2. D(v0 ) contains 3×n+m values with the following names: x1 , x2 , . . . , xn , x1 T,

x2 T, . . . , xn T, x1 F, x2 F, . . . , xn F, c1 , c2 , . . . , cm . All other domains in ICSP
only contain one value.
3. ∀i ∈ [1, n], xi T and xi F can never be BTP-merged. The idea here is to allow
exactly one BTP-merging among the three values xi , xi T and xi F : xi and
xi T if Xi is assigned True in the SAT instance ISAT , xi and xi F if Xi is
assigned False instead.
4. ∀(i, j) ∈ [1, n] × [1, m] such that Cj does not contain Xi (respectively Xi ),
cj can never be BTP-merged with xi T (respectively xi F ).
5. ∀(i, j) ∈ [1, n]×[1, m] such that Cj contains Xi (respectively Xi ), cj can only
be BTP-merged with xi T (respectively xi F ) if either cj or xi T (respectively
xi F ) has been previously BTP-merged with xi .
6. ∀i, j with 1 ≤ i < j ≤ n, the nine following couples can never be BTP-
merged: xi and xj , xi and xj T , xi and xj F , xi T and xj , xi T and xj T , xi T
and xj F , xi F and xj , xi F and xj T , xi F and xj F . The idea here is to prevent
any BTP-merging between two CSP values corresponding to two different
SAT variables.

When we say that two values a and b can never be BTP-merged, it means that
we add two variables va and vb , with only one value a in D(va ), and only one
value b in D(vb ), such that a is compatible with a and incompatible with b , b
is compatible with b and incompatible with a , a is compatible with b and all
other edges containing either a or b are incompatible. The purpose of making
a (respectively b ) incompatible will all values in the instance except a and b
(respectively b and a ) is twofold. First, it ensures that no future BTP-merging
can establish a compatibility between a (respectively b ) and b (respectively a)
and thus destroy the broken triangle. Second, it ensures that the only broken
triangle introduced by a and b is on a and b, so that the addition of a and b
does not prevent any other BTP-merging than the one between a and b.
When we say that two values a and b can only be BTP-merged if either a
or b has been previously BTP-merged with some third value c, it means that
we add two variables va and vb , with only one value a in D(va ), and only one
value b in D(vb ), such that a is compatible with a and incompatible with b ,
b is compatible with b and incompatible with a , c is compatible with both a
and b , a is compatible with b and all other edges containing either a or b are
incompatible. Here again, the purpose of making a (respectively b ) incompatible
will all values in the instance except a, b and c (respectively b, a and c) is
twofold. First, it ensures that no future BTP-merging that does not include c
can establish a compatibility between a (respectively b ) and b (respectively a)
and thus destroy the broken triangle. Second, it ensures that the only broken
triangle introduced by a and b is on a and b, so that the addition of a and b
does not prevent any other BTP-merging than the one between a and b.
For every couple of values that can never be BTP-merged, and for every
couple of values that can only be BTP-merged when one of them has been
previously BTP-merged with some third value, we add two new single-valued
variables to ICSP . Therefore, the third point in the definition of ICSP adds 2n
 Broken Triangles Revisited 67

variables to ICSP , the fourth and fifth points in the definition of ICSP add 4nm
variables to ICSP and the sixth point in the definition of ICSP adds 9n(n − 1)
variables to ICSP . Therefore, the total number of single-valued variables added
to ICSP is n(2 + 4m + 9(n − 1)), as expected from the first point in the definition
of ICSP .

– The number of BTP-mergings is limited by n + m:

From the third point in the definition of ICSP , for all i ∈ [1, n], we can BTP-
merge at most once within the triple {xi , xi T, xi F }. From the sixth point
in the definition of ICSP , we cannot BTP-merge two values within D(v0 ) if
they are associated to two different SAT variables Xi and Xj . Therefore, we
have at most m BTP-mergings remaining, one for each cj for 1 ≤ j ≤ m.
– If we can BTP-merge n + m times, then we have a solution for ISAT :
Since we have done the maximum number of BTP-mergings, we know that
for all i ∈ [1, n], xi has been BTP-merged with either xi T or xi F , but not
both. So we create the following solution for ISAT : ∀i ∈ [1, n], we assign T rue
to Xi if xi and xi T have been BTP-merged, and F alse otherwise. From the
fourth and fifth points in the definition of ICSP , we know that for each j
in [1, m], Cj is satisfied by the literal associated with the value Cj has been
BTP-merged with.
– If we have a solution for ISAT , then we can BTP-merge n + m times:
∀i ∈ [1, n], we BTP-merge xi with xi T if T rue has been assigned to Xi , with
xi F otherwise. ∀(i, j) ∈ [1, n] × [1, m], we BTP-merge cj and the value that
xi has been BTP-merged with if Xi is satisfied in Cj and cj has not been
BTP-merged yet. From the fifth point in the definition of ICSP , we know we
can BTP-merge each cj once.

Therefore ISAT is satisfiable if and only if we can perform k = n + m BTP-

mergings within D(v0 ), and we have the result.

We now generalise the result of Theorem 3 to arc-consistent binary CSP

instances.

Theorem 4. The problem of determining if it is possible to perform k BTP-

mergings within a same domain in an arc-consistent binary CSP instance is
NP-Complete.

Proof. We transform the binary CSP instance ICSP from the proof of Theorem 3

into an arc-consistent binary CSP instance ICSP . To do so, we add a new value
di in D(vi ) for 1 ≤ i ≤ n(2 + 4m + 9(n − 1)) such that all di are incompatible

with each other and compatible with all other points in ICSP . This ensures arc
consistency. It remains to show that:

1. For any couple of values (a, b) ∈ D(v0 ), adding the values di does not create
the broken triangle on a and b, even if a or b is the result of a previous
BTP-merging:
Suppose that we have two values a, b ∈ D(v0 ) such that adding the values di
68 M.C. Cooper et al.

creates a broken triangle on a and b. Let a ∈ D(va ) and b ∈ D(vb ) be the

other two values forming the broken triangle. Since it was the new values di
that created this particular broken triangle, either a or b is one of the di .
Without loss of generality, we assume that a is one of the di . But since the
di are compatible with all values from D(v0 ), both a and b are compatible
with a , even if a or b is the result of a previous BTP-merging. Therefore,
there cannot be any broken triangle on a and b caused by the new values di .
2. For all i ∈ [1, n(2 + 4m + 9(n − 1))], it is never possible to BTP-merge the
two values in D(vi ):
We assume, for simplicity of presentation and without loss of generality,
that the SAT instance ISAT has more than one variable and that no clause
contains both a variable and its negation. Let i ∈ [1, n(2 + 4m + 9(n − 1))].
Let a and di be the two points in D(vi ). From the proof of Theorem 3, we
know that a is compatible with only one value from D(v0 ). Let b be this
value. If b is associated with one of the SAT variables from ISAT , then from
the sixth point in the definition of ICSP in the proof of Theorem 3 we know
that there is at least one value c ∈ D(v0 ) that can never be BTP-merged
with b, and therefore will always be incompatible with a. If on the other
hand c is associated with one of the SAT clauses from ISAT , then from the
fourth point in the definition of ICSP in the proof of Theorem 3 we know
that there is at least one value c ∈ D(v0 ) that can never be BTP-merged
with b, and therefore will always be incompatible with a. Therefore, we have
a value c ∈ D(v0 ) that is always incompatible with a, even if c is the result
of a previous BTP-merging. Let j ∈ [1, n(2 + 4m + 9(n − 1))], such that
j = i. Since the di are incompatible with each other, and compatible with

all other values in ICSP , then dj is compatible with both a and c, and di
is compatible with c and incompatible with dj . Therefore we have a broken
triangle on a and di that can never be destroyed. Therefore a and di can
never be BTP-merged and we have the result.

One motivation for studying the single-variable version of the problem was
that if all values in D(x) can be BTP-merged, then the variable x can be elim-
inated since its domain becomes a singleton. Our proof of NP-hardness in the
single-variable case relied on a large domain which was not actually reduced
to a singleton. There remains therefore an interesting open question concerning
the complexity of eliminating the largest number of variables by sequences of
BTP-merging operations.

5 Virtual Interchangeability and Neighbourhood

Substitution

Since testing whether two values a, b ∈ D(x) are BTP-mergeable requires testing
all pairs of assignments to all pairs of distinct variables y, z = x, it is natural
to investigate weaker versions which are less costly to test. Two such weaker
versions are neighbourhood substitutability [6] and virtual interchangeability [8].
 Broken Triangles Revisited 69

Given two values a, b ∈ D(x), a is neighbourhood substitutable by b (NS), and

so can be merged with b, if for all variables y = x, ∀c ∈ D(y), (a, c) ∈ Rxy ⇒
(b, c) ∈ Rxy . Two values a, b ∈ D(x) are virtual interchangeable (VI), and so
can be merged, if for all variables y = x except at most one, ∀c ∈ D(y), (a, c) ∈
Rxy ⇔ (b, c) ∈ Rxy . Applying both VI and neighbourhood substitution (NS)
operations until convergence provides a weaker and less time-costly alternative
to applying BTP-merging operations until convergence. An interesting question
is therefore whether it is possible to find in polynomial time an optimal (i.e.
longest) sequence of VI and NS operations. The following theorem shows that
this problem is in fact also NP-hard.

Theorem 5. Determining whether there exists a sequence of VI and NS opera-

tions of length k that can be applied to a binary CSP instance is NP-complete.

Proof. Since checking the validity of a sequence of VI and NS operations can be

achieved in polynomial time, the problem is in NP. To complete the proof we
demonstrate a polynomial reduction from 3SAT. Let I3SAT be an instance of
3SAT. We will show how to construct a binary CSP instance ICSP and a value
k so that it is possible to perform k merges by VI and NS if and only if I3SAT
is satisfiable.
For each variable x in I3SAT , we introduce Boolean variables called x and x
in ICSP : variable x in I3SAT is assigned true (respectively, false) if and only if the
two domain values in D(x) (respectively, D(x)) are merged in the corresponding
sequence of merges in ICSP . Figure 7(a) shows the gadget for choosing the truth
value for x. The variables x and x are both connected to another variable, not
shown in this figure: this prevents the values in their domains being merged
before the values in the domains of t or t are merged. Initially, the only merges
(by VI and NS) which can occur are in the domain of variable s: we can either
merge values 1 and 2 or values 2 and 3. Once one of these merges is performed,
the other it not possible. Figure 7 shows the propagation of merges which occurs
in the second of these cases. Figure 7(b) shows the result of this merge in D(s).
In Figure 7(b) the only merge that is possible is the merging of values 1 and
3 (by NS) in D(t). The result is shown in Figure 7(c). Now, the two values in
the domain of x can be merged (by VI) since x is constrained by a single other
variable (not shown in the figure). It is important to note that no other merges
are possible. By a similar chain of propagations, if we had chosen to merge 1
and 2 in D(s), then the values 1 and 3 in D(t ) would have been merged, and
finally the two values in D(x). This gadget therefore allows us to choose a truth
value for the corresponding variable x of I3SAT .
In order to code the instance I3SAT , we need to be able to have several
copies of each variable. This is achieved by the gadget shown in Figure 8(a).
The variables x1 and x2 are each assumed to be constrained by a single other
variable not shown in the figure. The variable x is the variable in the gadget of
Figure 7. If the values in D(x) are merged then this allows the merging (by VI)
of the pair of values 0, 1 ∈ D(u) and the merging of the pair of values 2,3. In the
resulting instance, the two values in the domain of xi (i = 1, 2) can be merged.
70 M.C. Cooper et al.

 x
t  •

  

1 •   •
QQ   

Q
2  •
s  Q Q
 
Q Q 
1 •  • Q Q•
S
@  3  Q  
 Q

2 •
PPS
@  Q •
SP
@P  t   

(a) P   
• S@ P•Q
3  PP 
S@ P
Q P  x
S @• PP
P Q •
S PPQ PQ
S PQP•
• 
 x
 
•
   
t
1 •Q 


•
Q

  
 Q

2 •Q Q
s  
Q Q 

1 •  • Q Q•
S 
@ 3   Q 
 
Q


•P
{2, 3}  S
@  Q•
PS@
PP  t   

(b)  
S@PP•P  

S @ QQ PP
 PP  x
S @•PP Q P•
Q
S  PPQ
S PQP•
•
 
 x
•

 •
 
t
 
 

2 •
Q 

s  
Q
  
1 •  • Q •
S  
@{1, 3} Q 

 
Q
S

•P
{2, 3}  @  Q•

PS@
PP  t   

(c)  
S@PP•P  

S @ QQ PP
 PP  x
S @•P Q P•
S PPQ PQ
S PQP•
• 

Fig. 7. (a) Gadget for choosing a truth value for x: true if the two values in D(x) are
merged; false if the two values in D(x) are merged. This same gagdet (b) after merging
the values 2 and 3 in D(s), then (c) after merging the values 1 and 3 in D(t) .
 Broken Triangles Revisited 71

u
  x1
• •

x  0

•P • •
PP 1 
(a) PP 

x2
•P P• •
 PP 2
PP
P• •
 3 


x •
Q
Q 
•Q QQ •
 S
v 
 Q Q  v 

y • Q Q• 
•P • S
Q  
PP
 S S 
(b) 
Q  P PP
•  Q 
• •


P S
PP S •
S
   PP
z •  S•
P
 

•

Fig. 8. (a) Gadget for making copies of a variable x: if the two values in D(x) can be
merged, then the two values in D(x1 ) and the two values in D(x2 ) can be merged. (b)
Gadget used in the simulation of a clause: the variable v is true (i.e. the two values in
its domain can be merged) if and only if x, y, z are all true.

This gadget therefore allows us to make multiple copies of the variable x all with
the same truth value.
To complete the proof, we now show how to code each clause c of I3SAT . There
are exactly seven assignments to the variables in c which satisfy this clause. For
each of these assignments, we add a gadget of the form shown in Figure 8(b). The
variables x, y, z are the output of the gadgets introduced above and correspond
to the variables x, y, z occurring in the clause c in I3SAT . In the example shown
in the figure, the satisfying assignment is x = y = z = true. When the two
values in each of the domains of these three variables can be merged (and only
in this case), the values in the domain of v can also be merged. The triangle of
variables to the right of v in Figure 8(b) prevents the merging of the values in
D(v) when only two of the three variables x, y, z are assigned true.
In order to have the same number of copies of x and x in our construction,
we also add a gadget similar to Figure 8(b) for the one non-satisfying assignment
to the three variables of the clause c: in this case, the variable v is constrained
by two other variables (as is the variable v  in Figure 8(b)) which prevents the
merging of the values in D(v).
Suppose that there are n variables and m clauses in I3SAT . The maximum
total number of merges which can be performed in ICSP is 3 per gadget shown in
72 M.C. Cooper et al.

Figure 7, 4 per gadget shown in Figure 8(a) and 1 per gadget shown in Figure 8(b)
(provided the gadget corresponds to a truth assignment which satisfies the clause
c). Each clause c requires four copies of each of the three variables x occurring
in c (as well as four copies of x). For each copy of each literal assigned the value
true, there are 4 merges in the gadget of Figure 8(a). For the first occurrence
of each variable, produced by the gadget of Figure 7, there is one less merge (3
instead of 4). Finally, for each satisfied clause there is one merge. This implies
that we can perform a total of k = 48m − n + m = 49m − n merges in ICSP if
and only if I3SAT is satisfiable. Since this reduction is clearly polynomial, this
completes the proof.

6 Conclusion

We have investigated the possibility of maximising the number of domain reduc-

tion operations in binary CSP instances by choosing an optimal order in which
to apply them. Whereas for consistency and neighbourhood-substitution oper-
ations, the number of domain reduction operations can be maximised in poly-
nomial time, the problem becomes NP-hard when we allow merging operations,
such as virtual interchangeability or BTP-merging. We emphasise that this does
not detract from the possible utility of such value-merging operations in practice,
which is an independent question.
Different tractable subproblems of binary CSP have been defined based on
the absence of certain broken triangles [3,5,7]. Instances can be solved by elim-
inating variables one by one and, in each case, the recognition of instances in
the tractable class can be achieved in polynomial time by a greedy algorithm
since the elimination of one variable cannot prevent the elimination of another
variable. BTP-merging, on the other hand, performs reduction operations on a
lower level than the elimination of variables. Given the NP-completeness results
in this paper, recognizing those instances which can be reduced to a trivial prob-
lem with only singleton domains by some sequence of BTP-merges is unlikely to
be tractable, but this remains an open problem.

References
1. Cohen, D.A., Cooper, M.C.: Guillaume Escamocher and Stanislav Živný, Variable
and Value Elimination in Binary Constraint Satisfaction via Forbidden Patterns. J.
Comp. Systems Science (2015). http://dx.doi.org/10.1016/j.jcss.2015.02.001
2. Martin, C.: Cooper, Fundamental Properties of Neighbourhood Substitution in Con-
straint Satisfaction Problems. Artif. Intell. 90(1–2), 1–24 (1997)
3. Cooper, M.C., Jeavons, P.G., Salamon, A.Z.: Generalizing constraint satisfaction on
trees: Hybrid tractability and variable elimination. Artif. Intell. 174(9–10), 570–584
(2010)
4. Cooper, M.C., El Mouelhi, A., Terrioux, C., Zanuttini, B.: On broken triangles.
In: O’Sullivan, B. (ed.) CP 2014. LNCS, vol. 8656, pp. 9–24. Springer, Heidelberg
(2014)
 Broken Triangles Revisited 73

5. Cooper, M.C.: Beyond consistency and substitutability. In: O’Sullivan, B. (ed.) CP

2014. LNCS, vol. 8656, pp. 256–271. Springer, Heidelberg (2014)
6. Freuder, E.C.: Eliminating interchangeable values in constraint satisfaction prob-
lems. In: Proceedings AAAI 1991, pp. 227–233 (1991)
7. Jégou, P., Terrioux, C.: The extendable-triple property: a new CSP tractable class
beyond BTP. In: AAAI (2015)
8. Likitvivatanavong, C., Yap, R.H.C.: Eliminating redundancy in CSPs through merg-
ing and subsumption of domain values. ACM SIGAPP Applied Computing Review
13(2) (2013)
 A Microstructure-Based Family of Tractable
Classes for CSPs

Martin C. Cooper1(B) , Philippe Jégou2 , and Cyril Terrioux2

1
IRIT, University of Toulouse III, 31062 Toulouse, France
cooper@irit.fr
2
Aix-Marseille Université, CNRS, ENSAM, Université de Toulon, LSIS UMR 7296,
Avenue Escadrille Normandie-Niemen, 13397 Marseille Cedex 20, France
{philippe.jegou,cyril.terrioux}@lsis.org

Abstract. The study of tractable classes is an important issue in Arti-

ficial Intelligence, especially in Constraint Satisfaction Problems. In
this context, the Broken Triangle Property (BTP) is a state-of-the-art
microstructure-based tractable class which generalizes well-known and
previously-defined tractable classes, notably the set of instances whose
constraint graph is a tree. In this paper, we propose to extend and to
generalize this class using a more general approach based on a param-
eter k which is a given constant. To this end, we introduce the k-BTP
property (and the class of instances satisfying this property) such that
we have 2-BTP = BTP, and for k > 2, k-BTP is a relaxation of BTP in
the sense that k-BTP  (k + 1)-BTP. Moreover, we show that if k-TW
is the class of instances having tree-width bounded by a constant k, then
k-TW  (k + 1)-BTP. Concerning tractability, we show that instances
satisfying k-BTP and which are strong k-consistent are tractable, that
is, can be recognized and solved in polynomial time. We also study the
relationship between k-BTP and the approach of Naanaa who proposed
a set-theoretical tool, known as the directional rank, to extend tractable
classes in a parameterized way. Finally we propose an experimental study
of 3-BTP which shows the practical interest of this class, particularly
w.r.t. the practical solving of instances satisfying 3-BTP and for other
instances, w.r.t. to backdoors based on this tractable class.

1 Introduction
Finding islands of tractability, generally called tractable classes is an important
issue in Artificial Intelligence, especially in Constraint Satisfaction Problems
(CSPs [1]). Many studies have addressed this issue, from the very beginnings
of Artificial Intelligence. These results are often theoretical in nature with, in
certain cases, tractable classes which can be considered as somewhat artifi-
cial. But some tractable classes have actually been used in practice, such as
the classes defined by constraint networks with bounded tree-width [2,3]. More
recently, the concept of hybrid class has been defined, for example with the
Supported by ANR Project ANR-10-BLAN-0210 and EPSRC grant EP/L021226/1.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 74–88, 2015.
DOI: 10.1007/978-3-319-23219-5 6
 A Microstructure-Based Family of Tractable Classes for CSPs 75

class BTP [4]. This class strictly contains both structural tractable classes (such
as tree-structured CSPs) and tractable classes defined by language restrictions.
One major advantage of this class, in addition to its generalization of already-
known tractable classes, is related to its practical interest. Indeed, instances of
this class can be solved in polynomial time using algorithms, such as MAC (Main-
taining Arc-Consistency [5]) and RFL (Real Full Look-ahead [6]), implemented
in efficient solvers which allows it to be used directly in practice. In addition,
it may also help to explain theoretically the practical efficiency of solvers, even
though the theoretical complexity of the algorithms employed by the solvers is
exponential in the worst case.
In this paper, we return to this type of approach by generalizing the tractable
class BTP which is defined by a property excluding certain patterns (called
Broken Triangles) in the microstructure graph associated with a binary CSP
instance. Very recent work in this same direction introduced the class ETP
[7] which generalizes BTP by relaxing some of its conditions, since it tolerates
some broken triangles which are forbidden by BTP . Here we propose a broader
generalization called k-BTP which extends this previous work along two axes.
First, in the spirit of ETP , the new class allows the presence of a larger number of
broken triangles, generalizing strictly ETP (and thus BTP ). Secondly, the class
k-BTP is parameterized by a constant k, thus providing a generic version, which
may prove of theoretical interest for general values of k, although in practice we
consider the case k = 3 to be of most interest. Thus, while BTP is defined
for sets of three variables and ETP for sets of four variables, k-BTP is defined
on the basis of sets of k + 1 variables where k is a fixed constant. According
to this approach, BTP = 2-BTP while ETP  3-BTP . Thus, this approach
makes it possible to strictly generalize these two classes. Furthermore, k-BTP
retains some of their interesting properties and practical advantages mentioned
above. Notably, we show that classical algorithms such as MAC or RFL can solve
instances belonging to k-BTP in polynomial time, assuming that these instances
verify Strong k-Consistency [8]. Moreover, we highlight the relationships of this
class with known structural and hybrid classes. We show in particular that the
class of constraint networks whose tree-width is bounded by k is strictly included
in the class k-BTP . This result gives a first answer to a question recently asked by
M. Vardi about the relationships between ETP and the tractable class induced
by instances of bounded tree-width. We also highlight a recent but relatively
unknown result that was proposed by Naanaa [9] whose relationships with k-
BTP we investigate.
In Section 2 we recall the definitions of the tractable classes BTP and ETP . In
Section 3 we define the new class k-BTP and show that instances from this class
can be detected in polynomial time even when the variable order is not known
in advance. Furthermore, we show that, under the extra hypothesis of strong
k-consistency, such instances can be solved in polynomial time, and, in fact,
standard algorithms will solve them. In Section 4 we investigate relationships
between k-BTP and several known tractable classes and in Section 5 we report
results of experimental trials on benchmark problems.
76 M.C. Cooper et al.

2 Background
Formally, a constraint satisfaction problem also called constraint network is
a triple (X, D, C), where X = {x1 , . . . , xn } is a set of n variables, D =
(Dx1 , . . . , Dxn ) is a list of finite domains of values, one per variable, and
C = {c1 , . . . , ce } is a finite set of e constraints. Each constraint ci is a pair
(S(ci ), R(ci )), where S(ci ) = {xi1 , . . . , xik } ⊆ X is the scope of ci , and
R(ci ) ⊆ Dxi1 × · · · × Dxik is its compatibility relation. The arity of ci is |S(ci )|.
In this paper, we only deal with the case of binary CSPs, that is CSPs for which
all the constraints are of arity 2. Hence, we will denote by cij the constraints
involving xi and xj . The structure of a constraint network is represented by a
graph, called the constraint graph, whose vertices correspond to variables and
edges to the constraint scopes. An assignment to a subset Y of X is said to be
consistent if it does not violate any constraint whose scope is included in Y .
We use the notation R(cij )[a] to represent the set of values in Dxj compatible
with a ∈ Dxi . Thus, if there is a constraint with scope {i, j}, then R(cij )[a] =
{b ∈ Dxj |(a, b) ∈ R(cij )}; if there is no constraint with scope {i, j}, then, by
default, R(cij )[a] = Dxj . We recall the BTP property presented in [4].
Definition (BTP). A binary CSP instance (X, D, C) satisfies the Broken Tri-
angle Property (BTP) w.r.t. the variable ordering < if, for all triples of vari-
ables (xi , xj , xk ) s.t. i < j < k, if (vi , vj ) ∈ R(cij ), (vi , vk ) ∈ R(cik ) and
(vj , vk ) ∈ R(cjk ), then either (vi , vk ) ∈ R(cik ) or (vj , vk ) ∈ R(cjk ). If neither
of these two tuples exist, (vi , vj , vk , vk ) is called a broken triangle on xk w.r.t.
xi and xj .
If there exists at least one broken triangle on xk w.r.t. xi and xj , (xi , xj , xk ) is
called a broken triple on xk w.r.t. xi and xj . Let BTP be the set of the instances
for which BTP holds w.r.t. some variable ordering. The BTP property is related
to the compatibility between domain values, which can be graphically visualized
(Figure 1) on the microstructure graph. For example, in Figure 1 (a), there is
a broken triangle on x3 with respect to the variables x1 and x2 since we have
(v1 , v3 ) ∈
/ R(c13 ) and (v2 , v3 ) ∈
/ R(c23 ) while (v1 , v2 ) ∈ R(c12 ), (v1 , v3 ) ∈ R(c13 )
and (v2 , v3 ) ∈ R(c23 ) hold. So (x1 , x2 , x3 ) is a broken triple on x3 w.r.t. x1
and x2 . In contrast, in Figure 1 (b), if one of the two dashed edges (that is
binary tuples) appears in the microstructure, the BTP property holds for all
variable orderings.
Very recently, the property BTP has been relaxed to the Extendable-Triple
Property [7] by considering four variables rather than three, and allowing some
broken triangles.
Definition (ETP). A binary CSP instance P satisfies the Extendable-Triple
Property (ETP) with respect to the variable ordering < if, and only if, for all
subsets of four variables (xi , xj , xk , xl ) such that i < j < k < l, there is at most
one broken triple on xl among (xi , xj , xl ), (xi , xk , xl ) and (xj , xk , xl ).
In this way, a binary CSP can satisfy the ETP property while it contains two
broken triples among (xi , xj , xk , xl ), one on xk , and another one on xl , while
 A Microstructure-Based Family of Tractable Classes for CSPs 77

x1 x3 x1 x3
v3 v3
v1 v’3 v1 v’3

x2 v2 x2 v2
(a) (b)
Fig. 1. A non-BTP instance (a) and a BTP one (b) w.r.t. the order x1 < x2 < x3 if
one of the dashed lines occurs.

none is possible with BTP. So, ETP strictly generalizes BTP since each instance
satisfying BTP satisfies ETP while the reverse is false. So the class of instances
satisfying BTP (denoted BTP ) is strictly included in the class of instances sat-
isfying ETP (denoted ETP ) as indicated in Theorem 1 of [7] (BTP  ETP ).
As in the case of BTP, ETP allows us to define a tractable class but we need to
impose an additional property related to the level of local consistency which must
be verified. While the set of instances satisfying BTP define a tractable class, the
set of instances satisfying ETP must also satisfy Strong-Path-Consistency [8],
that is arc and path-consistency. Nevertheless, such instances have some of the
desirable properties of instances satisfying BTP, e.g. they can be solved in poly-
nomial time by usual algorithms such as MAC or RFL. In the next section, we
introduce a new property which generalizes BTP and ETP.

3 k-BTP: Definition and Properties

In this section, we introduce a new property k-BTP which generalizes previous

work along two axes. First, the property ETP is relaxed in the sense that we
allow more broken triangles than ETP when considering subsets of four variables.
But we also introduce a parameter k ≥ 2 allowing us to consider subsets of k + 1
variables, with k = 2 corresponding to BTP and k = 3 corresponding to a strict
generalization of ETP.
Definition (k-BTP). A binary CSP instance P satisfies the property k-BTP
for a given k (2 ≤ k < n) and with respect to the variable ordering < if, and
only if, for all subsets of k + 1 variables xi1 , xi2 , . . . xik+1 such that i1 < i2 <
· · · < ik−1 < ik < ik+1 , there is at least one triple of variables (xij , xij  , xik+1 )
with 1 ≤ j = j  ≤ k such that there is no broken triangle on xik+1 w.r.t. xij and
xij  . Let k-BTP be the set of the instances for which k-BTP holds w.r.t. some
variable ordering.
One can observe that 2-BTP is exactly BTP while 3-BTP includes ETP. So,
we can immediately extend Theorem 1 of [7] since BTP  ETP  3-BTP . But
above all, a more general result holds, which is an immediate consequence of the
definition of k-BTP:
78 M.C. Cooper et al.

Theorem 1. For all k ≥ 2, k-BTP  (k+1)-BTP

To analyze the tractability of k-BTP, we now show that the instances of this
class can be recognized in polynomial time:
Theorem 2. Given a binary CSP instance P = (X, D, C) and a constant k
with 2 ≤ k < n, there is a polynomial time algorithm to find a variable ordering
< such that P satisfies k-BTP w.r.t. <, or to determine that no such ordering
exists.
Proof: As in the corresponding proofs for BTP [4] and ETP [7], we define a
CSP instance Po which is consistent if and only if a possible ordering exists.
More precisely, this instance has a variable oi with domain {1, . . . , n} per vari-
able xi of X. The value of oi represents the position of the variable xi in
the ordering. We add a constraint involving {oi1 , oi2 , . . . oik , oik+1 } and impos-
ing the condition oik+1 < max(oi1 , oi2 , . . . oik ) for each k+1-tuple of variables
(xi1 , xi2 , . . . xik , xik+1 ) such that each triple of variables (xij , xij  , xik+1 ) with
1 ≤ j = j  ≤ k has at least one broken triangle on xik+1 w.r.t. xij and xij  .
If Po has a solution, then let < be any total ordering of the variables which is
a completion of the partial ordering given by the values of the variables oi . Then
for each k+1-tuple of variables (xi1 , xi2 , . . . xik , xik+1 ), with i1 < . . . < ik+1 ,
we have at least one triple of variables (xij , xij  , xik+1 ) with 1 ≤ j = j  ≤ k
which has no broken triangle on xik+1 w.r.t. xij and xij  . Indeed, if this were
not the case, then the constraint oik+1 < max(oi1 , oi2 , . . . oik ) would have been
imposed, which is in contradiction with i1 < . . . < ik+1 . So, if Po has a solution,
we have an ordering satisfying the k-BTP property. Conversely, let us consider
an ordering satisfying the k-BTP property and assume that Po has no solution.
It means that at least one constraint oik+1 < max(oi1 , oi2 , . . . oik ) is violated.
So each triple of variables (xij , xij  , xik+1 ) with 1 ≤ j = j  ≤ k has at least
one broken triangle on xik+1 , which is impossible since this ordering satisfies the
k-BTP property. Hence Po has a solution if and only if P admits an ordering
satisfying the k-BTP property.
We now prove that Po can be built and solved in polynomial time. Find-
ing all the broken triples can be achieved in O(n3 .d4 ) time, while defining the
constraints oik+1 < max(oi1 , oi2 , . . . oik ) can be performed in O(nk+1 ). So Po
can be computed in O(n3 .d4 + nk+1 ). Moreover, Po can be solved in polyno-
mial time by establishing generalized arc-consistency since its constraints are
max-closed [10]. 2
We analyze now the complexity of solving instances of the class k-BTP
(k ≥ 3). The following theorem shows that this is NP-hard since this is true
even for the smaller class ETP ⊂ 3-BTP .
Theorem 3. Deciding whether an instance of the class ETP is satisfiable is
NP-complete.
Proof: It sufficies to exhibit a polynomial reduction from binary CSP to its sub-
problem ETP . Given any binary CSP instance I, we can construct an equivalent
instance I  by
 A Microstructure-Based Family of Tractable Classes for CSPs 79

1. adding a new variable xij (with domain Dxij = Dxi ) for each constraint cij
2. adding a new equality constraint between each xi and xij
3. replacing each constraint ({xi , xj }, R) by the constraint ({xij , xj }, R).
Let < be any variable order in I  in which all the new variables xij occur after
all the original variables xk . Since each variable is constrained by at most two
variables which precede it in this order, we can easily deduce that I  satisfies
ETP. It follows from this polynomial reduction that deciding whether an instance
of the class ETP is satisfiable is NP-complete. 2
To ensure the tractability of the class k-BTP , we consider an additional
condition which is that instances satisfy Strong k-Consistency [8].
Definition (Strong k-Consistency). A binary CSP instance P satisfies i-
Consistency if any consistent assignment to i − 1 variables can be extended to
a consistent assignment on any ith variable. A binary CSP instance P satisfies
Strong k-Consistency if it satisfies i-Consistency for all i such that 1 < i ≤ k.
Strong k-Consistency and k-BTP allow us to define a new tractable class:
Theorem 4. Let P be a binary CSP instance P such that there exists a constant
k with 2 ≤ k < n for which P satisfies both Strong k-Consistency and k-BTP
w.r.t. the variable ordering <. Then P is consistent and a solution can be found
in polynomial time.
Proof: We consider an ordering for variable assignments corresponding to the
ordering <. As the instance satisfies Strong k-Consistency, it satisfies arc-
consistency and thus, no domain is empty and each value has a support in
each other domain. Moreover, as the instance satisfies Strong k-Consistency, we
have a consistent assignment on the k first variables. Now, and more generally,
suppose that we have a consistent assignment (u1 , u2 , . . . ul−1 , ul ) for the l first
variables x1 , x2 , . . . xl−1 , xl in the ordering, with k ≤ l < n. We show that this
assignment can be consistently extended to the variable xl+1 . To show this, we
must prove that ∩1≤i≤l R(cil+1 )[ui ] = ∅, that is there is at least one value in the
domain of xl+1 which is compatible with the assignment (u1 , u2 , . . . ul−1 , ul ).
We first prove this for l = k. Consider the consistent assignment
(u1 , u2 , . . . uk−1 , uk ) on the k first variables. Consider a k + 1th variable xk+1
appearing later in the ordering. Since P satisfies k-BTP , there exists at least
one triple of variables (xj , xj  , xk+1 ) with 1 ≤ j = j  ≤ k such that there is no
broken triangle on xk+1 w.r.t. xj and xj  . By Lemma 2.4 given in [4], we have:
(R(cjk+1 )[uj ] ⊆ R(cj  k+1 )[uj  ])
or
(R(cj  k+1 )[uj  ] ⊆ R(cjk+1 )[uj ])
Without loss of generality, assume that we have R(cjk+1 )[uj ] ⊆ R(cj  k+1 )[uj  ]
and j < j  . Since P satisfies Strong k-Consistency, we know that the sub-
assignment of (u1 , u2 , . . . , uj , . . . uk−1 , uk ) on k−1 variables excluding the assign-
ment uj  for xj  can be consistently extended to xk+1 . Moreover, we know that
80 M.C. Cooper et al.

R(cjk+1 )[uj ] ⊆ R(cj  k+1 )[uj  ] and by arc-consistency, R(cij ik+1 )[uj ] = ∅. Thus,
(u1 , u2 , . . . , uj , . . . , uj  , . . . , uk , uk+1 ) is a consistent assignment to the k + 1 first
variables.
Note that this proof holds for all subsets of k + 1 variables such that
xk+1 appears later in the ordering <, not only for the k + 1 first variables
x1 , x2 , . . . xk−1 , xk and xk+1 .
Now, we prove the property for l with k < l < n. That is, we show that a
consistent assignment (u1 , u2 , . . . ul−1 , ul ) can be extended to a (l +1)th variable.
As induction hypothesis, we assume that every consistent assignment on l − 1
variables can be extended to a lth variable, which appears later in the considered
ordering <.
Consider a consistent assignment (u1 , u2 , . . . ul−1 , ul ) on the l first vari-
ables. Let (ui1 , ui2 , . . . uik ) be a sub-assignment on k variables of the assign-
ment (u1 , u2 , . . . ul−1 , ul ). As P satisfies k-BTP, and as k < l < n, for all
subsets of k variables xi1 , xi2 , . . . xik , we know that there is a triangle which
is not broken in xl+1 w.r.t. xij and xij  , with xij and xij  appearing in the
variables xi1 , xi2 , . . . xik . So, without loss of generality, we can consider that
i1 ≤ ij < ij  ≤ ik ≤ l and we have R(cij l+1 )[uij ] ⊆ R(cij  l+1 )[uij  ]. Note that
xij and xij  can be interchanged in the ordering if necessary.
Now, consider the consistent assignment (u1 , u2 , . . . ul−1 , ul ) on the l
first variables. By the induction hypothesis, each partial assignment of
(u1 , u2 , . . . ul−1 , ul ) on l − 1 variables can be extended to a consistent assign-
ment on xl+1 with a compatible value ul+1 . So, consider the partial assign-
ment on l − 1 variables where uij  does not appear. This assignment is
for example (u1 , u2 , . . . uij , . . . ul−1 , ul , ul+1 ). As we have R(cij l+1 )[uij ] ⊆
R(cij  l+1 )[uij  ], the value uij  is also compatible with ul+1 , and thus the assign-
ment (u1 , u2 , . . . uij , . . . uij  , . . . ul−1 , ul , ul+1 ) on the l + 1 first variables is a con-
sistent assignment.
So, every consistent assignment (u1 , u2 , . . . ul−1 , ul ) on (x1 , x2 , . . . xl−1 , xl )
can be extended to a (l + 1)th variable, for all l with k < l < n. And more gener-
ally, we have shown that every consistent assignment on l variables, not necessar-
ily consecutive in the ordering (as are the l first variables), can be extended to a
consistent assignment for every (l+1)th variable which appears after these l vari-
ables in the ordering < associated with k-BTP. Thus, the induction hypothesis
holds for the next step.
Note that this proof also shows that an instance which satisfies Strong
k-Consistency and k-BTP (with respect to the variable ordering <) is consistent.
Finally, given the ordering <, we show that finding a solution can be per-
formed in polynomial time. Given a consistent assignment (u1 , u2 , . . . ul ) with
l < n, finding a compatible value ul+1 for the next variable xl+1 is feasible by
searching in its domain whose size is at most d. For each value, we need to ver-
ify the constraints connecting the variable xl+1 which can be done in O(el+1 )
if the next variable xl+1 has el+1 neighbors in the previous variables. Since
Σ1≤l<n el+1 = e, the total cost to find a solution is O((n + e).d). 2
 A Microstructure-Based Family of Tractable Classes for CSPs 81

In the sequel, we denote k-BTP -SkC , the class of instances satisfying k-

BTP and Strong k-Consistency. One of the most interesting properties of the
tractable class BTP is the fact that the instances of this class can be solved in
polynomial time using classical algorithms (such as MAC or RFL) implemented
in most solvers. The next property establishes that a similar result holds for
k-BTP -SkC . Indeed, the proof of Theorem 4 allows us to show that algorithms
such as BT (Backtracking), MAC and RFL can solve any instance of the class
k-BTP -SkC in polynomial time:
Theorem 5. Given a binary CSP instance P = (X, D, C) and a variable order-
ing < such that P satisfies k-BTP w.r.t. <, and is Strongly-k-Consistent, the
algorithms BT, MAC and RFL find a solution of the instance P in polynomial
time.
Proof: As the instance satisfies Strong k-Consistency, BT using the ordering <
for the variable assignment can find a consistent assignment on x1 , x2 , . . . xk−1
and xk . Moreover, given l, with k < l < n, it is shown in the proof of Theorem 4
that a consistent assignment (u1 , u2 , . . . ul−1 , ul ) on x1 , x2 , . . . xl−1 and xl can be
extended to a (l + 1)th variable, that is on xl+1 . To find the assignment of xl+1 ,
we need to look for a compatible value in its domain. This is feasible in O(el+1 .d)
assuming that xl+1 has el+1 neighbors in the previous variables. So, as for the
proof of Theorem 4, finding a solution of P is globally feasible in O((n + e).d).
If we consider now algorithms such as MAC or RFL, by the same reasoning, we
show that their complexity is bounded by O(n.(n + e).d2 ) due to the additional
cost of the arc-consistency filtering performed after each variable assignment. 2
In Section 5, we discuss the interest of the class k-BTP from a practical
viewpoint. In the next section, we study the relationships between k-BTP and
some tractable classes.

4 Relationship with Some Tractable Classes

We consider the important tractable class based on the notion of tree-
decomposition of graphs [11].
Definition (Tree-Decomposition). Given a graph G = (X, C), a tree-
decomposition of G is a pair (E, T ) with T = (I, F ) a tree and E = {Ei : i ∈ I}
a family of subsets of X, such that each subset (called cluster or bag in Graph
Theory) Ei is a node of T and satisfies:
1. ∪i∈I Ei = X,
2. for each edge {x, y} ∈ C, there exists i ∈ I with {x, y} ⊆ Ei , and
3. for all i, j, k ∈ I, if k is in a path from i to j in T , then Ei ∩ Ej ⊆ Ek .
The width of a tree-decomposition (E, T ) is equal to maxi∈I |Ei | − 1. The tree-
width w of G is the minimal width over all the tree-decompositions of G.
Let k-TW be the class of binary CSPs instances such that their tree-width
is less than or equal to a constant k. Recently, M. Vardi asked a question about
82 M.C. Cooper et al.

the relationships between k-TW and ETP or other generalizations of BTP . The
next theorems give a first partial answer to this question.
Theorem 6. k-TW  (k + 1)-BTP .
Proof: We show firstly that k-TW ⊆ (k + 1)-BTP . It is well known that if the
tree-width of a binary instance of CSP is bounded by k, there is an ordering
< on variables, such that for xi ∈ X, |{xj ∈ X : j < i and cji ∈ C}| ≤ k [2].
Now, consider a subset of k + 2 variables xi1 , xi2 , . . . xik , xik+1 , xik+2 such that
i1 < i2 < · · · < ik−1 < ik < ik+1 < ik+2 . Since the tree-width is bounded
by k, we know that there are at most k constraints cij ik+2 ∈ C. So, there is
at least one triple of variables (xij , xij  , xik+2 ) with 1 ≤ j = j  ≤ k such that
cij ik+2 ∈
/ C or cij  ik+2 ∈
/ C. Without loss of generality, assume that there is no
constraint cij ik+2 ∈ C. Thus, there is no broken triangle on xik+2 w.r.t. xij and
xij  because all the values of Dxij are compatible with all the values of Dxik+2 .
So, the considered instance of CSP satisfies the property (k + 1)-BTP . Finally,
it is easy to define instances whose tree-width is strictly greater than k which
satisfy the property (k + 1)-BTP . For example, we can consider an instance of
CSP with domains of size one, with the complete constraint graph, and with one
solution. The tree-width of this instance is n − 1 while it satisfies k-BTP for all
possible values of k. 2
The cost of checking for satisfiability of instances in k-TW has a similar
cost to that of achieving Strong (k+1)-Consistency, that is O(nk+1 dk+1 ). Nev-
ertheless, this does not allow us to establish a formal inclusion of k-TW in
(k+1)-BTP -S(k+1)C which is tractable while (k+1)-BTP is NP-complete for
k ≥ 2 by Theorem 3. But if we denote k-TW -S(k+1)C, the class of binary CSPs
instances belonging to k-TW and which satisfy Strong (k+1)-Consistency, the
next result holds:
Theorem 7. k-TW -S(k+1)C  (k + 1)-BTP -S(k+1)C.
The tractable class BTP has also recently been generalized in a different
way to that proposed in this paper, again by noticing that not all broken trian-
gles need to be forbidden [12]. We will show that these two generalizations are
orthogonal.
Definition (∀∃-BTP). A binary CSP instance P satisfies the property ∀∃-BTP
w.r.t. the variable ordering < if, and only if, for each pair of variables xi , xk such
that i < k, for all vi ∈ Dxi , ∃vk ∈ Dxk such that (vi , vk ) ∈ R(cik ) and for all xj
with j < k and j = i, and for all vj ∈ Dxj and for all vk ∈ Dxk , (vi , vj , vk , vk )
is not a broken triangle on xk w.r.t. xi and xj . Let ∀∃-BTP be the set of the
instances for which ∀∃-BTP holds w.r.t. some variable ordering.
The class ∀∃-BTP can be solved and recognized in polynomial time [12]. It
represents a tractable class which strictly includes BTP since it does not forbid
all broken triangles. Since k-BTP also does not forbid all broken triangles, it is
natural to compare these two classes. We do this for the special case k = 3, but
the same argument applies for any value of k ≥ 3.
 A Microstructure-Based Family of Tractable Classes for CSPs 83

Theorem 8. Even for sets of binary CSP instances which are strong path con-
sistent, the properties 3-BTP and ∀∃-BTP are incomparable.

Proof: Consider an instance P ∗ in which each domain Dxk contains a value a∗

such that for all other variables xi , for all values vi ∈ Dxi , (vi , a∗ ) ∈ R(cik ).
Then P ∗ satisfies ∀∃-BTP since there can be no broken triangle of the form
(vi , vj , a∗ , vk ), the value a∗ being compatible with all assignments to all other
variables. It is easy to complete such an instance P ∗ so that it does not satisfy
3-BTP for any variable ordering by adding broken triangles on domain elements
other than a∗ .
Consider a 3-variable binary CSP instance P3 with domains {0, 1, 2} and
the following three constraints: x1 = x2 , x1 = x3 , x2 = x3 , i.e. a 3-colouring
problem on a complete graph on three vertices. Then P3 is strong path consistent
and trivially satisfies 3-BTP (since there are only 3 variables), but P3 does not
satisfy ∀∃-BTP for any ordering i < j < k of the variables (due to the existence
of broken triangles on assignments (xi , a), (xj , b), (xk , a), (xk , b) for all pairs of
distinct colours a, b). 2
We now consider a very general tractable class recently discovered by
Naanaa [9] and which undoubtedly deserves to be better known.
Let E be a finite set and let {Ei }i∈I be a finite family of subsets of E. The
family {Ei }i∈I is said to be independent if and only if for all J ⊂ I,
 
Ei ⊂ Ej
i∈I j∈J

(where the notation A ⊂ B means that A is a proper subset of B). Observe that
{Ei }i∈I cannot be independent if ∃j = j  ∈ I such that Ej ⊆ Ej  , since in this
case and with J = I \ {j  } we would have
 
Ei = Ej .
i∈I j∈J

Definition (Directional Rank). Let P be a binary CSP instance whose vari-

ables are totally ordered by <. The directional rank of variable xm is the size k
of the largest consistent assignment (a1 , . . . , ak ) to a set of variables xi1 , . . . , xik
(with i1 < . . . < ik < m) such that the family of sets {R(cij m )[aj ]}j=1,...,k is
independent. The directional rank of P (w.r.t the ordering < of its variables) is
the maximum directional rank over all its variables.
Naanaa has shown that if P is a binary CSP instance which has directional
rank no greater than k and is directional strong (k + 1)-consistent then I is
globally consistent [9]. We denote DR-k, the set of these instances. Naanaa
points out that some known tractable classes, such as binary CSP instances
with connected row convex constraints [13], have bounded directional rank.
If a binary CSP instance P is (k + 1)-BTP, then no variable can have
a directional rank greater than k. This is because for any variable xm and
any assignments (a1 , . . . , ak+1 ) to any set of variables xi1 , . . . , xik+1 with
84 M.C. Cooper et al.

i1 < . . . < ik+1 < m, by the definition of (k+1)-BTP, we must have R(cij m )[aj ] ⊆
R(cij  m )[aj  ] for some j = j  ∈ {1, . . . , k + 1}; hence, as observed above, the sets
{R(cij m )[aj ]}j=1,...,k+1 cannot be independent. It follows that the tractability of
(k+1)-BTP -S(k+1)C is also a corollary of the result of Naanaa [9]. On the other
hand, the property (k + 1)-BTP, although subsumed by DR-k, can be detected in
time complexity O(nk dk + n3 d4 ) compared to O(nk+1 dk+1 ) for DR-k.

5 Experiments
In this section, we compare the tractable classes BTP , ETP -SPC , k-BTP -SkC
and DR-k-1 (where SPC stands for Strong Path Consistency) from a practical
viewpoint. We only consider the case k = 3, since strong k-consistency becomes
too expensive in time for k > 3 and may add constraints of arity k − 1.
Tractable classes are often critized for being artificial in the sense that their
underlying properties seldom occur in real instances. So, here, we first highlight
the existence of instances belonging to some of these classes among the bench-
mark instances classically exploited for solver evaluations and comparisons. More
precisely, our experiments involve 2,373 binary benchmarks from the third CSP
Solver Competition1 and cover all the benchmarks exploited in [7].
Then we will investigate the possible link between efficient solving and
belonging to these tractable classes.

5.1 Instances Belonging to Tractable Classes

Since the tractable classes ETP -SPC , 3-BTP -SPC and DR-2 require strong
path-consistency, we first achieve SPC on each instance before checking whether
it belongs to the considered classes, in the same spirit as [14,15]. In so doing, 628
instances are detected as inconsistent and so they trivially belong to all of these
tractable classes. 85 of the remaining instances belong to 3-BTP -SPC while 87
have directional rank at most two. Among these instances, we have respectively
71 and 76 instances in BTP -SPC and ETP -SPC . These differences between
these tractable classes are well highlighted by some instances of the bqwh-
15-106 family since we can observe all the possible configurations of the rela-
tions BTP -SPC  ETP -SPC  3-BTP -SPC  DR-2. For example, instance
bqwh-15-106-13 belongs to all the considered tractable classes while instances
bqwh-15-106-28, bqwh-15-106-16 and bqwh-15-106-76 only belong respectively
to three, two or one of these tractable classes. Table 1 presents some instances
belonging to classes ETP -SPC , 3-BTP -SPC or DR-2. It also provides the
tree-width w of these instances and their tree-width w once SPC is enforced.
When the exact tree-width is unknown (recall that computing an optimal tree-
decomposition is an NP-hard problem), we give a range. We can note the diver-
sity of these instances (academic, random or real-world instances). Some of these
instances belong to 3-BTP -SPC or DR-2 thanks to their structure. For instance,

1
See http://www.cril.univ-artois.fr/CPAI08.
 A Microstructure-Based Family of Tractable Classes for CSPs 85

Table 1. Some instances belonging to BTP -SPC , ETP -SPC , 3-BTP -SPC or DR-2
after the application of SPC with their tree-width w and the tree-width w of the
instances once SPC is enforced.

Instance n w w BTP -SPC ETP -SPC 3-BTP -SPC DR-2

bqwh-15-106-13 106 [7, 48] 104 yes yes yes yes
bqwh-15-106-16 106 [6, 45] 99 no no yes yes
bqwh-15-106-28 106 [7, 52] 105 no yes yes yes
bqwh-15-106-76 106 [6, 44] 100 no no no yes
bqwh-15-106-77 106 [7, 50] 100 no no yes yes
bqwh-18-141-33 141 [7, 64] 134 yes yes yes yes
bqwh-18-141-57 141 [7, 66] 137 yes yes yes yes
domino-100-100 100 2 2 yes yes yes yes
domino-5000-500 5000 2 2 yes yes yes yes
driverlogw-04c-sat 272 [19, 56] [214, 221] no no no yes
driverlogw-09-sat 650 [39, 108] 629 yes yes yes yes
fapp17-0300-10 300 [6, 153] [6, 154] yes yes yes yes
fapp18-0350-10 350 [5, 192] [12, 199] yes yes yes yes
fapp23-1800-9 1800 [6, 1325] [41, 1341] yes yes yes yes
graph12-w0 680 1 1 yes yes yes yes
graph13-w0 916 1 1 yes yes yes yes
hanoi-7 126 1 1 yes yes yes yes
langford-2-4 8 7 7 yes yes yes yes
lard-83-83 83 82 82 no no yes yes
lard-91-91 91 90 90 no no yes yes
os-taillard-4-100-0 16 [3, 9] 15 yes yes yes yes
os-taillard-4-100-9 16 [3, 9] 15 yes yes yes yes
scen5 400 [11, 32] [167, 188] no no yes yes

graph12-w0 and hanoi-7 have an acyclic constraint graph while the tree-width of
domino-100-100 and crossword-m1-uk-puzzle01 is two. However, most instances
have a tree-width greater than two. Moreover, in most cases, the application
of SPC may significantly increase the original tree-width of these instances. For
example, the tree-width of instance driverlogw-09-sat is initially bounded by 108
and is equal to 629 after the application of SPC. This increase is explained by
the pairs of values which are forbidden by SPC. When SPC forbids a pair of
values (vi , vj ) for a given pair of variables (xi , xj ), it removes (vi , vj ) from the
relation R(cij ) if the constraint cij exists. However, if the constraint cij does not
exist yet, SPC must first add it to the problem. In such a case, depending on the
added constraints and their number, the tree-width may significantly increase.
Note that the considered instances whose tree-width is initially at most two have
a tree-width unchanged by the application of SPC.

5.2 Link Between Efficient Solving and Belonging

to Tractable Classes
In this subsection, our aim is not to provide a new module based on tractable
classes in order to improve the efficiency of solvers but to see whether we
86 M.C. Cooper et al.

can exploit some tractable classes to explain the efficiency of solvers on some
instances. Indeed, we think that tractable classes are more useful from a practi-
cal viewpoint if they are implicitly handled by classical solvers than by ad-hoc
methods (as is generally the case). For instance, it is well kwown that MAC
can solve in backtrack-free manner any binary CSP whose constraint network is
acyclic without knowing that the instance has this particular feature [16].
Most state-of-the-art solvers rely on variants of MAC or RFL algorithms. In
the following, we focus our study on MAC but we have observed similar results
for RFL.
As far as solving is concerned, all the instances belonging to 3-BTP -SPC
or DR-2 are solved in a backtrack-free manner by MAC except the instance
driverlogw-04c-sat which needs one backtrack. Note that MAC has no knowledge
about the variable ordering used to satisfy 3-BTP or to obtain a directional rank
of at most two. In most cases, we have observed that the ordering CSP instance
built in the proof of Theorem 2 in order to compute a suitable variable ordering
has no constraints. So any variable ordering is suitable. In contrast, for about
a dozen instances, this CSP has several constraints but remains clearly under-
constrained and the constraint network has several connected components. This
ensues that the ordering CSP in general a huge number of solutions. So it is very
likely that MAC exploits implicitly one of these suitable variable orderings. For
example, the ordering CSP for checking whether the bqwh-15-106-76 instance
(which has 106 variables) has a directional rank at most two has 65 connected
components and admits more than 33 million solutions.
Some of the instances are solved efficiently by MAC in a backtrack-free man-
ner even though they do not belong to one of the studied tractable classes. Hence,
we now consider the notion of backdoor [17] with the aim in view to provide some
explanation about this efficiency in the same spirit as [7]. A backdoor is a set of
variables defined with respect to a class such that once the backdoor variables
are assigned, the problem falls in the class. Here, we are interested in back-
doors which are discovered implicitly by MAC when it assigns some variables.
Indeed, after some assignments and the associated filtering, the remaining part
of the problem may become tractable. So we assess the number of variables which
must be assigned before MAC finds implicitly a backdoor w.r.t. one of the studied
classes. In practice, over the 50 considered instances, we observe that MAC finds
a backdoor w.r.t. BTP after having assigned more variables than for the other
considered classes. The numbers of assigned variables required to find a backdoor
respectively for ETP and 3-BTP are very close, and even equal in most cases.
By considering DR-2, we save a few variables compared to ETP and 3-BTP . For
example, MAC needs to assign at most five variables before finding a backdoor
w.r.t. to 3-BTP or DR-2 for 14 instances compared to 12 and 4 instances, respec-
tively, for ETP and BTP 2 . Of course, the resulting instances do not necessarily
satisfy strong path-consistency and so we cannot exploit directly Theorem 5 to
explain the efficiency of MAC. Nevertheless, when the instance is 3-BTP and

2
Note that these instances do not include all the instances mentioned in [7] since some
of them belong to 3-BTP -SPC and/or DR-2.
 A Microstructure-Based Family of Tractable Classes for CSPs 87

strong path-consistent after having assigned some variables, MAC may exploit
implicitly a suitable variable ordering since, as evoked above, the corresponding
ordering CSP often admits a large number of solutions. Furthermore Theorem 5
provides sufficient conditions so that MAC solves some instances in polynomial
time, but these conditions are not always necessary. For instance, MAC solves
the instances which belong to BTP in polynomial time without requiring a suit-
able variable ordering or the satisfaction of strong path-consistency. Hence, one
part of the explanation of the practical efficiency of MAC may lie in its ability
to exploit implicitly different tractable classes.

6 Conclusion
This paper introduces a novel family of tractable classes for binary CSPs, denoted
k-BTP whose tractability is associated with a given level of strong k-consistency.
It is based on a hierarchy of classes of instances with the BTP class as the
base case. While BTP is defined on subsets of 3 variables, the k-BTP class is
defined on sets of k+1 variables, while relaxing the restrictive conditions imposed
by BTP which is the class 2-BTP . We showed that k-BTP inherits some of
the desirable properties of BTP, such as polynomial solvability using standard
algorithms such as MAC. We also showed that k-BTP strictly generalizes the
class of instances whose tree-width is bounded by a constant and we analyzed
the relationships with the class based on the notion of directional rank recently
introduced by Naanaa. To assess the practical interest of the k-BTP class, an
experimental analysis is presented focusing on the particular case of 3-BTP .
This analysis shows a significant advantage of 3-BTP compared to BTP and to
CSPs of bounded tree-width.
Further research is required to determine if the condition corresponding to
strong k-consistency is actually necessary or whether a weaker condition would
suffice. Indeed, experiments showed that MAC can solve without backtracking
certain instances belonging to 3-BTP even when they do not verify the cor-
responding level of consistency. From a practical point of view, an interesting
challenge is to find the minimum (generally) required level of consistency among
different kinds of local consistencies such as PIC [18], maxRPC [19] or SAC [20].
Note that, from a theoretical point of view, we can easily deduce from Theo-
rem 3 that any local consistency that only performs domain filtering (e.g. PIC,
maxRPC, SAC) cannot be sufficient (assuming P=NP) since ETP is invariant
under domain filtering operations.
Moreover, studying a relaxation of the k-BTP condition needs to be
addressed so as to further expand the class of instances that can be solved in
polynomial time, but along different avenues to the one proposed in [9], even if
further theoretical and experimental research are clearly required to fully appre-
ciate all the consequences of Naanaa’s result. Finally, it could be interesting
to investigate a similar approach to the one introduced in [21] which provides
a novel polynomial-time reduction operation based on the merging of domain
values.
88 M.C. Cooper et al.

References
1. Rossi, F., van Beek, P., Walsh, T.: Handbook of Constraint Programming. Elsevier
(2006)
2. Dechter, R., Pearl, J.: Tree-Clustering for Constraint Networks. Artificial Intelli-
gence 38, 353–366 (1989)
3. Gottlob, G., Leone, N., Scarcello, F.: A Comparison of Structural CSP Decompo-
sition Methods. Artificial Intelligence 124, 243–282 (2000)
4. Cooper, M.C., Jeavons, P., Salamon, A.: Generalizing constraint satisfaction on
trees: hybrid tractability and variable elimination. Artificial Intelligence 174,
570–584 (2010)
5. Sabin, D., Freuder, E.: Contradicting conventional wisdom in constraint satisfac-
tion. In: Proceedings of ECAI, pp. 125–129 (1994)
6. Nadel, B.: Tree Search and Arc Consistency in Constraint-Satisfaction Algorithms.
Search in Artificial Intelligence, pp. 287–342. Springer-Verlag (1988)
7. Jégou, P., Terrioux, C.: The extendable-triple property: a new CSP tractable class
beyond BTP. In: Proceedings of AAAI, pp. 3746–3754 (2015)
8. Freuder, E.: A Sufficient Condition for Backtrack-Free Search. Journal of the ACM
29(1), 24–32 (1982)
9. Naanaa, W.: Unifying and extending hybrid tractable classes of csps. Journal of
Experimental and Theoretical Artificial Intelligence 25(4), 407–424 (2013)
10. Jeavons, P., Cooper, M.: Tractable constraints on ordered domains. Artificial Intel-
ligence 79(2), 327–339 (1995)
11. Robertson, N., Seymour, P.D.: Graph minors II: Algorithmic aspects of treewidth.
Algorithms 7, 309–322 (1986)
12. Cooper, M.C.: Beyond consistency and substitutability. In: O’Sullivan, B. (ed.) CP
2014. LNCS, vol. 8656, pp. 256–271. Springer, Heidelberg (2014)
13. Deville, Y., Barette, O., van Hentenryck, P.: Constraint satisfaction over connected
row convex constraints. Artificial Intelligence 109(1–2), 243–271 (1999)
14. El Mouelhi, A., Jégou, P., Terrioux, C.: Hidden tractable classes: from theory to
practice. In: Proceedings of ICTAI, pp. 437–445 (2014)
15. El Mouelhi, A., Jégou, P., Terrioux, C.: Hidden Tractable Classes: from Theory to
Practice. Constraints (2015)
16. Sabin, D., Freuder, E.: Understanding and Improving the MAC Algorithm. In:
Smolka, G. (ed.) CP 1997. LNCS, vol. 1330, pp. 167–181. Springer, Heidelberg
(1997)
17. Williams, R., Gomes, C.P., Selman, B.: Backdoors to typical case complexity. In:
Proceedings of IJCAI, pp. 1173–1178 (2003)
18. Freuder, E., Elfe, C.D.: Neighborhood inverse consistency preprocessing. In:
Proceedings of AAAI, pp. 202–208 (1996)
19. Debruyne, R., Bessière, C.: From restricted path consistency to max-restricted
path consistency. In: Smolka, G. (ed.) CP 1997. LNCS, vol. 1330, pp. 312–326.
Springer, Heidelberg (1997)
20. Debruyne, R., Bessière, C.: Domain Filtering Consistencies. Journal of Artificial
Intelligence Research 14, 205–230 (2001)
21. Cooper, M.C., El Mouelhi, A., Terrioux, C., Zanuttini, B.: On broken triangles.
In: O’Sullivan, B. (ed.) CP 2014. LNCS, vol. 8656, pp. 9–24. Springer, Heidelberg
(2014)
 The Unary Resource with Transition Times

Cyrille Dejemeppe, Sascha Van Cauwelaert(B) , and Pierre Schaus

UCLouvain, ICTEAM, Place Sainte Barbe 2, 1348 Louvain-la-Neuve, Belgium

{cyrille.dejemeppe,sascha.cauwelaert,pierre.schaus}@uclouvain.be

Abstract. Transition time constraints are ubiquitous in scheduling

problems. They are said to be sequence-dependent if their durations
depend on both activities between which they take place. In this context,
we propose to extend the Θ-tree and Θ-Λ-tree data structures introduced
by Vilı́m in order to strengthen the bound computation of the earliest
completion time of a set of activities, by taking into account the sequence
dependent transition time constraints. These extended structures can
be substituted seamlessly in the state-of-the-art Vilı́m’s filtering algo-
rithms for unary resource constraints (Overload Checking, Detectable
Precedences, Not-First/Not-Last and Edge-Finding algorithms) with-
out changing their O(n log(n)) time complexities. Furthermore, this new
propagation procedure is totally independent from additional constraints
or the objective function to optimize. The proposed approach is able to
reduce the number of nodes by several order of magnitudes on some
instances of the job-shop with transition times problem, without intro-
ducing too much overhead on other instances for which it is less effective.

Keywords: Scheduling · Transition times · Global constraints ·

Constraint Programming

1 Introduction
This work extends the classic unary/disjunctive resource propagation algorithms
to include propagation over sequence-dependent transition times between activ-
ities. A wide range of real-world scheduling problems from the industry involves
transition times between activities. An example is the quay crane scheduling
problem in container terminals [21] where the crane is modeled as a unary
resource and transition times represent the moves of the crane on the rail to
move from one position to another along the vessel to load/unload containers.
We introduce filtering algorithms to tighten the bounds of (non-preemptive)
activities while taking into account the transition times between them. These
filtering algorithms are extensions of the unary resource propagation algo-
rithms (Overload Checking, Detectable Precedences, Not-First/Not-Last, Edge-
Finding) introduced in [18]. All these algorithms rely on an efficient computation
of the earliest completion time (ect) of a group of activities using the so-called
Theta tree and Theta-Lambda tree data structures. We demonstrate the effi-
ciency of the filtering on job-shop with transition times problem instances.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 89–104, 2015.
DOI: 10.1007/978-3-319-23219-5 7
90 C. Dejemeppe et al.

In Section 2, we give an overview of the tackled problems and of current

state-of-the-art techniques to solve them. In Section 3, we explain the require-
ments needed to integrate transition times propagation. Section 4 explains how
to obtain lower bounds for the time spent by transitions between activities from a
set. Then, Section 5 describes how to integrate this bound to efficiently compute
the ect of a set of activities with extended Θ-tree structures. Section 6 then
explains how classic unary algorithms can consider transition times by using
the extended Θ-tree structures. Finally, we report results obtained by the new
propagation procedure in Section 7.

2 Background
In Constraint Programming (CP), a scheduling problem is modeled by associat-
ing three variables to each activity Ai : starti , endi , and durationi representing
respectively the starting time, ending time and processing time of Ai . These
variables are linked together by the following relation:

starti + durationi = endi

Depending on the considered problem, global constraints linking the activity

variables are added to the model. In this work, we are interested in the unary
resource constraint. A unary resource, sometimes referred to as a machine, is a
resource allowing only a single activity to use it at any point in time. As such,
all activities demanding the same unary resource cannot overlap in time:

∀i, j i = j : (endi ≤ startj ) ∨ (endj ≤ starti )

The unary resource can be generalized by requiring transition times between

activities. A transition time tt i,j is a minimal amount of time that must occur
between two activities Ai and Aj if Ai ≺ Aj (precedes). These transition times
are described in a matrix M in which the entry at line i and column j represents
the minimum transition time between Ai and Aj , tt i,j . We assume that transition
times respect the triangular inequality. That is, inserting an activity between two
activities always increases the time between these activities:

∀i, j, k i = j = k : tt i,j ≤ tt i,k + tt k,j

The unary resource with transition times imposes the following relation:

∀i, j : (endi + tt i,j ≤ startj ) ∨ (endj + tt j,i ≤ starti ) (1)

2.1 Related Work

As described in a recent survey [2], scheduling problems with transition times
can be classified in different categories. First the activities can be in batch
(i.e. a machine allows several activities of the same batch to be processed
simultaneously) or not. Transition times may exist between successive batches.
 The Unary Resource with Transition Times 91

A CP approach for batch problems with transition times is described in [18]. Sec-
ondly the transition times may be sequence-dependent or sequence-independent.
Transition times are said to be sequence-dependent if their durations depend on
both activities between which they occur. On the other hand, transition times
are sequence-independent if their duration only depend on the activity after
which it takes place. The problem category we study in this article is non-batch
sequence-dependent transition time problems.
Several methods have been proposed to solve such problems. Ant Colony
Optimization (ACO) approaches were used in [9] and [15] while [6], [4], [13] and
[10] propose Local Search and Genetic Algorithm based methods. [13] proposes
a propagation procedure with the Iterative Flattening Constraint-Based Local
Search technique. The existing CP approaches for solving sequence-dependent
problems are [8], [3], [20] and [11].
Focacci et al [8] introduce a propagator for job-shop problems involving
alternative resources with non-batch sequence-dependent transition times. In
this approach a successor model is used to compute lower-bounds on the total
transition time. The filtering procedures are based on a minimum assignment
algorithm (a well known lower bound for the Travelling Salesman Problem). In
this approach the total transition time is a constrained variable involved in the
objective function (the makespan).
In [3], a Travelling Salesman Problem with Time Window (TSPTW) relax-
ation is associated to each resource. The activities used by a resource are rep-
resented as vertices in a graph and edges between vertices are weighted with
corresponding transition times. The TSPTW obtained by adding time windows
to vertices from bounds of corresponding activities is then resolved. If one of the
TSPTW is found un-satisfiable, then the corresponding node of the search tree
is pruned. A similar technique is used in [5] with additional propagation.
In [20], an equivalent model of multi-resource scheduling problem is proposed
to solve sequence-dependent transition times problems. Finally, in [11], a model
with a reified constraint for transition times is associated to a specific search to
solve job-shop with sequence-dependent transition times problems.
To the best of our knowledge, the CP filtering introduced in this article is
the first one proposing to extend all the classic filtering algorithms for unary
resources (Overload Checking [7], Detectable Precedences [17], Not-First/Not-
Last [19] and Edge Finding [19]) by integrating transition times, independently of
the objective function of the problem. This filtering can be used in any problem
involving a unary resource with sequence-dependent transition times.

2.2 Unary Resource Propagators in CP

The earliest starting time of an activity Ai denoted est i , is the time before which
Ai cannot start. The latest starting time of Ai , lst i , is the time after which Ai
cannot start. The domain of starti is thus the interval [est i ; lst i ]. Similarly the
earliest completion time of Ai , ect i , is the time before which Ai cannot end and
the latest completion time of Ai , lct i , is the time after which Ai cannot end.
92 C. Dejemeppe et al.

The domain of endi is thus the interval [ect i ; lct i ]. These definitions can be
extended to a set of activity Ω. For example, est Ω is defined as follows:

est Ω = min {est j |j ∈ Ω}

The propagation procedure for the unary resource constraint introduced

in [18] contains four different propagation algorithms all running with time com-
plexity in O(n log(n)) : Overload Checking (OC), Detectable Precedences (DP),
Not-First/Not-Last (NF/NL) and Edge Finding (EF). These propagation algo-
rithms all rely on an efficient computation of the earliest completion time of a
set of activities Ω using data structures called Theta Tree and Theta-Lambda
Tree introduced in [18]. Our contribution is a tighter computation of the lower
bound ect Ω taking into account the transition times between activities.

3 Transition Times Extension Requirements

The propagation procedure we introduce in this article relies on the computation
of ect Ω , the earliest completion time of a set of activities. This value depends on
the transition times between activities inside Ω. Let ΠΩ be the set of all possible
permutations of activities in Ω. For a given permutation π ∈ ΠΩ (where π(i)
is the activity taking place at position i), we can define the total time spent by
transition times, tt π , as follows:
|Ω|−1

tt π = tt π(i),π(i+1)
i=1

A lower bound for the earliest completion time of Ω can then defined as:
 
ect NP
Ω = max

est Ω  + pΩ  + min tt π (2)
Ω ⊆Ω π∈ΠΩ 

Unfortunately, computing this value is NP-hard. Indeed, computing the optimal

permutation π ∈ Π minimizing tt π is equivalent to solving a TSP. Since embed-
ding an exponential algorithm in a propagator is generally impractical, a looser
lower bound can be used instead:

ect Ω = max

{est Ω  + pΩ  + tt(Ω  )}
Ω ⊆Ω

where tt(Ω  ) is a lower bound of the total time consumed by transition times
between activities in Ω  :
tt(Ω  ) ≤ min tt π
π∈ΠΩ 

Our goal is to keep the overall O(n log(n)) time complexity of Vilı́m’s algo-
rithms. The lower bound tt(Ω  ) must therefore be available in constant time for
a given set Ω  . Our approach to obtain constant time lower-bounds for a given
set Ω  during search is to base its computation solely on the cardinality |Ω  |.
 The Unary Resource with Transition Times 93

More precisely, for each possible subset of cardinality k ∈ {1, . . . , n}, we pre-
compute the smallest transition time permutation of size k on Ω:
 
tt(k) =  min  min tt π
{Ω ⊆Ω: |Ω |=k} π∈ΠΩ 

For each k, the lower bound computation thus requires to solve a resource con-
strained shortest path problem (also NP-hard) with a fixed number of edges k
and with a free origin and destination. The next section proposes several ways
of pre-computing efficient lower bounds tt(k) for k ∈ {1, . . . , n}. Our formula
to compute a lower bound for the earliest completion time of a set of activities
(making use of pre-computed lower-bounds of transition times) becomes:

ectΩ = max

{est Ω  + pΩ  + tt(|Ω  |)} (3)
Ω ⊆Ω

4 Lower Bound of Transitions Times

The computation of tt(k) for all k ∈ {1, . . . , n} is NP-hard. This is a constrained
shortest path problem (for k = n it amounts to solving a TSP) in a graph
where each node corresponds to an activity and directed edges between nodes
represent the transition time between corresponding activities. Although these
computations are achieved at the initialization of the constraint, we propose to
use polynomial lower bounding procedures instead. Several approaches are used
and since none of them is dominated any other one, we simply take the maximum
of the computed lower bounds.

Minimum Weight Forest. A lower bound for tt(k) is a minimal subset of edges
of size k taken from this graph. We propose to strengthen this bound by using
Kruskal’s algorithm [12] to avoid selecting edges forming a cycle. We stop this
algorithm as soon as we have collected k edges. The result is a set of edges
forming a minimum weight forest (i.e. a set of trees) with exactly k edges.

Dynamic Programming. We can build the layered graph with exactly k lay-
ers and each layer containing all the activities. Arcs are only defined between
two successive layers with the weights corresponding to the transition times. A
shortest path on this graph between the first and last layer can be obtain with
Dijkstra. By construction this shortest path will use exactly k transitions, the
relaxation being that a same activity or transition can be used several times.

Minimum Cost Flow. Another relaxation is to keep the degree constraint but
relax the fact that selected edges must form a contiguous connected path. This
relaxation reduces to solving a minimum cost flow problem of exactly k units on
the complete bipartite graph formed by the transitions.
94 C. Dejemeppe et al.

Lagrangian Relaxation. As explained in Chapter 16 of [1], the Lagrangian relax-

ation (a single Lagrangian multiplier is necessary for the exactly k transitions
constraint) combined with a sub-gradient optimization technique can easily be
applied to compute a lower bound on the constrained shortest path problem.

5 Extending the Θ-tree with Transition Times

As introduced in [17], the O(n log n) propagation algorithms for unary resource
use the so-called Θ-tree data structure. We propose to extend it in order to
integrate transition times while keeping the same time complexities for all its
operations.
A Θ-tree is a balanced binary tree in which each leaf represents an activity
from a set Θ and internal nodes gather information about the set of (leaf) activ-
ities under this node. For an internal node v, we denote by Leaves(v), the leaf
activities under v. Leaves are ordered in non-decreasing order of the est of the
activities. That is, for two activities i and j, if est i < est j , then i is represented
by a leaf node that is at the left of the leaf node representing j. This ensures the
property:
∀i ∈ Left(v), ∀j ∈ Right(v) : est i ≤ est j
where left(v) and right(v) are respectively the left and right children of v, and
Left(v) and Right(v) denote Leaves(left(v)) and Leaves(right(v)).
A node v contains precomputed values about Leaves(v): ΣPv represents the
sum of the durations of activities in Leaves(v) and ect v is the ect of Leaves(v).
More formally, the values maintained in an internal node v are defined as follows:

ΣPv = pj
j∈Leaves(v)

ect v = ect Leaves(v) = max {est Θ + pΘ }

Θ  ⊆Leaves(v)

For a given leaf l representing an activity i, the values of ΣPl and ect l are pi
and ect i , respectively. In [18] Vilı́m has shown that for a node v these values
only depends on the values defined in both its left(v) and right(v) child. The
incremental update rules introduced in [18] are:

ΣPv = ΣPleft(v) + ΣPright(v)

 
ect v = max ect right(v) , ect left(v) + ΣPright(v)

An example of a classic Θ-tree is given in Figure 1.

When transition times are considered, the ect v value computed in the internal
nodes of the Θ-tree may only be a loose lower-bound since it is only based on
the earliest start times and the processing times. We strengthen the estimation
of the earliest computation times (denoted ect ∗ ) by also considering transition
times. We add another value inside the nodes: nv is the cardinality of Leaves(v)
(nv = |Leaves(v)|). The new update rules for the internal nodes of a Θ-tree are:
 The Unary Resource with Transition Times 95

Fig. 1. Classic Θ-tree as described in [18].

ΣPv = ΣPleft(v) + ΣPright(v)

nv = nleft(v) + nright(v)
 ∗ ∗
max{ectright(v ) , ectleft(v ) + ΣPright(v) + tt(nright(v) + 1)} : v internal
ectv∗ =
ect v : v leaf

As an example, let us consider the set of four activities used in the Θ-tree
example of Figure 1. Let us assume that the associated transition times are as
defined in the matrix M of Figure 2. The lower bounds for set of activities of
different cardinality are reported next to the matrix. With the new update rules
defined above, we obtain the extended Θ-tree presented in Figure 3. Note that
the values of ect ∗ in the internal nodes are larger than the values of ect reported
in the classic Θ-tree (Figure 1).

Fig. 2. Example of transition time matrix and associated lower bounds of transition
times permutations.

Lemma 1. ect v ≤ ectv∗ ≤ ectLeaves(v

 
) = maxΘ  ⊆Leaves(v) {est Θ +pΘ + tt(|Θ |)}
 

Proof. The proof is similar to the proof of Proposition 7 in [18], by also inte-
grating the inequality tt(|Θ |) ≥ tt(|Θ | ∩ Left(v)|) + tt(|Θ | ∩ Right(v)|), which
is itself a direct consequence of the fact that tt(k) is monotonic in k.

Since the new update rules are also executed in constant time for one node, we
keep the time complexities of the initial Θ-tree structure from [18] which are at
worst O(n log(n)) for the insertion of all activities inside the tree.
96 C. Dejemeppe et al.

Fig. 3. Extended Θ-tree for transition times. The ect ∗ values reported in the internal
nodes have been computed using the update rule of the extended Θ-tree.

Extending the Θ-Λ-tree with Transition Times

The Edge-Finding (EF) algorithm requires an extension of the original Θ-tree,
called Θ-Λ-tree [18]. This extension is used to obtain an efficient EF algorithm.
In this extension, in addition to the activities included in a Θ-tree, activities
can be marked as gray nodes. Gray nodes represent activities that are not really
in the set Θ. However, they allow to easily compute ect Θ if one of the gray
activities were included in Θ. If we consider the set of gray activities Λ such that
Λ ∩ Θ = ∅, we are interested in computing the largest ect obtained by including
one activity from Λ into Θ:
ect (Θ,Λ) = max ect Θ∪{i}
i∈Λ

In addition to ΣPv , ect v , the Θ-Λ-tree structure also maintains ΣPv and ect v ,
respectively corresponding to ΣPv and ect v , if the single gray activity in the
sub-tree rooted by v maximizing ect v were included:
  
∗ ∗ ∗
ect (Θ,Λ) = max ectΘ , max ectΘ∪{i}
i∈Λ

The update rule for ΣPv remains the same as the one described in [18]. However,
following a similar reasoning as the one used for the extended Θ-tree, we add
the nv value, and update rules are modified for ect v and n v . The rules become:
 
ΣPv = max ΣPleft(v) + ΣPright(v) , ΣPleft(v) + ΣPright(v)
⎧ ∗ ⎫
⎪ ect
⎨ right(v )
, ⎪
⎬
∗ ∗
ect v = max ect left(v ) + ΣP right(v) + tt(nright(v) + 1),
⎪
⎩ ∗ ⎪
⎭
ectleft(v ) + ΣPright(v) + tt(n right(v) + 1)

nv + 1 if the subtree rooted in v contains a gray node
nv =
nv otherwise
 The Unary Resource with Transition Times 97

This extended Θ-Λ-tree allows us to efficiently observe how the ect ∗ of a set of
activities is impacted if a single activity is added to this set. This information
allows the EF algorithm to perform propagation efficiently1 . An example of Θ-
Λ-tree based on the example from Figure 3 and Figure 2 is displayed in Figure 4.

Fig. 4. Extended Θ-Λ-tree with modified update rules.

Similarly to the reasoning applied for the Θ-tree, the time complexities
remain the same as the ones for the original Θ-Λ-tree structure from [18], which
are at worst O(n log(n)).

6 Disjunctive Propagation Algorithms with Transition

Times
In [18], a propagation procedure for the unary resource constraint is defined.
This propagation procedure consists of a propagation loop including Overload
Checking (OC), Detectable Precedences (DP), Not-First/Not-Last (NF/NL) and
Edge Finding (EF) propagation algorithms. The first three rely on the Θ-tree
while the latter employs the Θ-Λ-tree. Some small modifications can be done
to these algorithms to obtain an efficient propagation procedure making use of
knowledge about transition times.
1
Finding the “responsible” activity arg maxi ect Θ∪{i} (required by EF) is done simi-
larly to [18].
98 C. Dejemeppe et al.

6.1 Extension of Classic Unary Resource Propagation Algorithms

The four mentioned propagation algorithms use a Θ-tree or a Θ-Λ-tree to com-

pute ect Θ on a set of activities Θ. OC checks if ect Θ > lct Θ . DP, NF/NL and
EF rely on a set of rules that potentially allow to update the est or lct of an
activity. They all incrementally add/remove activities to a set of activities Θ
while maintaining the value ect Θ . When a rule is triggered by the consideration
of a given activity, the est or lct of this activity can be updated according to
the current value of ect Θ .
These four propagation algorithms can be used for the propagation of the
transition time constraints. To do so, we propose to substitute in the filtering
algorithms the Θ-tree and the Θ-Λ-tree structures by their extended versions.
∗
In the presence of transition times, ect ∗ /ect is indeed a stronger bound than
ect /ect . Furthermore, the update rules can be slightly modified to obtain an even
stronger propagation. When one of these algorithms detects that an activity i is
after all activities in a set Θ, the following update rule can be applied:
∗
est i ← max {est i , ectΘ }

In addition to all the transitions between activities of Θ - already taken into

∗
account in ectΘ - there must be a transition between one activity and i (not
necessarily from Θ, as we do not know which activity will be just before i in
the final schedule). It is therefore correct to additionally consider the minimal
transition from any activity to i. The update rule becomes:
 
∗
est i ← max est i , ectΘ + min tt j,i
j =i

An analogous reasoning can be applied to the update rule of the lct of an activity.
Similarly to the fix point propagation loop proposed in [18] for the unary
resource constraint, the four extended propagation algorithms are combined to
achieve an efficient propagation on transition time constraints. This allows to
obtain a global propagation procedure instead of the conjunction of pairwise
transition constraints described by Equation 1. The approach has however the
∗
disadvantage that the computation of ectΘ integrates a lower bound. This pre-
vents having the guarantee that sufficient propagation is achieved. The loop
must thus also integrate the conjunction of pairwise transition constraints given
in Equation 1. However, experimental results provided in Section 7 exhibits that
the supplementary global constraint reasoning can provide a substantial filtering
gain.

6.2 Detectable Precedences Propagation Example

Let us consider a small example (inspired from an example of [18]) with 3 activ-
ities, A, B and C whose domains are illustrated in Figure 5. The corresponding
transition matrix and lower bounds are given in Figure 6.
 The Unary Resource with Transition Times 99

Fig. 5. Example of extended Detectable Precedences with transition times. The

extended version updates est C from 11 to 17, while the joint use of transition time
binary constraints with the original unary constraint is not able to make this deduction.

Lower Bound k=1 k=2

⎛ ⎞ Min Weight Forest 2 5
046
Dynamic Programming 2 5
M = ⎝2 0 5⎠
Min Cost Flow 2 5
430
Lagrangian Relaxation 2 5
tt(k) 2 5

Fig. 6. Transition times for activities from Figure 5

Fig. 7. Comparison of classic and extended Θ-tree on the example described in Fig-
ures 5 and 6.

From this example, the Detectable Precedences algorithm will eventually

build a Θ-tree containing activities A and B. Figures 7a and 7b respectively
show the classic and the extended Θ-trees.
As one can see, ect ∗ is larger than ect as it is not agnostic about the transition
time constraints. Furthermore, the update rule of est C also includes the minimal
transition time from any activity to C. This leads to the following update of est C :
 
∗
est C = max est C , ectΘ + min tt i,C
i =C
= max {11, 12 + 5} = 17
100 C. Dejemeppe et al.

We finally obtain an updated est C , as shown by the red bold bracket in Figure 5.
Notice that the joint use of the constraints given in Equation 1 with the original
unary constraint of [18] would not make this deduction.

7 Evaluation
To evaluate our constraint, we used the OscaR solver [14] and ran instances on
AMD Opteron processors (2.7 GHz). For each considered instance, we used the
3 following filterings for the unary constraint with transition times:

1. Binary constraints2 (φb ) given in Equation 1.

2. Binary constraints given in Equation 1 with the Unary global constraint of
[18] (φb+u ).
3. The constraint introduced in this article (φuTT ). Based on our experience,
we slightly changed the propagation loop order: Edge-finding is put in first
position.

Considered Benchmarks. We constructed instances considering transition times

from famous JobShop benchmarks. For a given benchmark B, in each instance,
we added generated transition times between activities, while ensuring that tri-
angular inequality always hold. From B, we generated new benchmarks B(a,b)
inside which the instances are expanded by transition times uniformly picked
between a% and b% of D, where D is the average duration of all activities in
the original instance.
We generated instances from the well-known Taillard’s instances3 . From each
instance, we generated 2 instances for a given pair (a, b), where the following pairs
were used: (50, 100), (50, 150), (50, 200), (100, 150), (100, 200) and (150, 200).
This allowed us to create 960 new instances4 .

Comparison of the 3 models

In order to present fair results regarding the benefits that are only provided by
our constraint, we first followed the methodology introduced in [16]. Afterwards,
we made measurements using a static search strategy, as it cannot be influenced
by the additional pruning provided by our constraint.

Potential of the Constraint. In brief, the approach presented in [16] proposes to

pre-compute a search tree using the filtering that prunes the less - the baseline

2
For efficiency reason, dedicated propagators have been implemented instead of post-
ing reified constraint.
3
Available at http://mistic.heig-vd.ch/taillard/problemes.dir/ordonnancement.dir/
ordonnancement.html.
4
Available at http://becool.info.ucl.ac.be/resources/benchmarks-unary-resource-
transition-times
 The Unary Resource with Transition Times 101

propagator - and then to replay this search tree using the different studied fil-
tering procedures. The point is to only measure the time gain provided by the
propagation, by decoupling the gain provided by the search strategy (while still
being able to use dynamic ones) from the one provided by the propagation. We
used φb as the baseline filtering, and the SetTimes (st) search strategy to con-
struct the search tree, as this strategy is recognized to provide good performances
in Scheduling. The search tree construction time was limited to 600 seconds. We
then constructed performance profiles as described in [16]. Basically, those are
cumulative distribution functions of a performance metric τ . Here, τ is the ratio
between the solution time (or number of backtracks) of a target approach (i.e.
φb+u or φuTT ) and that of the baseline (i.e. φb ). For time (similar for number of
backtracks), the function is defined as:
 
1  t(replay(st), M ∪ φi ) 

Fφi (τ ) =  M ∈ M : ≤ τ  (4)
|M| t(replay(st), M )

where M is the set of considered instances while t(replay(st), M ∪ φi ) and

t(replay(st), M ) are respectively the time required to replay the generated search
tree with the studied model (model using φi , i.e. φb+u or φuTT ) and with the
baseline model.
Figures 8a and 8b respectively provide the profiles for time and backtrack
for all the 960 instances5 . Figure 8c provides a “long-term” view of Figure 8a.
From Figure 8a, we can first conclude that φb+u is clearly worse than φuTT
and φb from a time perspective. Moreover, Figure 8b shows that φb+u rarely
offers more pruning than φb .
In comparison, we can see from Figure 8a that for ∼ 20% of the instances,
φuTT is about 10 times faster than φb , and that we solve ∼ 35% of the instances
faster (see FφuTT (1)). Moreover, it offers more pruning for ∼ 75% of the instances
(see Figure 8b).
From Figure 8c, we can see that the constraint does not have too much
overhead, as φuTT is at worst about 7.5 times slower than φb for ∼ 45% percent
of the instances (FφuTT (7.5) − FφuTT (1)). It is a bit slower for the remaining

1.00
φuTT
0.75 φb+u
% instances

0.50

0.25

0.00
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0 0.0 2.5 5.0 7.5 10.0 12.5
τ τ τ

(a) τ is a time ratio. (b) τ is a backtrack ratio. (c) “Long-term” profile (τ

is a time ratio).

Fig. 8. Performance profiles for the 960 generated instances.

5
When instances were separated by number of jobs, the profiles had similar shapes.
102 C. Dejemeppe et al.

Table 1. Best time results for φuTT compared to φb . The problem is to find the given
makespan m using a binary static search strategy. Time is in seconds.
φuTT φb φb+u
Instance m
Time #Fails Time #Fails Time #Fails
15 15-3 225 50 100-1 2,344 1.12 2,442 117.92 980,330 432.07 911,894
50 15-8 750 50 100-2 6,682 2.11 744 182.27 1,127,272 999.79 1,127,272
20 15-7 300 150 200-2 4,784 0.24 449 17.63 168,466 62.27 168,466
15 15-6 225 50 100-1 2,398 3.90 5,593 187.93 889,079 534.20 602,591
50 20-3 1000 50 150-2 7,387 2.96 1,709 126.61 584,407 829.25 584,407
100 20-4 2000 150 200-1 18,595 11.59 885 340.32 332,412 1225.44 206,470
30 15-3 450 50 200-1 4,643 1.97 1,178 39.23 226,700 314.34 226,700
15 15-5 225 100 150-2 3,320 0.91 2,048 16.40 119,657 63.38 119,657
50 20-2 1000 50 100-1 6,979 3.79 1,680 63.16 878,162 4.63 1,695
30 15-10 450 100 200-1 5,586 0.74 687 9.24 106,683 41.25 106,683

Table 2. Worst time results for φuTT compared to φb . The problem is to find the given
makespan m using a binary static search strategy. Time is in seconds.
φuTT φb φb+u
Instance m
Time #Fails Time #Fails Time #Fails
15 15-10 225 50 200-2 2,804 645.26 546,803 127.38 546,803 572.81 546,803
50 15-9 750 50 200-1 6,699 954.77 164,404 174.63 164,437 1,108.43 164,437
20 20-5 400 100 150-2 4,542 213.54 78,782 38.26 78,968 180.20 78,968
20 20-8 400 100 150-2 4,598 147.55 164,546 26.42 164,576 175.69 164,576
15 15-2 225 50 100-2 2,195 178.37 96,821 31.23 96,821 139.84 96,821
20 20-6 400 100 200-1 4,962 11.15 8,708 1.94 8,745 11.87 8,745
30 20-8 600 50 200-1 5,312 18.63 6,665 3.15 6,687 19.93 6,687
20 15-10 300 50 200-2 3,571 85.84 61,185 14.24 61,185 65.12 61,185
50 20-8 1000 100 200-1 9,186 286.61 88,340 46.17 88,340 180.23 88,340
20 15-1 300 100 150-1 3,557 189.37 208,003 29.55 209,885 157.33 209,885

∼ 20%, which roughly corresponds to the percentage of instances for which

φuTT provides no extra pruning (see FφuTT (1) in Figure 8b).

Evaluation Over a Static Search Strategy. We here present results in a more

“traditional” fashion. We compute the best makespan m that can be obtained
with φb within 600 seconds, using the following binary static search strategy:
fixed variable order, left branch assigns starti to est i , right branch removes est i
from the domain of starti . Then, the time and number of failures required by
each model to find this solution are computed. We filtered out instances for
which the solution was found by φb in less than 1 seconds and we computed the
time ratio between φuTT and φb . From this perspective, the 10 best and worst
results are reported in tables 1 and 2, respectively. On the 10 best instances, the
gains (the number of failures and time) are significant (sometimes two orders
of magnitude). On the 10 worst instances, the times obtained with φuTT are
 The Unary Resource with Transition Times 103

similar to the results using the classical unary resource (i.e. φb+u ), while they are
at worst around 6.4 times slower than the simple binary decomposition (i.e. φb ).

8 Conclusion

In this paper, we proposed to extend classic unary resource propagation algo-

rithms such that they consider transition times. We first stated that a lower
bound of the time taken by transitions between activities from a set Ω is required
to have a tighter bound of ect Ω . We described several possible methods to com-
pute these lower bounds. We then proposed to extend the Θ-tree and Θ-Λ-tree
structures to integrate these lower bounds. These extended structures can then
be used in unary propagation algorithms: OC, DP, NF/NL and EF. The new
obtained propagation procedure has the advantage that it can be used conjointly
with any other constraint and that it is completely independent from the objec-
tive to optimize. We have demonstrated that the additional pruning achieved
by this propagation can dramatically reduce the number of nodes (and thus the
time taken to solve the problem) on a wide range of instances.
Future work would analyze the possibility to integrate tighter incremental
lower bounds in Θ-tree and Θ-Λ-tree structures. The order and real usefulness
of the propagators (OC, DP, NF/NL, EF) should also be studied in order to
acquire the most efficient fixpoint propagation loop. Finally, we would like to
experiment on a new update rule in Θ-tree and Θ-Λ-tree to be able to obtain
tighter lower bounds for ect Ω .

References
1. Ahuja, R.K., Magnanti, T.L., Orlin, J.B.: Network Flows: Theory, Algorithms, and
Applications. Prentice-Hall Inc, Upper Saddle River (1993)
2. Allahverdi, A., Ng, C., Cheng, T.E., Kovalyov, M.Y.: A survey of scheduling prob-
lems with setup times or costs. European Journal of Operational Research 187(3),
985–1032 (2008)
3. Artigues, C., Belmokhtar, S., Feillet, D.: A new exact solution algorithm for the job
shop problem with sequence-dependent setup times. In: Régin, J.-C., Rueher, M.
(eds.) CPAIOR 2004. LNCS, vol. 3011, pp. 37–49. Springer, Heidelberg (2004)
4. Artigues, C., Buscaylet, F., Feillet, D.: Lower and upper bound for the job shop
scheduling problem with sequence-dependent setup times. In: Proceedings of the
Second Multidisciplinary International Conference on Scheduling: Theory and
Applications, MISTA 2005 (2005)
5. Artigues, C., Feillet, D.: A branch and bound method for the job-shop problem with
sequence-dependent setup times. Annals of Operations Research 159(1), 135–159
(2008)
6. Balas, E., Simonetti, N., Vazacopoulos, A.: Job shop scheduling with setup times,
deadlines and precedence constraints. Journal of Scheduling 11(4), 253–262 (2008)
7. Baptiste, P., Le Pape, C., Nuijten, W.: Constraint-based scheduling: applying con-
straint programming to scheduling problems, vol. 39. Springer Science & Business
Media (2001)
104 C. Dejemeppe et al.

8. Focacci, F., Laborie, P., Nuijten, W.: Solving scheduling problems with setup times
and alternative resources. In: AIPS, pp. 92–101 (2000)
9. Gagné, C., Price, W.L., Gravel, M.: Scheduling a single machine with sequence
dependent setup time using ant colony optimization. Faculté des sciences de
l’administration de l’Université Laval, Direction de la recherche (2001)
10. González, M.A., Vela, C.R., Varela, R.: A new hybrid genetic algorithm for the job
shop scheduling problem with setup times. In: ICAPS, pp. 116–123 (2008)
11. Grimes, D., Hebrard, E.: Job shop scheduling with setup times and maximal time-
lags: a simple constraint programming approach. In: Lodi, A., Milano, M., Toth, P.
(eds.) CPAIOR 2010. LNCS, vol. 6140, pp. 147–161. Springer, Heidelberg (2010)
12. Kruskal, J.B.: On the shortest spanning subtree of a graph and the traveling
salesman problem. Proceedings of the American Mathematical society 7(1), 48–50
(1956)
13. Oddi, A., Rasconi, R., Cesta, A., Smith, S.F.: Exploiting iterative flattening search
to solve job shop scheduling problems with setup times. PlanSIG2010, p. 133 (2010)
14. OscaR Team: OscaR: Scala in OR (2012). https://bitbucket.org/oscarlib/oscar
15. Tahar, D.N., Yalaoui, F., Amodeo, L., Chu, C.: An ant colony system minimizing
total tardiness for hybrid job shop scheduling problem with sequence dependent
setup times and release dates. In: Proceedings of the International Conference on
Industrial Engineering and Systems Management, pp. 469–478 (2005)
16. Van Cauwelaert, S., Lombardi, M., Schaus, P.: Understanding the potential of
propagators. In: Proceedings of the Twelfth International Conference on Integra-
tion of Artificial Intelligence and Operations Research techniques in Constraint
Programming (2015)
17. Vilı́m, P.: O (nlog n) filtering algorithms for unary resource constraint. In:
Régin, J.-C., Rueher, M. (eds.) CPAIOR 2004. LNCS, vol. 3011, pp. 335–347.
Springer, Heidelberg (2004)
18. Vilım, P.: Global constraints in scheduling. Ph.D. thesis, Charles University in
Prague, Faculty of Mathematics and Physics, Department of Theoretical Computer
Science and Mathematical Logic, KTIML MFF, Universita Karlova, Praha (2007)
19. Vilı́m, P., Barták, R., Čepek, O.: Extension of o (n log n) filtering algorithms for
the unary resource constraint to optional activities. Constraints 10(4), 403–425
(2005)
20. Wolf, A.: Constraint-based task scheduling with sequence dependent setup times,
time windows and breaks. GI Jahrestagung 154, 3205–3219 (2009)
21. Zampelli, S., Vergados, Y., Van Schaeren, R., Dullaert, W., Raa, B.: The berth
allocation and quay crane assignment problem using a CP approach. In: Schulte, C.
(ed.) CP 2013. LNCS, vol. 8124, pp. 880–896. Springer, Heidelberg (2013)
 A Global Constraint for a Tractable Class
of Temporal Optimization Problems

Alban Derrien1 , Jean-Guillaume Fages2 , Thierry Petit1,3 ,

and Charles Prud’homme1(B)
1
TASC (CNRS/INRIA), Mines Nantes, 44307 Nantes, France
{alban.derrien,charles.prudhomme}@mines-nantes.fr
2
COSLING S.A.S., 44307 Nantes, France
jg.fages@cosling.com
3
Foisie School of Business, WPI, Worcester, MA 01609, USA
TPetit@wpi.edu

Abstract. This paper is originally motivated by an application where

the objective is to generate a video summary, built using intervals
extracted from a video source. In this application, the constraints used
to select the relevant pieces of intervals are based on Allen’s algebra.
The best state-of-the-art results are obtained with a small set of ad hoc
solution techniques, each specific to one combination of the 13 Allen’s
relations. Such techniques require some expertise in Constraint Program-
ming. This is a critical issue for video specialists. In this paper, we design
a generic constraint, dedicated to a class of temporal problems that cov-
ers this case study, among others. ExistAllen takes as arguments a vector
of tasks, a set of disjoint intervals and any of the 213 combinations of
Allen’s relations. ExistAllen holds if and only if the tasks are ordered
according to their indexes and for any task at least one relation is satis-
fied, between the task and at least one interval. We design a propagator
that achieves bound-consistency in O(n + m), where n is the number of
tasks and m the number of intervals. This propagator is suited to any
combination of Allen’s relations, without any specific tuning. Therefore,
using our framework does not require a strong expertise in Constraint
Programming. The experiments, performed on real data, confirm the
relevance of our approach.

1 Introduction
The study of temporal relations between elements of a process is a very active
topic of research, with a wide range of applications: biomedical informatics [5,20],
law [17], media [2,4,8] etc. Temporal reasoning enables to analyze the content of
a document in order to infer high-level information. In particular, a summary of
a tennis match may be generated from the match recording with Artificial Intel-
ligence techniques [2,4]. The summarization requires to extract some noticeable
time intervals and annotate them with qualitative attributes using signal recog-
nition techniques. Then, the summary generation may be formulated as a Con-
straint Satisfaction Problem (CSP) where variables are the video segments to

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 105–120, 2015.
DOI: 10.1007/978-3-319-23219-5 8
106 A. Derrien et al.

be displayed whereas constraints stem from different considerations: displaying

relevant information, balancing the selected content, having nice transitions etc.
Then, the CSP may be solved quite efficiently with a Constraint-Programming
(CP) solver. At first sight, this seems an easy task.
Unfortunately, things get harder when it comes to practice. Designing a CP
model requires some expert knowledge, in order to achieve good performances
on hard problems. This is particularly true when one has to design a global
constraint that would be missing in the solver. For instance, the summariza-
tion model of [2] relied on many disjunctions that are poorly propagated by
constraint engines and thus lead to unsatisfiable performances. Therefore, the
authors collaborated with CP experts to design ad hoc global constraints, which
lead to significant speedups [4]. Alternatively, one may have used temporal logic
models, to benefit from solution techniques and solvers dedicated the tempo-
ral CSP [6,7,9,18]. As the video summarization gets more sophisticated, all
these approaches suffer from the need of specific and often intricate propaga-
tors/models. This is a critical issue for video specialists, who are rarely CP
experts. Furthermore, one may need to include in her model other features avail-
able in state-of-the-art constraint solvers, such as standard global constraints and
predefined search strategies. What is missing is a both expressive and efficient
global constraint for modeling a relevant class of problems on time intervals.
We introduce the ExistAllen constraint, defined on a vector of tasks T and
a set of disjoint intervals I, respectively of size n and m. Given a subset R
of Allen’s relations [1], ExistAllen is satisfied if and only if the two following
properties are satisfied:

1. For any task in T at least one relation in R is satisfied between this task
and at least one interval in I.
2. Tasks in T are ordered according to their indexes in the vector given as
argument, i.e., for any integer i, 1 ≤ i < n, the task Ti should end before or
at the starting time of Ti+1 .

In the context of video-summarization, tasks in T are the video segments

that compose the summary. Fixed video sequences in I are extracted from the
source according to some precise features. In this way, it is possible to constrain
the content of the summary with qualitative information.
Considering the invariability of task processing times, we introduce a bound-
consistency propagator for this constraint, suited to any of the 213 subsets of
Allen’s relations. The time complexity of the most natural algorithm for this
propagator is O(n×m). We propose an improved algorithm, running in O(n+m)
time. While ExistAllen may be used in different contexts, e.g., online schedul-
ing, this paper is motivated by video-summarization. Our experiments on the
Boukadida et al.’s application [4] demonstrate that using our generic constraint
and its linear propagator is significantly better than the models built with stan-
dard constraints of the solver, and competitive with the ad hoc global constraint
approach.
 A Global Constraint for a Tractable Class 107

Table 1. Allen’s temporal algebra relations.

Symbol Relation
T (1)
T (2)
T (1) p T (2) (T (2) pi T (1) ) T (1) precedes T (2)
T (1)
T (2)
T (1) m T (2) (T (2) mi T (1) ) T (1) meets T (2)
T (1)
T (2)
T (1) o T (2) (T (2) oi T (1) ) T (1) overlaps T (2)
T (1)
T (2)
T (1) s T (2) (T (2) si T (1) ) T (1) starts T (2)
T (1)
T (2)
T (1) d T (2) (T (2) di T (1) ) T (1) during T (2)
T (1)
(2)
T
T (1) f T (2) (T (2) f i T (1) ) T (1) finishes T (2)
T (1)
T (2)
T (1) eq T (2) T (1) equal to T (2)

2 Background

In this section we give some background and fix the notations used in this paper.

2.1 Temporal Constraint Networks

Temporal reasoning has been an important research topic for the last thirty
years. One may distinguish qualitative temporal reasoning, based on relations
between intervals and/or time points, from quantitative reasoning where dura-
tion of a given event is represented in a numerical fashion. Allen’s algebra [1]
represents qualitative temporal knowledge by interval constraint networks. An
interval constraint network is a directed graph where nodes represent intervals
and edges are labelled with disjunctions of Allen’s relations. Table 1 details those
relations. Many state-of-the-art papers deal with generic solving techniques and
tractability of temporal networks [6,7,9,18], including temporal problems with
quantified formulas [10,12]. Most of these methods make no strong restriction
on the constraint network to be solved.
A few techniques, more specialized, focus on optimization problems. The two
most related to this paper are the following. Kumar et al. [11] consider temporal
problems with “taboo” regions, minimizing the number of intersections between
108 A. Derrien et al.

tasks and a set of fixed intervals. These problems occur in scheduling applica-
tions, among others. In the context of video summarization, a recent paper [4]
proposes the idea of using global constraints involving a set of ordered intervals.
Each constraint is restricted to one specific relation in the 213 combinations of
Allen’s relations. The propagators are not described.

2.2 Constraint Programming (CP)

CP is a problem solving framework where relations between variables are stated
in the form of constraints, which together form a constraint network. Each vari-
able x has a domain D(x), whose minimum value is x and maximum value is x.
A task T in a set T is an object represented by three integer variables: sT , its
starting time, eT , its ending time, and pT , its processing time. The task should
satisfy the constraint sT +pT = eT . An interval I in a set I is a fixed task, defined
by integer values instead of variables. A propagator is an algorithm associated
with a constraint, stated on a set of variables. This propagator removes from
domains values that cannot be part of a solution to that constraint. The notion
of consistency characterizes propagator effectiveness. In this paper, we consider
bound (Z)-consistency [3]. When domains are exclusively represented by their
bounds (i.e., have no holes), bound (Z)-consistency ensures that for each variable
x, x and x can be part of a solution of the constraint.

3 The ExistAllen Constraint

This section introduces the ExistAllen constraint and its propagator. Let T =
{T1 , T2 , . . . , Tn } be a set of tasks, such that any task Ti+1 must be scheduled
at or after the end of task Ti . Similarly, we define a set of ordered Intervals
I = {I1 , I2 , . . . , Im }. From a subset R of Allen’s relations, ExistAllen ensures
that any task in T is related to at least one interval in I.
Definition 1 (ExistAllen). ExistAllen(T , R, I) ⇔
 
∀T ∈ T , T RI
R∈R I∈I

∧ (∀i, 1 ≤ i < n, sTi+1 ≥ eTi )

Example 1. When summarizing the video of a tennis match, it may be required
that each segment (task) selected to be part the summary contains an applause.
Applause intervals can be preprocessed [4]. Such requirement can then be for-
mulated with a ExistAllen constraint, where T are the selected segments, I are
the applause segments and R is set to {f i, di, eq, si}.
A symmetry can be used, where the problem is seen in a mirror: starting
variables become ending variables and the lower bounds filtering of the mirror
relation is applied to the upper bounds (e.g. starts is propagated onto upper
bounds using finishes, see Table 1). Therefore, we put the focus on the algorithms
for the lower bounds of starting/ending task variables in T . We consider here
that processing times are exclusively updated by the constraints sT + pT = eT .
 A Global Constraint for a Tractable Class 109

3.1 Basic Filtering: One Relation, One Task, One Interval

The basic filtering rule specific to each Allen’s relation can be derived from time-
point logic relations [19], in order to state lower and upper bounds of starting and
ending times of tasks. For instance, consider a task T ∈ T , an interval I ∈ I, and
the relation starts. The relation is satisfied if and only if two conditions are met:

T s I ⇔ sT = sI ∧ eT < eI

The only filtering on the lower bounds of task T induced by the two conditions of
relation starts is sT ≥ sI . On the same basis, we define in Table 2 the conditions
and filtering rules for the 13 Allen’s algebra relations between a task T and an
interval I.

Table 2. Allen’s algebra lower bound filtering on each variable of a task T .

Relation Conditions Filtering Relation Conditions Filtering

T pI eT < sI sT = sI sT ≥ sI
T si I
T pi I eI < sT sT > eI eT > eI eT > eI
T mI eT = sI eT ≥ sI sT > sI sT > sI
T dI
T mi I sT = eI sT ≥ eI eT < eI
sT < sI sT < sI
T di I
T oI eT < eI eT > eI eT > eI
eT > sI eT > sI sT > sI sT > sI
T f I
sT < eI eT = eI eT ≥ eI
T oi I sT > sI sT > sI sT < sI
T fi I
eT > eI eT > eI eT = eI eT ≥ eI
sT = sI sT ≥ sI sT = sI sT ≥ sI
T sI T eq I
eT < eI eT = eI eT ≥ eI

3.2 A First Propagator

We propose a propagator based on two procedures. Again, we only present the

algorithms adjusting lower bounds of starting times and ending times.
The main procedure, ExistAllenQuadratic (Algorithm 1), takes as argu-
ments the sets T and I and a subset R of Allen’s relations. Algorithm 1 considers
all tasks in T and checks the relations according to intervals in I. At the end of
the procedure, the bounds of variables in T are updated according to the earliest
support. If no support has been found, a domain is emptied and the solver raises
a failure exception. The order between the tasks is maintained by the procedure
propagate at line 10, whose propagation is obvious.
Algorithm 1 calls the procedure AllenAllRelation(Algorithm 2, line 6),
which performs the check for one task and one interval, with all the relations
in R. In order to define a procedure instead of a function returning a pair of
bounds, we use two global variables s∗ and e∗ , storing permanently the two
lowest adjustments, respectively for the lower bounds of the starting and ending
110 A. Derrien et al.

Require: Global Variable : s∗ , e∗

1: procedure ExistAllenQuadratic(T , R, I)
2: for i = 1 to n do  Loop over tasks
3: s∗ ← sTi + 1  Intialize to a value out of the domain
4: e∗ ← eTi + 1
5: for j = 1 to m do  Loop over Intervals
6: AllenAllRelation(Ti , R, Ij )
7: end for
8: sTi ← s∗
9: eTi ← e∗
10: if i < n − 1 then propagate(eTi , ≤, sTi+1 ) end if
11: end for
12: end procedure
Algorithm 1: Main procedure.

time of the current task. An adjustment of the bound of one such variable is
made in Algorithm 2 if and only if the current relation gives a support which is
less than the lowest previously computed support.

Require: Global Variable : s∗ , e∗

1: procedure AllenAllRelation(T, R, I)
2: for all r ∈ R do
3: if checkCondition(T, r, I) then
4: s∗ ← min(s∗ , seekSupportStart(T, r, I))
5: e∗ ← min(e∗ , seekSupportEnd(T, r, I))
6: end if
7: end for
8: end procedure
Algorithm 2: Update of s∗ and e∗ .

The function checkCondition(T, r, I) (line 3) returns true if and only if a

support can be found. Consider again the example of relation starts. From the
condition induced by Table 2, col. 2, we have:

checkCondition(T , s, I) ⇔ sT ≤ sI ∧ sT ≥ sI ∧ eT < eI .

If the conditions are met then a support exists. seekSupportStart(T , s, I)

returns the lowest support for the starting time variable, that is, sI . As no fil-
tering is directly induced for the ending time variable, the minimal support
returned, for the relation s by seekSupportEnd(T , s, I) is max(eT , sI +pT ).
For each Allen’s algebra relation, the three functions are similarly derived from
Table 2.

Lemma 1. The time complexity of Algorithm 1 is in O(n × m).

 A Global Constraint for a Tractable Class 111

Proof. As the number of Allen’s relations is constant (at most 13), and the call
of the three functions (lines 3, 4 and 5 in Algorithm 2) are in O(1), the whole
filtering of lower bounds is performed in O(n × m) time complexity.
Theorem 1. The propagator based on Algorithm 1 and its symmetric calls for
upper bounds of starting/ending variables in T , ensure bounds(Z)-consistency
if processing times are fixed integers.
Proof. After the execution of Algorithm 1, the optimal update has been done for
each lower bound, according to the current upper bounds. The filtering of upper
bounds is symmetric. Therefore, providing that durations of tasks are fixed, the
propagator ensures bound (Z)-consistency when a fixpoint is reached.
A fixpoint is not necessarily reached after running Algorithm 1 twice, once
to filter lower bounds and once to filter upper bounds. Indeed, the pass on
upper bounds can filter values which were previously supports for lower bounds.
Let’s consider ExistAllen(T , R, I) depicted in Figure 1 wherein: T = {T1 =
sT1 = [0, 2], pT1 = [2, 4], eT1 = [4,6] , T2 = sT2 = [5, 6], pT2 = [4, 5], eT2 = [10, 10] },
R = {d, di} and I = {I 1 = [1, 5], I 2 = [6, 8]}.

T1 = sT1 = [0, 2], eT1 = [4,6] T1 for T1 on di = [0, 6[

T1 for T1 on d = [2, 4[

T2 = sT2 = [5, 6], eT2 = [10, 10] T2

I 1 = [1, 5], I 2 = [6, 8] I1 I2

0 1 2 3 4 5 6 7 8 9 10

Fig. 1. Several phases may be required to get a fixpoint when the processing times of
tasks are not fixed.

We now simulate the calls of Algorithm 1 required to reach a fixpoint. No

values are filtered during the run of Algorithm 1 on lower bounds, since di
provides the minimal support for sT1 , d provides the minimal support for eT1
and di provides the minimal supports for T2 . On the run of Algorithm 1 on
upper bounds, since no relation provides support for 6 from sT2 , the value is
removed. The ordering constraint (Algorithm 1, line 10) removes 6 from eT1 .
Thus, di is no longer valid to provide a support for T1 . Consequently, a second
call to Algorithm 1 on lower bounds has to be done and the minimal value for
sT1 will then be 2, given by relation d.
However, when task processing times are constants, the minimum support of
the ending date of an activity is the minimum support of its starting date plus the
constant. Therefore, the issue mentioned in the previous example cannot occur.
The fixpoint can be reached in two passes. One may note that, in this case, we
could improve the algorithm by ordering Allen’s relations. As our target appli-
cation involves variable processing times, we do not detail this simplification.
112 A. Derrien et al.

3.3 A Linear Propagator

This section introduces an improved propagator, running in O(n + m) time
complexity.
First, one may observe that the satisfaction of the relation precedes can be
done in constant time. Indeed, if a task T can precede the last interval Im ,
the lower bounds are a support. And if not, task T cannot precede any inter-
val. The same way relation precedes inverse can be symmetrically checked with
the first intervals. Therefore, to simplify the presentation and without loss of
generality, we now consider that relations p and pi can be isolated and treated
separately. For each task, they can be checked in constant time. In this section
we exclude them from R.
Second, as the sets T and I are chronologically ordered, we can exploit
dominance properties that lead to a linear propagator. We now provide those
properties and their proof, as well as the ExistAllenLinear procedure. As in
Section 3.2, we focus on lower bounds of starting/ending variables of tasks in T .

Property 1. An interval I which starts after a support y for the ending time of
a task T cannot provide a support for eT lower than y.

Proof. Consider a task T , R ⊆ {m, mi, o, oi, s, si, d, di, f, f i, eq}, y a support for
eT and I an interval, such that y < sI . From filtering rules in Table 2, col. 3, no
relation in R provides a support lower than y for eT . 


Property 2. For any of the relations in {mi, oi, s, si, d, f, eq}, an interval I which
starts after a support y for the ending time of a task T cannot provide a support
for sT lower than y.

Proof. Similar to proof of Property 1. 



Property 3. For any of the relations in {m, o, di, f i}, an interval I which starts
after a support for the ending time of a task T can provide a support for sT . The
lowest support for T is then given by the interval I in I having the lowest sI .

Proof. Let T be a task and r be a relation, r ∈ R = {m, o, di, f i}, x be a support

for sT , y be a support for eT and I be the interval in I with the lowest sI , such
that sI > y. For each r ∈ R, we distinguish two cases.
Case 1 (T is related to I with relation r, T r I). As the support for eT is at
least sI and no rule on starting time is explicitly defined by r (Table 2, col. 3),
then the support for sT is at least sI − pT . Given that all intervals from I are
chronologically ordered, no interval greater than I can provide a lower x.

Case 2 (T is not related to I with relation r: ¬(T r I)). Consider the m relation.
Task T can meet interval I only if sI ∈ [eT , eT ] (Table 2 col. 2). As it exists a
support y with a value e∗ < sI and the relation m is not satisfied, we have
eT < sI . Given that all intervals are chronologically ordered, no interval with a
greater index can meet the task T . A similar reasoning can be applied to the
other relations in R. 

 A Global Constraint for a Tractable Class 113

Thanks to Properties 1, 2 and 3, we can improve the Algorithm 1 by stopping

the iteration over intervals in I for a given task T (Line 5) if e∗ < sI . We now
provide two properties from which, for a given task T, the iteration over intervals
in I does not have to always start at the first interval of I.

Property 4. An interval I which ends before the ending time of a task Ti cannot
provide a support for the next task Ti+1 .

Proof. Let Ti be a task and I an interval such that eI < eTi . Then eI < sTi+1 .
Ti+1 cannot be in relation with I. 


Property 5. Given a task T , there exists at most one interval between the interval
I i with the highest ending time such that eI i < eT and the interval I j with the
lowest starting time such that sI j > eT .

Proof. Let T be a task, let I i be the interval with the highest eI such that
eI < eT , we have then eI i+1 > eT , and let I j be the interval with the lowest
sI j such that sI j > eT , we have then sI j−1 < eT . Given that all intervals are
chronologically ordered, and that sI j−1 < eI i+1 , we have that j − 1 ≤ i + 1, that
is j − i ≤ 2. As the difference between indices i and j is at most 2, there is at
most one interval between i and j. 


Thanks to Properties 4 and 5, we know that the next task cannot be in

relation with any interval whose index is lower than or equal to j − 2. We can
improve the Algorithm 1 by starting the iteration over intervals for a given task
at j − 1.

Require: Global Variable : s∗ , e∗

1: procedure ExistAllenLinear(T , R, I)
2: j←0
3: for i = 1 to n do  Loop over Tasks
4: s∗ ← sTi + 1  Intialize to a value out of the domain
5: e∗ ← eTi + 1
6: repeat  Loop over Intervals
7: j ←j+1
8: AllenAllRelation(Ti , R, Ij )
9: until j < m and e∗ < sIj  Stop iteration, see Properties 1, 2 and 3
10: sTi ← s∗
11: eTi ← e∗
12: if i < n − 1 then propagate(eTi , ≤, sTi+1 ) end if
13: j ← max(0, j − 2)  Set next interval index, see Properties 4 and 5
14: end for
15: end procedure
Algorithm 3: Linear Algorithm for Main Procedure.

By construction, Algorithms 3 and 1 do exactly the same filtering.

Theorem 2. The time complexity of Algorithm 3 is in O(n + m).

114 A. Derrien et al.

Proof. The number of evaluated intervals is equal to m+2×(n−1): every time a

new task is evaluated, the algorithm goes two intervals back. The new algorithm
is then in O(n + m). 


3.4 Improvements

In this section, we describe three improvements brought to the ExistAllen con-

straint in order to improve its efficiency in practice.
First, the Algorithm 2 can be adapted to store, for a given task, the index
of the first interval which satisfied the conditions of a relation. Indeed, intervals
located before that interval do not provide a support for the task (and they will
never do in the current search sub-tree). By doing so, useless calls to check-
Condition can be avoided since they will always return false. In practice, an
operation is added after the line 7 in Algorithm 3 to put in j the maximum
between j and the first satisfying interval for the task evaluated. These indices
are automatically updated upon backtrack.
Similarly, the tasks whose variables have been modified since the last call
of the procedure have to be memorized. Thus, the for-loop (line 3-14 in Algo-
rithm 3) can start from the index of the first modified task. Moreover, following
tasks that have not been modified can be skipped safely.
Finally, our generic framework enables to replace some remarkable combina-
tions of the Allen’s algebra relations with meta-relations. By doing so, even if
the complexity of the Algorithm 3 remains the same, the number of operations
made to check conditions and seek supports for the combined relations may be
reduced. For instance, a “contains” meta-relation, as described in Example 1,
which expresses {f i, di, eq, si}, can save up to three calls of the methods in the
for-loop in Algorithm 2, lines 2-7. Note that since {p, pi} are handled in a par-
ticular way by the constraint, it is even more efficient to limit the combinations
to relations in {m, mi, o, oi, s, si, d, di, f, f i, eq}. Adding meta-relations is easy in
our implementation since we use a facade design pattern to define the 13 rela-
tions. Some meta-relations may require to define their inverse, in order to filter
on upper bounds. This is not the case for “contains”, though. Indeed, the mirror
relation of f i is si, the mirror relation of si is f i, while the mirror relation of di
is di and the mirror relation of eq is eq.

4 Evaluation

The main contribution of this work is an “expressivity gain”, which leads to

reducing the investment necessary to build and maintain a model. Nevertheless,
it is important to check if this gain does not come at the price of efficiency.
In this section, we empirically evaluate the impact of the proposed filtering
algorithm. First, we recall the video summarization problem. Second, we show
that the expressive ExistAllen constraint we introduced is very competitive with
the state-of-the-art dedicated approach.
 A Global Constraint for a Tractable Class 115

4.1 Problem Description

The video summarization problem of [4] consists in extracting audio-visual fea-

tures and computing segments from an input video. The goal is to provide a
summary of the video. More precisely, they consider tennis match records.
Several features (games, applause, speech, dominant color, etc.) are extracted
as a preprocessing step, in order to compute time intervals that describe the
video. Then, the problem is to compute segments of the video that will consti-
tute the summary. The number of segments to compute and their minimal and
maximal duration are given as parameter, as well as the summary duration. In
this case-study, the purpose is to build a tennis match summary with a duration
between four and five minutes, composed of ten segments, whose duration varies
between 10 and 120 seconds.
In order to obtain a good summary (from the end-user point of view), this
process is subject to constraints, such as:

– (1a) a segment should not cut a speech interval,

– (1b) a segment should not cut a game interval,
– (2) each selected segment must contain an applause interval,
– (3) the cardinality of the intersection between the segments and the dominant
color intervals must be at least one third of the summary,

On the assumption that an applause indicates an interesting action, the pres-

ence of applause in the summary must be maximized, i.e., the cardinality of the
intersection between the computed segments and the applause time intervals
must be as large as possible.

Table 3. Match features.

Name Total duration # Speech # Applause # Dominant color # Games

M1 2h08 571 271 1323 156
M2 1h22 332 116 101 66
M3 3h03 726 383 223 194

4.2 Benchmark

We consider the model implementation as well as a 3-instance dataset (see

Table 3) kindly provided by Boukadida et. al. [4].
The model is encoded using integer variables. A segment is represented by
three variables to indicate its start, duration and end. If constraints (1a) and (1b)
are easily ensured by forbidding values for the segment start and end variables,
most constraints have been encoded using ad hoc propagators. This holds on
constraint (2), whereas it could be handled with a single ExistAllen constraint
wherein T is the set of selected segments, R is equal to {f i, di, eq, si} and I is
the set of applause time intervals. Therefore, to evaluate the practical impact of
the linear-time ExistAllen propagator, four models are considered.
116 A. Derrien et al.

1. decomp: constraint (2) is explicitly represented by the disjunction depicted

in Section 3, Definition 1, using primitive constraints of the solver.
2. allen(n.m): constraint (2) is represented by an ExistAllen constraint, using
the quadratic propagator presented in section 3.2,
3. allen(n+m): constraint (2) is represented by an ExistAllen constraint, using
the linear propagator described in sections 3.3 and 3.4.
4. ad hoc: constraint (2) is represented with the dedicated constraints of [4].
Such model is given for reference only as neither its complexity nor its con-
sistency level are known.

100 M1

95

90

85

102
100 M2
98
objective

96
94
92
90

110
105 M3
100
95
90 ad hoc
85 allen(n+m)
80 allen(n.m)
75 decomp

1 10 100 1000
time (sec)

Fig. 2. Comparative evaluation of decomp, ad hoc, allen(n.m) and allen(n+m) on

the three matches M 1, M 2 and M 3 with a static search heuristic. The plots report
the evolution of the objective value to be maximized with respect to the cpu time in
seconds. The x-axis are in logscale.
 A Global Constraint for a Tractable Class 117

Each of the four models has been implemented in Choco-3.3.0 [16]. Each of
the instances was executed with a 15 minutes timeout, on a Macbook Pro with
8-core Intel Xeon E5 at 3Ghz running a MacOS 10.10, and Java 1.8.0 25. Each
instance was run on its own core, each with up to 4096MB of memory.
In order to compare the efficiency on the four models, we first consider a static
search heuristic: the variables representing the segment bounds are selected in a
lexicographic order and assigned to their lower bounds. In this way, we ensure
that the same search tree is explored, finding the same solutions in the same

220
190
160
130 M1
100
70
40
10

180
170
160
objective

150 M2
140
130
120

240
210
180
150
M3
120 ad hoc
90 allen(n+m)
allen(n.m)
60
decomp
30
1 10 100 1000
time (sec)

Fig. 3. Comparative evaluation of decomp, ad hoc,allen(n.m) and allen(n+m) on the

three matches M 1, M 2 and M 3 with ABS and PGLNS. The plots report the evolution
of the objective value to be maximized with respect to the cpu time in seconds. The
x-axis are in logscale.
118 A. Derrien et al.

order, and that only the running time is evaluated. The comparative evaluations
of the four models are in reported in Figure 2. Each plot is associated with
an instance and indicates the improvement of the objective function over time.
Recall that the x-axis are in logscale. The three plots are similar.
First of all, a strict reformulation (decomp) is clearly not competitive with
the models with specific constraints: decomp is always the slowest. This has to
do with the huge number of constraints and additional variables it requires to
express the ExistAllen constraint. As an example in the match M 1 where there
are 10 tasks, 271 intervals and 4 relations, each triplet T, R, I is expressed by
three new binary constraints, each of them reified by a new boolean variable.
Second, the quadratic propagator improves the filtering of a strict reformulation.
Third, as expected, the performances are even better with the linear propaga-
tor. For information purpose, the results of the ad hoc model (as used in [4])
are reported. allen(n+m), our generic constraint is shown to be the most effec-
tive model, mildly faster than ad hoc. This confirms that our generic linear
propagator, besides being expressive and flexible and offering guarantees on its
complexity and consistency level, is very efficient in practice.
The four models have also been evaluated with the configuration described
in [4], that is, using Activity-based search [13] combined with Propagation-
Guided Large Neighborhood Search [14].1 Due to the intrinsic randomness of
ABS and PGLNS, each resolution was run 30 times. Thus, the average objec-
tive values are reported in Figure 3. Recall that the x-axis are in logscale. The
dynamic strategy offered by the combination of ABS and PGLNS enables to
reduce the differences between the various models. Although to a slightly lesser
extent, the order between efficiency of the four models is preserved when apply-
ing a more aggressive search strategy heuristic.

5 Conclusion
We introduced ExistAllen, a generic constraint defined on a vector of tasks
and a set of disjoint intervals, which applies on any of the 213 combinations
of Allen’s algebra relations. This constraint is useful to tackle many problems
related to time intervals, such as the video-summarization problem [4], used as a
case study. From a technical viewpoint, we proposed a generic propagator that
achieves bound-consistency in O(n + m) worst-case time, where n is the number
of tasks and m the number of intervals, whereas the most natural implemen-
tation requires O(n × m) worst-case time. Our experiments demonstrate that
using our technique is very competitive with the best ad hoc approach, specific
to one particular combination of relations, while being much more expressive.
Future work includes the extension of this approach to several task sets,
in order to tackle problems beyond the context of video-summarization. In the
Satellite Data Download Management Problem [15], Earth observation satellites
acquire data that need to be downloaded to the ground. The download of data
is subject to temporal constraints, such as fitting in visibility windows. Using
1
The configuration for ABS and PGLNS are the default one described in [13] and [14].
 A Global Constraint for a Tractable Class 119

ExistAllen constraint, the visibility windows are the intervals, the amount of
acquired data fixes the duration of the download tasks, and the relation required
is during. The tasks can be ordered with respect to the type of acquired data
and their need to be scheduled.

Acknowledgments. The authors thank Haykel Boukadida, Sid-Ahmed Berrani and

Patrick Gros for helping us modeling the tennis match summarization problem and for
having providing us their dataset.

References
1. Allen, J.F.: Maintaining knowledge about temporal intervals. Commun. ACM
26(11), 832–843 (1983)
2. Berrani, S.-A., Boukadida, H., Gros, P.: Constraint satisfaction programming for
video summarization. In: Proceedings of the 2013 IEEE International Symposium
on Multimedia, ISM 2013, Washington, DC, USA, pp. 195–202. IEEE Computer
Society (2013)
3. Bessière, C.: Constraint propagation. Research report 06020. In: Rossi, F.,
van Beek, P., Walsh, T. (eds.) Handbook of Constraint Programming, LIRMM,
chapter 3. Elsevier (2006)
4. Boukadida, H., Berrani, S.-A., Gros, P.: A novel modeling for video summarization
using constraint satisfaction programming. In: Bebis, G., et al. (eds.) ISVC 2014,
Part II. LNCS, vol. 8888, pp. 208–219. Springer, Heidelberg (2014)
5. Bramsen, P., Deshp, P., Lee, Y.K., Barzilay, R.: Finding temporal order in dis-
charge summaries, pp. 81–85 (2006)
6. Choueiry, B.Y., Lin, X.: An efficient consistency algorithm for the temporal con-
straint satisfaction problem. AI Commun. 17(4), 213–221 (2004)
7. Dechter, R., Meiri, I., Pearl, J.: Temporal constraint networks. Artif. Intell.
49(1–3), 61–95 (1991)
8. Ibrahim, Z.A.A., Ferrane, I., Joly, P.: Temporal relation analysis in audiovi-
sual documents for complementary descriptive information. In: Detyniecki, M.,
Jose, J.M., Nürnberger, A., van Rijsbergen, C.J.K. (eds.) AMR 2005. LNCS,
vol. 3877, pp. 141–154. Springer, Heidelberg (2006)
9. Koubarakis, M.: From local to global consistency in temporal constraint networks.
Theor. Comput. Sci. 173(1), 89–112 (1997)
10. Koubarakis, M., Skiadopoulos, S.: Querying temporal and spatial constraint
networks in PTIME. Artif. Intell. 123(1–2), 223–263 (2000)
11. Satish Kumar, T.K., Cirillo, M., Koenig, S.: Simple temporal problems with taboo
regions. In: des Jardins, M., Littman, M.L. (eds.) Proceedings of the Twenty-
Seventh AAAI Conference on Artificial Intelligence, July 14–18, 2013, Bellevue,
Washington, USA. AAAI Press (2013)
12. Ladkin, P.B.: Satisfying first-order constraints about time intervals. In:
Shrobe, H.E., Mitchell, T.M., Smith, R.G. (eds.) Proceedings of the 7th
National Conference on Artificial Intelligence. St. Paul, MN, August 21–26, 1988,
pp. 512–517. AAAI Press/The MIT Press (1988)
13. Michel, L., Van Hentenryck, P.: Activity-based search for black-box constraint
programming solvers. In: Beldiceanu, N., Jussien, N., Pinson, É. (eds.) CPAIOR
2012. LNCS, vol. 7298, pp. 228–243. Springer, Heidelberg (2012)
120 A. Derrien et al.

14. Perron, L., Shaw, P., Furnon, V.: Propagation guided large neighborhood search.
In: Proceedings of the Principles and Practice of Constraint Programming - CP
2004, 10th International Conference, CP 2004, Toronto, Canada, pp. 468–481,
September 27–October 1, 2004
15. Pralet, C., Verfaillie, G., Maillard, A., Hebrard, E., Jozefowiez, N., Huguet, M.-J.,
Desmousceaux, T., Blanc-Paques, P., Jaubert, J.: Satellite data download man-
agement with uncertainty about the generated volumes. In: Chien, S., Do, M.B.,
Fern, A., Ruml, W. (eds.) Proceedings of the Twenty-Fourth International Con-
ference on Automated Planning and Scheduling, ICAPS 2014, Portsmouth, New
Hampshire, USA, June 21–26, 2014. AAAI (2014)
16. Prud’homme, C., Fages, J.-G., Lorca, X.: Choco3 Documentation. http://www.
choco-solver.org. TASC, INRIA Rennes, LINA CNRS UMR 6241, COSLING
S.A.S. (2014)
17. Schilder, F.: Event extraction and temporal reasoning in legal documents. In:
Schilder, F., Katz, G., Pustejovsky, J. (eds.) Annotating, Extracting and Rea-
soning about Time and Events. LNCS (LNAI), vol. 4795, pp. 59–71. Springer,
Heidelberg (2007)
18. van Beek, P., Cohen, R.: Exact and approximate reasoning about temporal rela-
tions. Computational Intelligence 6, 132–144 (1990)
19. Vilain, M., Kautz, H., van Beek, P.: Readings in qualitative reasoning about
physical systems. In: Constraint Propagation Algorithms for Temporal Reasoning:
A Revised Report, pp. 373–381. Morgan Kaufmann Publishers Inc., San Francisco
(1990)
20. Zhou, L., Hripcsak, G.: Methodological review: Temporal reasoning with medical
data-a review with emphasis on medical natural language processing. J. of Biomed-
ical Informatics 40(2), 183–202 (2007)
 Exploiting GPUs in Solving (Distributed)
Constraint Optimization Problems
with Dynamic Programming

Ferdinando Fioretto1,2(B) , Tiep Le1 , Enrico Pontelli1 , William Yeoh1 ,

and Tran Cao Son1
1
Department of Computer Science, New Mexico State University,
Las Cruces, NM, USA
{ffiorett,tile,epontell,wyeoh,tson}@cs.nmsu.edu
2
Department of Mathematics and Computer Science,
University of Udine, Udine, Italy

Abstract. This paper proposes the design and implementation of a

dynamic programming based algorithm for (distributed) constraint opti-
mization, which exploits modern massively parallel architectures, such as
those found in modern Graphical Processing Units (GPUs). The paper
studies the proposed algorithm in both centralized and distributed opti-
mization contexts. The experimental analysis, performed on unstruc-
tured and structured graphs, shows the advantages of employing GPUs,
resulting in enhanced performances and scalability.

1 Introduction

The importance of constraint optimization is outlined by the impact of its appli-

cation in a range of Constraint Optimization Problems (COPs), such as sup-
ply chain management (e.g., [15,27]) and roster scheduling (e.g., [1,8]). When
resources are distributed among a set of autonomous agents and communication
among the agents are restricted, COPs take the form of Distributed Constraint
Optimization Problems (DCOPs) [21,33]. In this context, agents coordinate their
value assignments to maximize the overall sum of resulting constraint utilities.
DCOPs are suitable to model problems that are distributed in nature, and where
a collection of agents attempts to optimize a global objective within the confines
of localized communication. They have been employed to model various dis-
tributed optimization problems, such as meeting scheduling [20,32,35], resources
allocation [13,36], and power network management problems [17].
Dynamic Programming (DP) based approaches have been adopted to solve
COPs and DCOPs. The Bucket Elimination (BE) procedure [10] iterates over
the variables of the COP, reducing the problem at each step by replacing a

This research is partially supported by the National Science Foundation under grant
number HRD-1345232. The views and conclusions contained in this document are
those of the authors and should not be interpreted as representing the official policies,
either expressed or implied, of the sponsoring organizations, agencies, or the U.S.
government.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 121–139, 2015.
DOI: 10.1007/978-3-319-23219-5 9
122 F. Fioretto et al.

variable and its related utility functions with a single new function, derived
by optimizing over the possible values of the replaced variable. The Dynamic
Programming Optimization Protocol (DPOP) [25] is one of the most efficient
DCOP solvers, and it can be seen as a distributed version of BE, where agents
exchange newly introduced utility functions via messages.
The importance of DP-based approaches arises in several optimization fields
including constraint programming [2,28]. For example, several propagators adopt
DP-based techniques to establish constraint consistency; for instance, (1) the
knapsack constraint propagator proposed by Trick applies DP techniques to
establish arc consistency on the constraint [31]; (2) the propagator for the regular
constraint establishes arc consistency using a specific digraph representation of
the DFA, which has similarities to dynamic programming [24]; (3) the context
free grammar constraint makes use of a propagator based on the CYK parser
that uses DP to enforce generalized arc consistency [26].
While DP approaches may not always be appropriate to solve (D)COPs, as
their time and space requirements may be prohibitive, they may be very effective
in problems with particular structures, such as problems where their underlying
constraint graphs have small induced widths or distributed problems where the
number of messages is crucial for performance, despite the size of the messages.
The structure exploited by DP-based approaches in constructing solutions makes
it suitable to exploit a novel class of massively parallel platforms that are based
on the Single Instruction Multiple Thread (SIMT) paradigm—where multiple
threads may concurrently operate on different data, but are all executing the
same instruction at the same time. The SIMT-based paradigm is widely used
in modern Graphical Processing Units (GPUs) for general purpose parallel com-
puting. Several libraries and programming environments (e.g., Compute Unified
Device Architecture (CUDA)) have been made available to allow programmers
to exploit the parallel computing power of GPUs.
In this paper, we propose a design and implementation of a DP-based algo-
rithm that exploits parallel computation using GPUs to solve (D)COPs. Our
proposal aims at employing GPU hardware to speed up the inference process of
DP-based methods, representing an alternative way to enhance the performance
of DP-based constraint optimization approaches. This paper makes the following
contributions: (1) We propose a novel design and implementation of a central-
ized and a distributed DP-based algorithm to solve (D)COPs, which harnesses
the computational power offered by parallel platforms based on GPUs; (2) We
enable the use of concurrent computations between CPU(s) and GPU(s), during
(D)COP resolution; and (3) We report empirical results that show significant
improvements in performance and scalability.

2 Background
2.1 Centralized Constraint Optimization Problems (COPs)
A (centralized) Constraint Optimization Problem (COP) is defined as (X, D, C)
where: X = {x1 , . . . , xn } is a set of variables; D = {D1 , . . . , Dn } is a set of
Exploiting GPUs in Solving (Distributed) Constraint Optimization Problems 123

x1 x2 Utilities x3
x1 for i < j
x1 0 0 max(5+5, 8+8) = 16 1
xi x j Utilities 0 1 max(5+20, 8+3) = 25 0
0 0 5 1 0 max(20+5, 3+8) = 25 0
x2 x2 0 1 8 1 1 max(20+20, 3+3) = 40 0
1 0 20
x1 Utilities x2 x3
1 1 3
x3 0 max(5+16, 8+25) = 33 1 0
x3
1 max(20+25, 3+40) = 45 0 0
(a) (b) (c) (d)

Fig. 1. Example (D)COP (a-c) and UTIL phase computations in DPOP (d).

domains for the variables in X, where Di is the set of possible values for the
variable xi ; C is a finite set of utility functions on variables in X, with fi :
xj ∈xi Dj → R+ ∪ {−∞}, where xi ⊆ X is the set of variables relevant to fi ,
referred to as the scope of fi , and −∞ is used to denote that a given combination
of values for the variables in xi is not allowed.1 A solution is a value assignment
for a subset of variables from X that is consistent
n with their respective domains;
i.e., it is a partial function θ : X → i=1 Di such that, for each xj ∈ X, if
θ(xj ) is defined, then θ(xj ) ∈ Dj . A solution is complete if it assigns a value to
each variable in X. We will use the notation σ to denote a complete solution,
and, for a set of variables V = {xi1 , . . . , xih } ⊆ X, σV = σ(xi1 ), . . . , σ(xih ),
where i1 < · · · < ih . The goal for a COP is to find a complete solution σ ∗
that maximizes the  total problem utility expressed by its utility functions, i.e.,
σ ∗ = argmaxσ∈Σ fi ∈C fi (σxi ), where Σ is the state space, defined as the set of
all possible complete solutions.
Given a COP P , GP = (X, EC ) is the constraint graph of P , where {x, y} ∈
EC iff ∃fi ∈ C such that {x, y} ⊆ xi . Fig. 1(a) shows the constraint graph of a
simple COP with three variables, x1 , x2 , and x3 . The domain of each variable is
the set {0, 1}. Fig. 1(c) describes the utility functions of the COP.

Definition 1 (Projection). The projection of a utility function fi on a set of

variables V ⊆ xi is a new utility function fi|V : V → R+ ∪ {−∞}, such that for
each possible assignment θ ∈ xj ∈V Dj , fi|V (θ) = max fi (σxi ).
σ∈Σ,σV =θ

In other words, fi|V is constructed from the tuples of fi , removing the values of
the variable that do not appear in V and removing duplicate values by keeping
the maximum utility of the original tuples in fi .
Definition 2 (Concatenation). Let us consider two assignments θ , defined
for variables V , and θ , defined for variables W , such that for each x ∈ V ∩ W
we have that θ (x) = θ (x). Their concatenation is an assignment θ · θ defined
for V ∪ W , such as for each x ∈ V (resp. x ∈ W ) we have that θ · θ (x) = θ (x)
(resp. θ · θ (x) = θ (x)).
1
For simplicity, we assume that tuples of variables are built according to a predefined
ordering.
124 F. Fioretto et al.

Algorithm 1. BE
1 for i ← n downto 1 do
2 Bi ← {fj ∈ C | xi ∈ xj ∧ i = max{k | xk ∈ xj }}
 
3 fˆi ← π−xi fj
fj ∈Bi
4 X ← X \ {xi }
5 C ← (C ∪ {fˆi }) \ Bi

We define two operations on utility functions:

• The aggregation of two functions fi and fj , is a function fi + fj : xi ∪ xj →
R+ ∪ {−∞}, such that ∀θ ∈ xk ∈xi Dk and ∀θ ∈ xk ∈xj Dk , if θ · θ is
defined, then we have that (fi + fj )(θ · θ ) = fi (θ ) + fj (θ ).
• Projecting out a variable xj ∈ xi from a function fi , denoted as π−xj (fi ),
produces a new function with scope xi \ {xj }, and defined as the projection
of fi on xi \ {xj }, i.e., π−xj (fi ) = fi|xi {xj } .
Bucket Elimination (BE): BE [10,11] is a dynamic programming based pro-
cedure that can be used to solve COPs. Algorithm 1 illustrates its pseudocode.
Given a COP (X, D, C) and an ordering o = x1 , . . . , xn  on the variables in X,
we say that a variable xi has a higher priority with respect to variable xj if xi
appears after xj in o. BE operates from the highest to lowest priority variable.
When operating on variable xi , it creates a bucket Bi , which is the set of all
utility functions that involve xi as the highest priority variable in their scope
(line 2). The algorithm then computes a new utility function fˆi by aggregating
the functions in Bi and projecting out xi (line 3). Thus, xi can be removed from
the set of variables X to be processed (line 4) and the new function fˆi replaces
in C all the utility functions that appear in Bi (line 5). In our example, BE
operates, in order, on the variables x3 , x2 , and x1 . When x3 is processed, the
bucket B3 is {f13 , f23 }, and the fˆ3 utility function is shown in Fig. 1(d) top.
The rightmost column shows the values for x3 after its projection. BE updates
the sets X = {x1 , x2 } and C = {f12 , fˆ3 }. When x2 is processed, B2 = {f12 , fˆ3 }
and fˆ2 is shown in Fig. 1(d) bottom. Thus, X = {x1 } and C = {fˆ2 }. Lastly, the
algorithm processes x1 , sets B1 = {fˆ2 }, and fˆ1 contains one value combination
σ ∗ = 1, 0, 0, which corresponds to an optimal solution to the problem.
The complexity of the algorithm is bounded by the time needed to process
a bucket (line 3), which is exponential in number of variables in the bucket.

2.2 Distributed Constraint Optimization Problems (DCOPs)

In a Distributed Constraint Optimization Problem (DCOP) [21,25,33], the vari-

ables, domains, and utility functions of a COP are distributed among a collection
of agents. A DCOP is defined as (X, D, C, A, α), where X, D, and C are defined
as in a COP, A = {a1 , . . . , ap } is a set of agents, and α : X → A maps each
variable to one agent. Following common conventions, we restrict our attention
Exploiting GPUs in Solving (Distributed) Constraint Optimization Problems 125

to binary utility functions and assume that α is a bijection: Each agent con-
trols exactly one variable. Thus, we will use the terms “variable” and “agent”
interchangeably and assume that α(xi ) = ai . This is a common assumption in
the DCOP literature as there exist pre-processing techniques that transform a
general DCOP into this more restrictive DCOP [7,34]. In DCOPs, solutions are
defined as for COPs, and many solution approaches emulate those proposed in
the COP literature. For example, ADOPT [21] is a distributed version of Itera-
tive Deepening Depth First Search, and DPOP [25] is a distributed version of BE.
The main difference is in the way the information is shared among agents. Typi-
cally, a DCOP agent knows exclusively its domain and the functions involving its
variable. It can communicate exclusively with its neighbors (i.e., agents directly
connected to it in the constraint graph2 ), and the exchange of information takes
the form of messages. Given a DCOP P , a DFS pseudo-tree arrangement for GP
is a spanning tree T = X, ET  of GP such that if fi ∈ C and {x, y} = xi , then x
and y appear in the same branch of T . Edges of GP that are in (resp. out of)
ET are called tree edges (resp. backedges). The tree edges connect parent-child
nodes, while backedges connect a node with its pseudo-parents and its pseudo-
children. We use N (ai ) = {aj ∈ A|{xi , xj } ∈ ET } to denote the neighbors of agent
ai ; Ci , P Ci , Pi , and PPi to denote the set of children, pseudo-children, parent,
and pseudo-parents of agent ai ; and sep(ai ) to denote the separator of agent
ai , which is the set of ancestor agents that are constrained (i.e., they are linked
in GP ) with agent ai or with one of its descendant agents in the pseudo-tree.
Fig. 1(b) shows one possible pseudo-tree for the problem, where the agent a1
has one pseudo-child a3 (the dotted line is a backedge).
Dynamic Programming Optimization Protocol (DPOP): DPOP [25] is a
dynamic programming based DCOP algorithm that is composed of three phases.
(1) Pseudo-tree generation: Agents coordinate to build a pseudo-tree, realized
through existing distributed pseudo-tree construction algorithms [16]. (2) UTIL
propagation: Each agent, starting from the leaves of the pseudo-tree, computes
the optimal sum of utilities in its subtree for each value combination of variables
in its separator. The agent does so by aggregating the utilities of its functions
with the variables in its separator and the utilities in the UTIL messages received
from its child agents, and then projecting out its own variable. In our example
problem, agent a3 computes the optimal utility for each value combination of
variables x1 and x2 (Fig. 1(d) top), and sends the utilities to its parent agent
a2 in a UTIL message. When the root agent a1 receives the UTIL message from
each of its children, it computes the maximum utility of the entire problem.
(3) VALUE propagation: Each agent, starting from the root of the pseudo-
tree, determines the optimal value for its variable. The root agent does so by
choosing the value of its variable from its UTIL computations—selecting the
value with the maximal utility. It sends the selected value to its children in a
VALUE message. Each agent, upon receiving a VALUE message, determines
the value for its variable that results in the maximum utility given the variable

2
The constraint graph of a DCOP is equivalent to that of the corresponding COP.
126 F. Fioretto et al.

assignments (of the agents in its separator) indicated in the VALUE message.
Such assignment is further propagated to the children via VALUE messages.
The complexity of DPOP is dominated by the UTIL propagation phase, which
is exponential in the size of the largest separator set sep(ai ) for all ai ∈ A. The
other two phases require a polynomial number of linear size messages, and the
complexity of the local operations is at most linear in the size of the domain.
Observe that the UTIL propagation phase of DPOP emulates the BE process
in a distributed context [6]. Given a pseudo-tree and its preorder listing o, the
UTIL message generated by each DPOP agent ai is equivalent to the aggregated
and projected function fˆi in BE when xi is processed according to the ordering o.

2.3 Graphical Processing Units (GPUs)

Modern GPUs are multiprocessor devices, offering hundreds of computing cores
and a rich memory hierarchy to support graphical processing. We consider the
NVIDIA CUDA programming model [29], which enables the use of the mul-
tiple cores of a graphics card to accelerate general (non-graphical) applica-
tions. The underlying model of parallelism is Single-Instruction Multiple-Thread
(SIMT), where the same instruction is executed by different threads that run
on identical cores, grouped in Streaming Multiprocessors (SMs), while data and
operands may differ from thread to thread.
A typical CUDA program is a C/C++ program. The functions in the pro-
gram are distinguished based on whether they are meant for execution on the
CPU (referred to as the host) or in parallel on the GPU (referred as the device).
The functions executed on the device are called kernels, and are executed by sev-
eral threads. To facilitate the mapping of the threads to the data structures being
processed, threads are grouped in blocks, and have access to several memory lev-
els, each with different properties in terms of speed, organization, and capacity.
CUDA maps blocks (coarse-grain parallelism) to the SMs for execution. Each
SM schedules the threads in a block (fine-grain parallelism) on its computing
cores in chunks of 32 threads (warps) at a time. Threads in a block can com-
municate by reading and writing a common area of memory (shared memory).
Communication between blocks and communication between the blocks and the
host is realized through a large slow global memory. The development of CUDA
programs that efficiently exploit SIMT parallelism is a challenging task. Several
factors are critical in gaining performance. Memory levels have significantly dif-
ferent sizes (e.g., registers are in the order of dozens per thread, shared memory is
in the order of a few kilobytes per block) and access times, and various optimiza-
tion techniques are available (e.g., coalesced of memory accesses to contiguous
locations into a single memory transaction).

3 GPU-Based (Distributed) Bucket Elimination

(GPU-(D)BE)
Our GPU-based (Distributed) Bucket Elimination framework, extends BE (resp.
DPOP) by exploiting GPU parallelism within the aggregation and projection
Exploiting GPUs in Solving (Distributed) Constraint Optimization Problems 127

Algorithm 2. GPU-(D)BE
(1) Generate pseudo-tree
2 GPU-Initialize( )
3 if Ci = ∅ then
4 UTILxi ⇔ ParallelCalcUtils( )
(5) Send UTIL message (xi , UTILxi ) to Pi
6 else
7 Activate UTILMessageHandler(·)
(8) Activate VALUEMessageHandler(·)

operations. These operations are responsible for the creation of the functions fˆi
in BE (line 3 of Algorithm 1) and the UTIL tables in DPOP (UTIL propagation
phase), and they dominate the complexity of the algorithms. Thus, we focus on
the details of the design and the implementation relevant to such operations.
Due to the equivalence of BE and DPOP, we will refer to the UTIL tables and
to the aggregated and projected functions fˆ of Algorithm 1, as well as variables
and agents, interchangeably. Notice that the computation of the utility for each
value combination in a UTIL table is independent of the computation in the
other combinations. The use of a GPU architecture allows us to exploit such
independence, by concurrently exploring several combinations of the UTIL table,
computed by the aggregation operator, as well as concurrently projecting out
variables.
Algorithm 2 illustrates the pseudocode, where we use the following notations:
Line numbers in parenthesis denote those instructions required exclusively in the
distributed case. Starred line numbers denote those instructions executed concur-
rently by both the CPU and the GPU. The symbols ← and ⇔ denote sequential
and parallel (multiple GPU-threads) operations, respectively. If a parallel oper-
D←H H←D
ation requires a copy from host (device) to device (host), we write ⇔ ( ⇔ ).
Host to device (resp. device to host) memory transfers are performed immedi-
ately before (resp. after) the execution of the GPU kernel. Algorithm 2 shows
the pseudocode of GPU-(D)BE for an agent ai . Like DPOP, also GPU-(D)BE
is composed of three phases; the first and third phase are executed exclusively
in the distributed version. The first phase is identical to that of DPOP (line 1).
In the second phase:
• Each agent ai calls GPU-Initialize() to set up the GPU kernel. For example,
it determines the amount of global memory to be assigned to each UTIL table
and initializes the data structures on the GPU device memory (line 2).
• Each agent ai aggregates the utilities for the functions between its variables
and its separator, projects its variable out (line 4), and sends them to its
parent (line 5). The MessageHandlers of lines 7 and 8 are activated for each
new incoming message.
By the end of the second phase (line 11), the root agent knows the overall utility
for each values of its variable xi . It chooses the value that results in the maximum
128 F. Fioretto et al.

Procedure UTILMessageHandler(ak , UTILak )

(9) Store UTILak
10 if received UTIL message from each child ac ∈ Ci then
11 UTILai ⇔ ParallelCalcUtils( )
12 if Pi = NULL then
13 d∗i ← ChooseBestValue(∅)
(14) foreach ac ∈ Ci do
(15) VALUEai ← (xi , d∗i )
(16) Send VALUE message (ai , VALUEai ) to ac
(17) else Send UTIL message (ai , UTILai ) to Pi

Procedure VALUEMessageHandler(ak , VALUEak )

(18) VALUEai ← VALUEak
(19) d∗i ← ChooseBestValue(VALUEai )
(20) foreach ac ∈ Ci do
(21) VALUEai ← {(xi , d∗i )} ∪ {(xk , d∗k ) ∈ VALUEak | xk ∈ sep(ac )}
(22) Send VALUE message (ai , VALUEai ) to ac

utility (line 13). Then, in the distributed version, it starts the third phase by
sending to each child agent ac the value of its variable xi (lines 14-16). These
operations are repeated by every agent receiving a VALUE message (lines 18-22).
In contrast, in the centralized version, the value assignment for each variable is
set by the root agent directly.

3.1 GPU Data Structures

In order to fully capitalize on the parallel computational power of GPUs, the

data structures need to be designed in such a way to limit the amount of infor-
mation exchanged between the CPU host and the GPU device, and in order to
minimize the accesses to the (slow) device global memory (and ensure that they
are coalesced). To do so, each agent identifies the set of relevant static entities,
i.e., information required during the GPU computation, which does not mutate
during the resolution process. The static entities are communicated to the GPU
once at the beginning of the computation. This allows each agent running on
a GPU device to communicate with the CPU host exclusively to exchange the
results of the aggregation and projection processes. The complete set of utility
functions, the constraint graph, and the agents ordering, all fall in such category.
Thus, each agent ai stores:
• The set of utility functions involving exclusively xi and a variable in ai ’s
separator set: Si = {fj ∈ C | xi ∈ xj ∧ sep(ai ) ∩ xj = ∅}. For a given function
fj ∈ Si , its utility values are stored in an array named gFunc j .
Exploiting GPUs in Solving (Distributed) Constraint Optimization Problems 129

• The domain Di of its variable (for simplicity assumed to be all of equal car-
dinality).
• The set Ci of ai ’s children.
• The separator sets sep(ai ), and sep(ac ), for each ac ∈ Ci .
The GPU-Initialize() procedure of line 2, invoked after the pseudo-tree
construction, stores the data structures above for each agent on the GPU device.
As a technical detail, all the data stored on the GPU global memory is organized
in mono-dimensional arrays, so as to facilitate coalesced memory accesses. In
particular, the identifier and scope of the functions in Si as well as identifiers and
separator sets of child agents in Ci are stored within a single mono-dimensional
array. The utility values stored in the rows of each function are padded to ensures
that a row is aligned to a memory word—thus minimizing the number of memory
accesses.
GPU-Initialize() is also responsible for reserving a portion of the GPU
global memory to store the values for the agent’s UTIL table, denoted by gUtils i ,
and those of its children, denoted by gChUtils c , for each ac ∈ Ci . As a techni-
cal note, an agent’s UTIL table is mapped onto the GPU device to store only
the utility values, not the associated variables values. Its j-th entry is associated
with the j-th permutation of the variable values in sep(ai ), in lexicographic order.
This strategy allows us to employ a simple perfect hashing to efficiently associate
row numbers with variables’ values and vice versa. Note that the agent’s UTIL
table size grows exponentially with the size of its separator set; more precisely,
after projecting out xi , it has |Di |sep(ai ) entries. However, the GPU global mem-
ory is typically limited to a few GB (e.g., in our experiments it is 2GB). Thus,
each agent, after allocating its static entities, checks if it has enough space to
allocate its children’s UTIL tables and a consistent portion (see next subsection
for details) of its own UTIL table. In this case, it sets the project on device flag
to true, which signals that both aggregate and project operations can be done
on the GPU device.3 Otherwise it sets the flag to false and bounds the device
UTIL size table to the maximum storable space on the device. In this case, the
aggregation operations are performed only partially on the GPU device.

3.2 Parallel Aggregate and Project Operations

The ParallelCalcUtils procedure (executed in lines 4 and 11) is responsible
for performing the aggregation and projection operations, harnessing the paral-
lelism provided by the GPU. Due to the possible large size of the UTIL tables,
we need to separate two possible cases and devise specific solutions accordingly:
(a) When the device global memory is sufficiently large to store all ai ’s children
UTIL tables as well as a significant portion of ai ’s UTIL table4 (i.e., when
3
If the UTIL table of agent ai does not fit in the global memory, we partition such
table in smaller chunks, and iteratively execute the GPU kernel until all rows of the
table are processed.
4
In our experiments, we require that at least 1/10 of the UTIL table can be stored
in the GPU. We experimentally observed that a partitioning of the table in at most
130 F. Fioretto et al.

Procedure ParallelCalcUtils( )
23 if project on device then
D←H
24 gChUTILac ⇔ UTILac for all ac ∈ Ci
25 R ← 0 ; U T ILai ← ∅
26 while R < |Di |sep(ai ) do
27 if project on device then
H←D
28* UTILai ⇔ GPU-Aggregate-Project(R)
29 else
H←D
30* UTILai ⇔ GPU-Aggregate(R)
31* UTILai ← AggregateCh-Project(ai , UTILai , UTILac ) for all
ac ∈ Ci
32* UTILai ← U T ILai ∪ Compress(UTILai )
33 R ← R + |UTILai |
34 return UTILai

project on device = true), both aggregation and projection of the agent’s

UTIL table are performed in parallel on the GPU. The procedure first stores
the UTIL tables received from the children of ai into their assigned locations
in the GPU global memory (lines 23-24). It then iterates through successive
GPU kernel calls (line 28) until the UTILai table is fully computed (lines 26-
33). Each iterations computes a certain number of rows of the UTILai table
(R serves as counter).
(b) When the device global memory is insufficiently large to store all ai ’s chil-
dren UTIL tables as well as a significant portion of ai ’s UTIL table (i.e.,
when project on device = false), the agent alternates the use of the GPU
and the CPU to compute UTILai . The GPU is in charge of aggregating the
functions in Si (line 30), while the CPU aggregates the children UTIL table,5
projecting out xi . Note that, in this case, the UTILai storage must include
all combinations of values for the variables in sep(xi ) ∪ {xi }, thus the pro-
jection operation is performed on the CPU host. As in the previous case,
the UTILai is computed incrementally, given the amount of available GPU
global memory.
To fully capitalize on the use of the GPU, we exploit an additional level of
parallelism, achieved by running GPU kernels and CPU computations concur-
rently; this is possible when the UTILai table is computed in multiple chunks.
Fig. 2 illustrates the concurrent computations between the CPU and GPU. After
transferring the children UTIL tables into the device memory (Init)—in case
10 chunks provides a good time balance between memory transfers and actual com-
putation.
5
The CPU aggregates only those child UTIL table that could not fit in the GPU
memory. Those that fit in memory are integrated through the GPU computation as
done in the previous point.
Exploiting GPUs in Solving (Distributed) Constraint Optimization Problems 131

(Init) Compress U 1 Compress U 2

CPU
(Host) Copy Execute K 1 Copy Execute K 2 Copy

GPU
(Device)
H D Compute U 1 D H Compute U 2 D H

Update Global Mem. Update Global Mem.

Fig. 2. Concurrent computation between host and device.

(a) only—the execution of kernel K1 produces the update of the first chunk of
UTILai , denoted by U1 in Fig. 2, which is transferred to the CPU host. The
successive parallel operations are performed asynchronously with respect to the
GPU, that is, the execution of the j-th CUDA kernel Kj (j > 1), returns the con-
trol immediately to the CPU, which concurrently operates a compression opera-
tion on the previously computed UTILai chunk (line 32), referred to as Uk−1 in
Fig. 2. For case (b), the CPU also executes concurrently the AggregateCh-
Project of line 31. We highlight the concurrent operations by marking with a
∗
symbol their respective lines in the procedure ParallelCalcUtils.
Technical Details: We now describe in more detail how we divide the workload
among parallel blocks, i.e., the mapping between the UTIL table rows and the
CUDA blocks. A total of T = 64 · k (1 ≤ k ≤ 16) threads (a block) are associ-
ated to the computation of T permutations of values for sep(ai ). The value k
depends on the architecture and it is chosen to maximize the number of con-
current threads running at the same time. In our experiments, we set k = 3.
The number of blocks is chosen so that the corresponding aggregate number
of threads does not exceed the total number of UTILai permutations currently
stored in the device. Let h be the number of stream multiprocessors of the GPU.
Then, the maximum number of UTIL permutations that can be computed con-
currently is M = h · T . In our experiments h = 14, and thus, M = 2688. Fig. 3
provides an illustration of the UTIL permutations computed in parallel on GPU.
The blocks Bi in each row are executed in parallel on different SMs. Within each
block, a total of (at most) 192 threads operate on as many entries of the UTIL
table.
The GPU kernel procedure is shown in lines 35-49. We surround line numbers
with | · | to denote parts of the procedure executed by case (b). The kernel takes
as input the number R of the UTIL table permutations computed during the
previous kernel calls. Each thread identifies its entry index rid within the table
chunk UTILai (line 35). It then assigns the shared memory allocated to local
arrays to store the static entities Si , Ci , and sep(ac ), for each ac ∈ Ci . In addi-
tion it reserves the space θ to store the assignments corresponding to the UTIL
permutation being computed by each thread, which is retrieved using the thread
entry index and the offset R (line 38). Decode implements a minimal perfect
hash function to convert the entry index of the UTIL table to its associated vari-
ables value permutation. Each thread aggregates the functions in Si (lines 42-44)
and the UTIL tables of ai ’s children (lines 45-47), for each element of its domain
132 F. Fioretto et al.

Fig. 3. GPU kernel parallel computations.

Procedure GPU-Aggregate-Project(R)
|35| rid ← the thread’s entry index of UTILi
|36| did ← the thread’s value index of Di
|37| |θ, Si |, Ci , sep(xc ) ← AssignSharedMem() for all xc ∈ Ci
|38| θ ← decode(R + rid )
|39| util ← −∞
40 foreach did ∈ Di do
|41| util did ← 0
|42| foreach fj ∈ Si do
|43| ρj ← encode(θxj | xi = did )
|44| util did ← util did + gFunc j [ρj ]
45 foreach ac ∈ Ci do
46 ρc ← encode(θsep(ac ) | xi = did )
47 util did ← util did + gChUtils c [ρc ]
|48| util ← max(util , util did )
|49| gUtils i [rid ] ← util

(lines 40-48). The Encode routine converts a given assignments for the variables
in the scope of a function fj (line 43), or in the separator set of child ac (line 46),
to the corresponding array index, sorted in lexicographic order. The value for
the variable xi within each input, is updated at each iteration of the for loop.
The projection operation is executed in line 48. Finally, the thread stores the
best utility in the corresponding position of the array gUtils i
The GPU-Aggregate procedure (called in line 30), is illustrated in
lines 35-49—line numbers surrounded by | · |. Each thread is in charge of
a value combination in sep(ai ) ∪ {xi }, thus, the foreach loop of lines 40-
48 is operated in parallel by |Di | threads. Lines 45-47 are not executed. The
AggregateCh-Project procedure (line 31), which operates on the CPU, is
similar to the GPU-Aggregate-Project procedure, except that lines 36-37,
and 42-44, are not executed.
The proposed kernel has been the result of several investigations. We exper-
imented with other levels of parallelism, e.g., by unrolling the for-loops among
Exploiting GPUs in Solving (Distributed) Constraint Optimization Problems 133

groups of threads. However, these modifications create divergent branches, which

degrade the parallel performance. We experimentally observed that such degra-
dation worsen consistently as the size of the domain increases.

3.3 General Observations

Observation 1. GPU-DBE requires the same number of messages as those

required by DPOP, and it requires messages of the same size as those required
by DPOP.
Observation 2. The UTIL messages constructed by each GPU-DBE agent are
identical to those constructed by each corresponding DPOP agent.
The above observations follow from the pseudo-tree construction and VALUE
propagation GPU-DBE phases, which are identical to those of DPOP. Thus, their
corresponding messages and message sizes are identical in both algorithms. More-
over, given a pseudo-tree, each DPOP/GPU-DBE agent computes the UTIL
table containing each combination of values for the variables in its separator set.
Thus, the UTIL messages of GPU-DBE and DPOP are identical.
Observation 3. The memory requirements of GPU-(D)BE is, in the worst case,
exponential in the induced width of the problem (for each agent).
This observation follows from the equivalence of the UTIL propagation phase of
DPOP and BE [6] and from Observation 2.
Observation 4. GPU-(D)BE is complete and correct.
The completeness and correctness of GPU-(D)BE follow from the completeness
and correctness of BE [10] and DPOP [25].

4 Related Work

The use of GPUs to solve difficult combinatorial problems has been explored
by several proposals in different areas of constraint optimization. For instance,
Meyer et al. [18] proposed a multi-GPU implementation of the simplex tableau
algorithm which relies on a vertical problem decomposition to reduce communi-
cation between GPUs. In constraint programming, Arbelaez and Codognet [3]
proposed a GPU-based version of the Adaptive Search that explores several large
neighborhoods in parallel, resulting in a speedup factor of 17. Campeotto et al. [9]
proposed a GPU-based framework that exploits both parallel propagation and
parallel exploration of several large neighborhoods using local search techniques,
leading to a speedup factor of up to 38. The combination of GPUs with dynamic
programming has also been explored to solve different combinatorial optimiza-
tion problems. For instance, Boyer et al. [5] proposed the use of GPUs to compute
the classical DP recursion step for the knapsack problem, which led to a speedup
factor of 26. Pawlowski et al. [23] presented a DP-based solution for the coalition
structure formation problem on GPUs, reporting up to two orders of magnitude
134 F. Fioretto et al.

of speedup. Differently from other proposals, our approach aims at using GPUs
to exploit SIMT-style parallelism from DP-based methods to solve general COPs
and DCOPs.

5 Experimental Results
We compare our centralized and distributed versions of GPU-(D)BE with BE [10]
and DPOP [25] on binary constraint networks with random, scale-free, and reg-
ular grid topologies. The instances for each topology are generated as follows:
Random: We create an n-node network, whose density p1 produces n (n−1) p1 
edges in total. We do not bound the tree-width, which is based on the underlying
graph.
Scale-free: We create an n-node network based on the Barabasi-Albert
model [4]: Starting from a connected 2-node network, we repeatedly add a new
node, randomly connecting it to two existing nodes. In turn, these two nodes
are selected with probabilities that are proportional to the numbers of their
connected edges. The total number of edges is 2 (n − 2) + 1.
Regular grid: We create an n-node network arranged as a rectangular grid,
where each internal node is connected to four neighboring nodes, while nodes
on the grid edges (resp. corners) are connected to two (resp. three) neighboring
nodes.
We generate 30 instances for each topology, ensuring that the underlying graph
is connected. The utility functions are generated using random integer costs in
[0, 100], and the constraint tightness (i.e., ratio of entries in the utility table
different from −∞) p2 is set to 0.5 for all experiments. We set as default param-
eters, |A| = |X| = 10, |Di | = 5 for all variables, and p1 = 0.3 for random networks,
and |A| = |X| = 9 for regular grids. Experiments for GPU-DBE are conducted
using a multi-agent DCOP simulator, that simulates the concurrent activities of
multiple agents, whose actions are activated upon receipt of a message. We use
the publicly-available implementation of DPOP available in the FRODO frame-
work v.2.11 [19], and we use the same framework to run the BE algorithm, in a
centralized setting.
Since all algorithms are complete, our focus is on runtime. Performance of
the centralized algorithms are evaluated using the algorithm’s wallclock runtime,
while distributed algorithms’ performances are evaluated using the simulated
runtime metric [30]. We imposed a timeout of 300s of wallclock (or simulated)
time and a memory limit of 32GB. Results are averaged over all instances and
are statistically significant with p-values < 1.638 e−12 .6 These experiment are
performed on an AMD Opteron 6276, 2.3GHz, 128GB of RAM, which is equipped
with a GPU device GeForce GTX TITAN with 14 multiprocessors, 2688 cores,
and a clock rate of 837MHz.
6
t-test performed with null hypothesis: GPU-based algorithms are faster than non-
GPU ones.
Exploiting GPUs in Solving (Distributed) Constraint Optimization Problems 135

Fig. 4. Runtimes for COPs (top) and DCOPs (bottom) at varying number of vari-
ables/agents.

Fig. 4 illustrates the runtime, in seconds, for random (a), scale-free (b), and
regular grid (c) topologies, varying the number of variables (resp. agents) for the
centralized (resp. distributed) algorithms. The centralized algorithms (BE and
GPU-BE) are shown at the top of the figure, while the distributed algorithms
(DPOP and GPU-DBE) are illustrated at the bottom. All plots are in log-scale.
We make the following observations:
• The GPU-based DP-algorithms (for both centralized and distributed cases)
are consistently faster than the non-GPU-based ones. The speedups obtained
by GPU-BE vs. BE are, on average, and minimum (showed in parenthesis)
69.3 (16.1), 34.9 (9.5), and 125.1 (42.6), for random, scale-free, and regular grid
topologies, respectively. For the distributed algorithms, the speedups obtained
by GPU-DBE vs. DPOP are on average (minimum) 44.7 (14.7), 22.3 (8.2), and
124.2 (38.8), for random, scale-free, and regular grid topologies, respectively.
• In terms of scalability, the GPU-based algorithms scale better than the non-
GPU-based ones. In addition, their scalability increases with the level of struc-
ture exposed by each particular topology. On random graphs, which have
virtually no structure, the GPU-based algorithms reach a timeout for instances
with small number of variables (25 variables—compared to 20 variables for the
non-GPU-based algorithms). On scale-free networks, the GPU-(D)BE algo-
rithms can solve instances up to 50 variables,7 while BE and DPOP reach a
timeout for instances greater than 40 variables. On regular grids, the GPU-
based algorithms can solve instances up to 100 variables, while the non-GPU-
based ones, fail to solve any instance with 36 or more variables.
We relate these observations to the size of the separator sets and, thus, the size
of the UTIL tables that are constructed in each problem. In our experiments,

7
With 60 variables, we reported 12/30 instances solved for GPU-(D)BE.
136 F. Fioretto et al.

Fig. 5. Runtimes for COPs (top) and DCOPs (bottom) at varying number of vari-
ables/agents.

we observe that the average sizes of the separator sets are consistently larger
in random graphs, followed by scale-free networks, followed by regular grids.
• Finally, the trends of the centralized algorithms are similar to those of the
distributed algorithms: The simulated runtimes of the DCOP algorithms are
consistently smaller than the wallclock runtimes of the COP ones.
Fig. 5 illustrates the behavior of the algorithms when varying the graph den-
sity p1 for the random graphs (a), and the domains size for random graphs (b)
and regular grids (c). As for the previous experiments, the centralized (resp. dis-
tributed) algorithms are shown on the top (resp. bottom) of the figure. We can
observe:
• The trends for the algorithms runtime, when varying both p1 and domains
size, are similar to those observed in the previous experiments.
• GPU-(D)BE achieves better speed-up for smaller p1 (Fig. 4 (a)). The result
is explained by observing that small p1 values correspond to smaller induced
width of the underlying constraint graph. In turn, for small p1 values, GPU-
(D)BE agents construct smaller UTIL tables, which increases the probabil-
ity of performing the complete inference process on the GPU, through the
GPU-Aggregate-Project procedure. This observation is also consistent
with what observed in the previous experiments in terms of scalability.
• GPU-(D)BE achieves greater speedups in presence of large domains. This is
due to the fact that large domains correspond to large UTIL tables, enabling
the GPU-based algorithms to exploit a greater amount of parallelism, provided
that the UTIL tables can be stored in the global memory of the GPU.

6 Conclusions and Discussions

In this paper, we presented an investigation of the use of GPUs to exploit
SIMT-style parallelism from DP-based methods to solve COPs and DCOPs.
Exploiting GPUs in Solving (Distributed) Constraint Optimization Problems 137

We proposed a procedure, inspired by BE (for COPs) and DPOP (for DCOPs),

that makes use of multiple threads to parallelize the aggregation and projection
phases. Experimental results show that the use of GPUs may provide signifi-
cant advantages in terms of runtime and scalability. The proposed results are
significant—the wide availability of GPUs provides access to parallel computing
solutions that can be used to improve efficiency of (D)COP solvers. Furthermore,
GPUs are renowned for their complex architectures (multiple memory levels with
very different size and speed characteristics; relatively slow cores), which often
create challenges to the effective exploitation of parallelism from irregular appli-
cations; the strong experimental results indicate that the proposed algorithms
are well-suited to GPU architectures. While envisioning further research in this
area, we anticipate several challenges:
• In terms of implementation, GPU programming can be more demanding when
compared to a classical sequential implementation. One of the current limita-
tions for (D)COP-based GPU approaches is the absence of solid abstractions
that allow component integration, modularly, without restructuring the whole
program.
• Exploiting the integration of CPU and GPU computations is a key factor to
obtain competitive solvers performance. Complex and repeated calculations
should be delegated to GPUs, while simpler and memory intensive operations
should be assigned to CPUs. It is however unclear how to determine good
tradeoffs of such integrations. For instance, repeatedly invoking many mem-
ory demanding GPU kernels could be detrimental to the overall performance,
due to the high cost of allocating the device memory (e.g., shared memory).
Creating lightweight communication mechanisms between CPU and GPU (for
instance, by taking advantage of the asynchronism of CUDA streams) to allow
active GPU kernels to be used in multiple instances could be a possible solu-
tion to investigate.
• While this paper describes the applicability of our approach to BE and DPOP,
we believe that analogous techniques can be derived and applied to other
DP-based approaches to solve (D)COPs—e.g., to implement the logic of DP-
based propagators. We also envision that such technology could open the door
to efficiently enforcing higher form of consistencies than domain consistency
(e.g., path consistency [22], adaptive consistency [12], or the more recently
proposed branch consistency for DCOPs [14]), especially when the constraints
need to be represented explicitly.

References
1. Abdennadher, S., Schlenker, H.: Nurse scheduling using constraint logic program-
ming. In: Proceedings of the Conference on Innovative Applications of Artificial
Intelligence (IAAI), pp. 838–843 (1999)
2. Apt, K.: Principles of constraint programming. Cambridge University Press (2003)
3. Arbelaez, A., Codognet, P.: A GPU implementation of parallel constraint-based
local search. In: Proceedings of the Euromicro International Conference on Parallel,
Distributed and network-based Processing (PDP), pp. 648–655 (2014)
138 F. Fioretto et al.

4. Barabási, A.-L., Albert, R.: Emergence of scaling in random networks. Science

286(5439), 509–512 (1999)
5. Boyer, V., El Baz, D., Elkihel, M.: Solving knapsack problems on GPU. Computers
& Operations Research 39(1), 42–47 (2012)
6. Brito, I., Meseguer, P.: Improving DPOP with function filtering. In: Proceedings
of the International Conference on Autonomous Agents and Multiagent Systems
(AAMAS), pp. 141–158 (2010)
7. Burke, D., Brown, K.: Efficiently handling complex local problems in distributed
constraint optimisation. In: Proceedings of the European Conference on Artificial
Intelligence (ECAI), pp. 701–702 (2006)
8. Burke, E.K., De Causmaecker, P., Berghe, G.V., Van Landeghem, H.: The state of
the art of nurse rostering. Journal of scheduling 7(6), 441–499 (2004)
9. Campeotto, F., Dovier, A., Fioretto, F., Pontelli, E.: A GPU implementation of
large neighborhood search for solving constraint optimization problems. In: Pro-
ceedings of the European Conference on Artificial Intelligence (ECAI), pp. 189–194
(2014)
10. Dechter, R.: Bucket elimination: a unifying framework for probabilistic inference.
In: Learning in graphical models, pp. 75–104. Springer (1998)
11. Dechter, R.: Constraint Processing. Morgan Kaufmann Publishers Inc., San
Francisco (2003)
12. Dechter, R., Pearl, J.: Network-based heuristics for constraint-satisfaction prob-
lems. Springer (1988)
13. Farinelli, A., Rogers, A., Petcu, A., Jennings, N.: Decentralised coordination of
low-power embedded devices using the Max-Sum algorithm. In: Proceedings of
the International Conference on Autonomous Agents and Multiagent Systems
(AAMAS), pp. 639–646 (2008)
14. Fioretto, F., Le, T., Yeoh, W., Pontelli, E., Son, T.C.: Improving DPOP with
branch consistency for solving distributed constraint optimization problems. In:
O’Sullivan, B. (ed.) CP 2014. LNCS, vol. 8656, pp. 307–323. Springer, Heidelberg
(2014)
15. Gaudreault, J., Frayret, J.-M., Pesant, G.: Distributed search for supply chain
coordination. Computers in Industry 60(6), 441–451 (2009)
16. Hamadi, Y., Bessière, C., Quinqueton, J.: Distributed intelligent backtracking.
In: Proceedings of the European Conference on Artificial Intelligence (ECAI),
pp. 219–223 (1998)
17. Kumar, A., Faltings, B., Petcu, A.: Distributed constraint optimization with struc-
tured resource constraints. In: Proceedings of the International Conference on
Autonomous Agents and Multiagent Systems (AAMAS), pp. 923–930 (2009)
18. Lalami, M.E., El Baz, D., Boyer, V.: Multi GPU implementation of the simplex
algorithm. Proceedings of the International Conference on High Performance Com-
puting and Communication (HPCC) 11, 179–186 (2011)
19. Léauté, T., Faltings, B.: Distributed constraint optimization under stochastic
uncertainty. In: Proceedings of the AAAI Conference on Artificial Intelligence
(AAAI), pp. 68–73 (2011)
20. Maheswaran, R., Tambe, M., Bowring, E., Pearce, J., Varakantham, P.: Taking
DCOP to the real world: Efficient complete solutions for distributed event schedul-
ing. In: Proceedings of the International Conference on Autonomous Agents and
Multiagent Systems (AAMAS), pp. 310–317 (2004)
21. Modi, P., Shen, W.-M., Tambe, M., Yokoo, M.: ADOPT: Asynchronous distributed
constraint optimization with quality guarantees. Artificial Intelligence 161(1–2),
149–180 (2005)
Exploiting GPUs in Solving (Distributed) Constraint Optimization Problems 139

22. Montanari, U.: Networks of constraints: Fundamental properties and applications

to picture processing. Information sciences 7, 95–132 (1974)
23. Pawlowski, K., Kurach, K., Michalak, T., Rahwan, T.: Coalition structure genera-
tion with the graphic processor unit. Technical Report CS-RR-13-07, Department
of Computer Science, University of Oxford (2014)
24. Pesant, G.: A regular language membership constraint for finite sequences of vari-
ables. In: Proceedings of the International Conference on Principles and Practice
of Constraint Programming (CP), pp. 482–495 (2004)
25. Petcu, A., Faltings, B.: A scalable method for multiagent constraint optimiza-
tion. In: Proceedings of the International Joint Conference on Artificial Intelligence
(IJCAI), pp. 1413–1420 (2005)
26. Quimper, C.-G., Walsh, T.: Global grammar constraints. In: Benhamou, F. (ed.)
CP 2006. LNCS, vol. 4204, pp. 751–755. Springer, Heidelberg (2006)
27. Rodrigues, L.C.A., Magatão, L.: Enhancing supply chain decisions using constraint
programming: a case study. In: Gelbukh, A., Kuri Morales, Á.F. (eds.) MICAI 2007.
LNCS (LNAI), vol. 4827, pp. 1110–1121. Springer, Heidelberg (2007)
28. Rossi, F., van Beek, P., Walsh, T. (eds.) Handbook of Constraint Programming.
Elsevier (2006)
29. Sanders, J., Kandrot, E.: CUDA by Example. An Introduction to General-Purpose
GPU Programming. Addison Wesley (2010)
30. Sultanik, E., Modi, P.J., Regli, W.C.: On modeling multiagent task scheduling as
a distributed constraint optimization problem. In: Proceedings of the International
Joint Conference on Artificial Intelligence (IJCAI), pp. 1531–1536 (2007)
31. Trick, M.A.: A dynamic programming approach for consistency and propagation
for knapsack constraints. Annals of Operations Research 118(1–4), 73–84 (2003)
32. Yeoh, W., Felner, A., Koenig, S.: BnB-ADOPT: An asynchronous branch-and-
bound DCOP algorithm. Journal of Artificial Intelligence Research 38, 85–133
(2010)
33. Yeoh, W., Yokoo, M.: Distributed problem solving. AI Magazine 33(3), 53–65
(2012)
34. Yokoo, M. (ed.): Distributed Constraint Satisfaction: Foundation of Cooperation
in Multi-agent Systems. Springer (2001)
35. Zivan, R., Okamoto, S., Peled, H.: Explorative anytime local search for distributed
constraint optimization. Artificial Intelligence 212, 1–26 (2014)
36. Zivan, R., Yedidsion, H., Okamoto, S., Glinton, R., Sycara, K.: Distributed con-
straint optimization for teams of mobile sensing agents. Journal of Autonomous
Agents and Multi-Agent Systems 29(3), 495–536 (2015)
 Conflict Ordering Search
for Scheduling Problems

Steven Gay1(B) , Renaud Hartert1 , Christophe Lecoutre2 , and Pierre Schaus1

1
UCLouvain, ICTEAM, Place Sainte Barbe 2, 1348 Louvain-la-Neuve, Belgium
{steven.gay,renaud.hartert,pierre.schaus}@uclouvain.be
2
CRIL-CNRS UMR 8188, Université d’Artois, 62307 Lens, France
lecoutre@cril.fr

Abstract. We introduce a new generic scheme to guide backtrack

search, called Conflict Ordering Search (COS), that reorders variables
on the basis of conflicts that happen during search. Similarly to gen-
eralized Last Conflict (LC), our approach remembers the last variables
on which search decisions failed. Importantly, the initial ordering behind
COS is given by a specified variable ordering heuristic, but contrary to
LC, once consumed, this first ordering is forgotten, which makes COS
conflict-driven. Our preliminary experiments show that COS – although
simple to implement and parameter-free – is competitive with special-
ized searches on scheduling problems. We also show that our approach
fits well within a restart framework, and can be enhanced with a value
ordering heuristic that selects in priority the last assigned values.

1 Introduction
Backtracking search is a central complete algorithm used to solve combinato-
rial constrained problems. Unfortunately, it suffers from thrashing – repeat-
edly exploring the same fruitless subtrees – during search. Restarts, adaptive
heuristics, and strong consistency algorithms are typical Constraint Program-
ming (CP) techniques used to cope with thrashing.
Last Conflicts (LC) [9] has been shown to be highly profitable to complete
search algorithms, both in constraint satisfaction and in automated artificial
intelligence planning. The principle behind LC is to select in priority the last
conflicting variables as long as they cannot be instantiated without leading to
a failure. Interestingly enough, last conflict search can be combined with any
underlying variable ordering heuristic. In normal mode, the underlying heuristic
selects the variables to branch on, whereas in conflict mode, variables are directly
selected in a conflict set built by last conflict.
While last conflict uses conflicts to repair the search heuristic, we show in this
paper that conflicts can also be used to drive the search process by progressively
replacing the initial variable heuristic. Basically, the idea behind our approach –
namely, Conflict Ordering Search – is to reorder variables according to the most
recent conflict they were involved in. Our experiments highlight that this simple
reordering scheme, while being generic, can outperform domain specific heuristics
for scheduling problems.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 140–148, 2015.
DOI: 10.1007/978-3-319-23219-5 10
 Conflict Ordering Search for Scheduling Problems 141

2 Related Works
We start by providing a quick overview of general-purpose search heuristics and
schemes since our approach is definitively one of them. The simple variable
ordering heuristic dom [5] – which selects variables by their domain size – has
long been considered as the most robust backtrack search heuristic. However,
a decade ago, modern adaptive heuristics were introduced. Such heuristics take
into account information related to the part of the search space already explored.
The two first proposed generic adaptive heuristics are impact [16] and wdeg [1].
The former relies on a measure of the effect of any assignment, and the latter
associates a counter with each constraint (and indirectly, with each variable)
indicating how many times any constraint led to a domain wipe-out. Counting-
based heuristics [14] and activity-based search [11] are two recently introduced
additional adaptive techniques to guide the search process.
Interestingly, Last Conflict (LC) [9] is a search mechanism that can be applied
on top of any variable ordering heuristic. Precisely, the generalized form LC(k)
works by recording and assigning first the k variables involved in the k last deci-
sions that provoked a failure after propagation. The underlying search heuristic
is used when all the last k conflicting variables have been assigned. While the
ability of relying on an underlying search heuristic is a strong point of LC, setting
the parameter k can be problematic, as we shall see later. The related scheme
introduced in this paper goes further and orders permanently all conflicting vari-
ables using the time of their last conflicts, eventually becoming independent of
the helper variable ordering heuristic.

3 Guiding Search by Timestamping Conflicts

We first introduce basic concepts. Then, we introduce Conflict Ordering Search

and highlight its benefits within a context of restarts. We conclude this section
by discussing the differences between LC(k) and COS.

3.1 Background

CSP. A Constraint Satisfaction Problem (CSP) P is a pair (X , C), where X is

a finite set of variables and C is a finite set of constraints. Each variable x ∈ X
has a domain dom(x) that contains the allowed values for x. A valuation on a
subset X ⊆ X of variables maps each variable x ∈ X with a value in dom(x).
Each constraint c ∈ C has a scope scp(c) ⊆ X , and is semantically defined by a
set of allowed valuations on scp(c); the valuations that satisfy c. A valuation on
X is a solution of P iff it satisfies each constraint of P . A CSP is satisfiable iff
it admits at least one solution.

Tree-Search. One can solve CSPs by using backtrack search, a complete depth-
first exploration of the search space, with backtracking when a dead-end occurs.
At each search node, a filtering process φ can be performed on domains by
142 S. Gay et al.

soliciting propagators associated with constraints. A CSP is in failure, denoted ⊥,

when unsatisfiability is detected by φ. A branching heuristic is a function that
maps a non-failed CSP to an ordered sequence of constraints, called decisions.
In this work, we only consider binary variable-based branching heuristics, i.e.,
heuristics that always generate sequences of decisions of the form x ∈ D, x ∈/ D,
where x is a variable of X and D a strict subset of dom(x). A search tree
is the structure explored by backtrack search through its filtering capability
and its branching heuristic. A failed node in the search tree is a node where
unsatisfiability has been detected by φ.

3.2 Conflict Ordering

Considering a variable-based branching heuristic, we can associate a failed search
node with the variable involved in the decision leading to it. This allows us to
timestamp variables with the number of the last conflict they caused (see Fig. 1).
The basic idea behind Conflict Ordering Search is to leverage this timestamping
mechanism to reorder the variables during search.

Fig. 1. Conflict numbering and timestamps associated with each variable. Variables
are stamped with the number of their latest conflict (or 0 by default).

Algorithm 1 describes Conflict Ordering Search. For simplicity, we only con-

sider classical binary branching with decisions of the form x ≤ v and x > v.
We use an integer nConflicts to count the number of conflicts and a reference
lastVar to the last variable involved in a decision (initially null at the root of the
search-tree). We also consider a one-dimensional array stamps that associates
with each variable x ∈ X the last time variable x was involved in a conflict.
They are all initialized to 0. We suppose that φ corresponds to a domain filter-
ing consistency, which is at least as strong as the partial form of arc consistency
ensured by the forward checking algorithm [5].
If the resulting CSP at line 1 is trivially inconsistent (⊥), false is returned
(line 6). If the failure is due to a previous decision (line 3), the number of conflicts
is incremented and the conflicting variable timestamped with this number (lines 4
and 5). Otherwise, COS returns true if a solution has been found, i.e., the domain
of each variable in X is a singleton (lines 7 and 8). The selection of the next
variable to branch on is performed between lines 9 and 13. Here, the timestamps
are used to select the unbound variable involved in the latest conflict. If no
unbound variable ever conflicted, the search falls back to the bootstrapping
 Conflict Ordering Search for Scheduling Problems 143

heuristic varHeuristic. When a new value has been selected by the heuristic
valHeuristic[x], we recursively call COS . One can observe that the complexity
of selecting a variable is linear in time and space, hence scaling well.1

Algorithm 1. COS (P = (X , C): CSP)

Output: true iff P is satisfiable
1 P ← φ(P )
2 if P = ⊥ then
3 if lastVar = null then
4 nConflicts ← nConflicts + 1
5 stamps[lastVar] ← nConflicts
6 return false
7 if ∀x ∈ X , |dom(x)| = 1 then
8 return true
9 failed ← {x ∈ X : stamps[x] > 0 ∧ |dom(x)| > 1}
10 if failed = ∅ then
11 lastVar ← varHeuristic.select()
12 else
13 lastVar ← argmax x∈failed {stamps[x]}
14 v ← valHeuristic[lastVar].select()
15 return COS (P|lastVar≤v ) ∨ COS (P|lastVar>v )

Example 1. Let us consider a toy CSP with n “white” variables, and m “black”
variables. White variables have a binary domain while black variables have
{1, 2, . . . , m − 1} as domain. We also add a binary difference constraint on each
pair of black variables (thus making the CSP unsatisfiable), but no constraint
at all on the white variables. Let us also assume a variable ordering heuristic
that selects the white variables first, then the black variables. Hence, proving
unsatisfiability using this heuristic requires to prove the “black conflict” for the
2n valuations of the white variables (see left part of Fig. 2). Using COS on top
of this heuristic allows one to detect unsatisfiability quickly. Indeed, the m − 2
first conflicting black variables will be prioritized as a white variable cannot be
involved in a conflict. The number of times the “black conflict” as to be proven
thus becomes linear in the number of white variables n (see right part of Fig. 2).

COSPhase: a Variant. Because it is known that remembering last assigned val-

ues for later priority uses can be worthwhile (see for example phase saving [15]
in SAT), we propose such a variant for COS. So, when a positive decision x ≤ v
succeeds, we record its value v. Then, when branching is performed on a times-
tamped variable x, we exploit the associated recorded value v. If v is still in
1
The time complexity could be improved if an ordered linked-list is used instead of
the array stamps.
144 S. Gay et al.

Without Conflict Ordering With Conflict Ordering

. . .

Fig. 2. Conflict Ordering Search reorders the variables to reduce the number of times
the inconsistent black subtree has to be explored.

the domain of x, we use interval splitting on it, i.e., we branch with decisions
x ≤ v and x > v, otherwise the value heuristic is solicited. Observe that this
mechanism follows the first-fail/best-first principle.

3.3 Restarts and Timestamp Ordering

Depth-first search is far from always being the most efficient way to solve a CSP.
In some cases, it may suffer from heavy-tailed distributions [4]. While restarting
the search with a randomized heuristic is a typical way to avoid the worst parts
of a long-tailed curve, nogood learning mitigates the effect of cutting a search
process short by remembering parts of the search space already explored [8].2
In the context of COS, we observe that our approach not only remembers the
sources of conflicts but also produce a variable ordering that yields a behavior
similar to randomization. We thus propose to use no additional randomization
when restarting, only using the natural randomizing effect of conflict ordering
instead. The rationale is that while conflict ordering is good at finding a set of
conflicting variables in a given context – i.e., a sequence of previous decisions –
restarting with conflict ordering has the effect of trying the latest conflict set in
other contexts.

Example 2. Let us consider the toy CSP described in Example 1. The time
required to prove unfeasibility could be drastically reduced if a restart occurs
after having explored the inconsistent black subtree at least once. Indeed, in this
context, the second restart will directly explore the black search tree without
even considering the white variables that have been “disculpated”.

3.4 Differences with Last Conflict Search

Although similar, we show that COS and LC are rather different. Indeed, LC
relies on a parameter k that corresponds to the maximum size of the conflict
sets that can be captured by the search process. The value of k is of importance
as setting k too low may not allow LC to capture the encountered conflict sets.
For instance, LC(k) cannot capture the conflict set in Example 1 if k is lower
than m − 2. COS, however, does not require any parameter and is able to handle
2
Similar frameworks are typically used by SAT solvers [12].
 Conflict Ordering Search for Scheduling Problems 145

conflict sets of any size. While setting the parameter k to the number of decision
variables may solve the problem, LC still suffers from resets of its conflict set
that occur each time the conflicting variables have been successfully assigned.
Conversely, COS does not forget conflicting variables and progressively reorders
those variables to give priority to the recently conflicting ones. This is particu-
larly important with restarts as LC is not designed to focus on conflict sets in
such contexts (see Example 2).

4 Experiments

We have tested our approach on RCPSP (Resource-Constrained Project Schedul-

ing Project) instances from PSPLIB [6]. We used a computer equipped with
a i7-3615QM processor running at 2.30GHz. The problem has been modeled
in the open-source solver OscaR [13], using precedence and cumulative con-
straints. Precedences are simple binary precedence constraints, and cumulative
constraints use the Time-Tabling propagator presented in [2]. Both overload
checking [19] or time-table edge-finding [17] were tested but energy-based rea-
soning does not help much on PSPLIB instances, whose optimal solutions typ-
ically waste capacity. Adding TTDR [3] helps even with learning searches, but
it makes the experiments harder to reproduce, thus we chose to not use it.

4.1 Branch-and-Bound

The goal of this first experiment is to find an optimal solution using a pure
branch-and-bound search. We compare five search solving methods. The first
is a simple min/min scheme, which selects the variable with the smallest min-
imal value and chooses the smallest value (for assignment). The second one is
the scheduling-specialized SetTimes heuristic with dominances [7], which is a
min/min scheme that simply postpones assignments (instead of making value
refutations) when branching at right, fails when a postponed task can no longer
be woken up, and assigns the tasks that are not postponed and cannot be
disturbed by other tasks. Finally, the last three heuristics correspond to conflict-
based reasoning searches, namely, LC(k) for the best value of k, our main con-
tribution COS, and COSPhase based on the variant presented in Section 15. All
these have a min/min helper heuristic.
We have observed that the ranking of these five search methods is the same
on the four RCPSP benchmarks (J30, J60, J90, J120) from PSPLIB. Results
are represented on the left part of Fig. 3 where the y-axis is the cumulated
number of solved instances and the x-axis is the CPU time. SetTimes is clearly
an improvement on min/min, as it finishes ahead and seems to continue its
course on a better slope than min/min. In turn, the well-parameterized LC
(we empirically chose the best possible value for k) fares better than SetTimes.
Finally, COS allows us to close a higher number of instances, and the variant
COSPhase improves COS even further.
146 S. Gay et al.

4.2 Branch-and-Bound with Restarts

As illustrated in Example 2, keeping the conflict ordering between restarts could
drastically reduce search efforts. The aim of this second experiment is to compare
the performance of COS if the ordering is kept between restarts or not. Experi-
mental settings are the same as before except that we use restarts and nogood
recording has explained in [10]. The first iteration is limited to 100 failures and
increases by a 1.15 factor. We compared the performance of COS and COSPhase
with and without reset (we add “rst-” as prefix for the resetting version). All
searches rely on minrnd /min – a randomized version of min/min that breaks ties
randomly – as helper heuristic.
Results are presented in the right part of Fig. 3. First, we observe that restarts
do have a positive effect as minrnd /min obtains better results than min/min in
the previous experiment. Next, we see that resetting the conflict order has a bad
effect on the search process. Indeed, these variant obtain worse results than in
the pure branch-and-bound framework. This highlights that using conflict-based
search as a full heuristic can yield much better results than using it as a repairing
patch. Finally, phase recording does not seem to help anymore.

COSPhase
235 COSPhase 237
COS
228 COS
211 rst-COSPhase
206 rst-COS
185 LC(k)
175 SetTimes 172 minrnd/min
155 min/min

0 0
0 10 ms 1 sec 1 min 0 10 ms 1 sec 1 min

Search #closed Σ Makespans Search #closed Σ Makespans

COSPhase 235 77789 COS 237 77770
COS 228 77952 COSPhase 236 77721
LC(k) 185 79010 rst-COSPhase 211 78301
SetTimes 175 80970 rst-COS 206 78476
min/min 155 80038 minrnd /min 172 79510

Fig. 3. On the left, results obtained for pure branch-and-bound, and on the right,
results obtained with branch-and-bound with restarts. Graphs at the top show the
cumulated number of solved RCPSP instances from PSPLIB120. The tables at the
bottom compare branching heuristics at the end of the 60s timeout, giving the number
of closed instances and the sum of makespans.

4.3 Destructive Lower Bounds.

We performed similar experiments for destructive lower bounds. We added the
Time-Table Edge-Finding propagator presented in [17] since it has a large impact
 Conflict Ordering Search for Scheduling Problems 147

in this case. The results are similar. We also compared COS to the recently intro-
duced Failure Directed Search [18] by implementing it directly in CP Optimizer.
Unfortunately our COS implementation in CP Optimizer was not able to obtain
results competitive with FDS.

5 Conclusion

In this paper, we have proposed a general-purpose search scheme that can be

combined with any variable ordering heuristic. Contrary to Last Conflict, Con-
flict Ordering Search is very aggressive, discarding progressively the role played
by the heuristic. Besides, by means of simple timestamps, all variables recorded
in the global conflict set stay permanently ordered, the priority being modified
at each new conflict. We have shown that on some structured known problems
our approach outperforms other generic and specific solving methods. So, COS
should be considered as one new useful technique to be integrated in the outfit
of constraint systems.

Acknowledgments. Steven Gay is financed by project Innoviris 13-R-50 of the

Brussels-Capital region. Renaud Hartert is a Research Fellow of the Fonds de la
Recherche Scientifiques - FNRS. Christophe Lecoutre benefits from the financial sup-
port of CNRS and OSEO (BPI France) within the ISI project ?Pajero?. The authors
would like to thank Petr Vilı́m for his help with the comparison to FDS in CP
Optimizer.

References
1. Boussemart, F., Hemery, F., Lecoutre, C., Sais, L.: Boosting systematic search by
weighting constraints. In: Proceedings of ECAI 2004, pp. 146–150 (2004)
2. Gay, S., Hartert, R., Schaus, P.: Simple and scalable time-table filtering for
the cumulative constraint. In: Pesant, G. (ed.) Proceedings of CP 2015. LNCS,
vol. 9255, pp. 149–157. Springer, Heidelberg (2015)
3. Gay, S., Hartert, R., Schaus, P.: Time-table disjunctive reasoning for the cumula-
tive constraint. In: Michel, L. (ed.) CPAIOR 2015. LNCS, vol. 9075, pp. 157–172.
Springer, Heidelberg (2015)
4. Gomes, C., Selman, B., Crato, N., Kautz, H.: Heavy-tailed phenomena in satisfia-
bility and constraint satisfaction problems. Journal of Automated Reasoning 24,
67–100 (2000)
5. Haralick, R.M., Elliott, G.L.: Increasing tree search efficiency for constraint satis-
faction problems. Artificial Intelligence 14, 263–313 (1980)
6. Kolisch, R., Schwindt, C., Sprecher, A.: Benchmark instances for project scheduling
problems. In: Project Scheduling, pp. 197–212. Springer (1999)
7. Le Pape, C., Couronné, P., Vergamini, D., Gosselin, V.: Time-versus-capacity com-
promises in project scheduling (1994)
8. Lecoutre, C., Sais, L., Tabary, S., Vidal, V.: Nogood recording from restarts. In:
Proceedings of IJCAI 2007, pp. 131–136 (2007)
148 S. Gay et al.

9. Lecoutre, C., Sais, L., Tabary, S., Vidal, V.: Reasonning from last conflict(s) in
constraint programming. Artificial Intelligence 173(18), 1592–1614 (2009)
10. Lecoutre, C., Sais, L., Tabary, S., Vidal, V.: Recording and minimizing nogoods
from restarts. Journal on Satisfiability, Boolean Modeling and Computation 1,
147–167 (2007)
11. Michel, L., Van Hentenryck, P.: Activity-based search for black-box constraint
programming solvers. In: Beldiceanu, N., Jussien, N., Pinson, É. (eds.) CPAIOR
2012. LNCS, vol. 7298, pp. 228–243. Springer, Heidelberg (2012)
12. Moskewicz, M.W., Madigan, C.F., Zhao, Y., Zhang, L., Malik, S. Chaff: engineering
an efficient SAT solver. In: Proceedings of DAC 2001, pp. 530–535 (2001)
13. OscaR Team. OscaR: Scala in OR (2012). bitbucket.org/oscarlib/oscar
14. Pesant, G., Quimper, C.-G., Zanarini, A.: Counting-based search: Branching
heuristics for constraint satisfaction problems. Journal of Artificial Intelligence
Research 43, 173–210 (2012)
15. Pipatsrisawat, K., Darwiche, A.: A lightweight component caching scheme for sat-
isfiability solvers. In: Marques-Silva, J., Sakallah, K.A. (eds.) SAT 2007. LNCS,
vol. 4501, pp. 294–299. Springer, Heidelberg (2007)
16. Refalo, P.: Impact-based search strategies for constraint programming. In:
Wallace, M. (ed.) CP 2004. LNCS, vol. 3258, pp. 557–571. Springer, Heidelberg
(2004)
17. Vilı́m, P.: Timetable edge finding filtering algorithm for discrete cumulative
resources. In: Achterberg, T., Beck, J.C. (eds.) CPAIOR 2011. LNCS, vol. 6697,
pp. 230–245. Springer, Heidelberg (2011)
18. Vilı́m, P., Laborie, P., Shaw, P.: Failure-directed search for constraint-based
scheduling. In: Michel, L. (ed.) CPAIOR 2015. LNCS, vol. 9075, pp. 437–453.
Springer, Heidelberg (2015)
19. Wolf, A., Schrader, G.: O(n log n) overload checking for the cumulative constraint
and its application. In: Umeda, M., Wolf, A., Bartenstein, O., Geske, U., Seipel, D.,
Takata, O. (eds.) INAP 2005. LNCS (LNAI), vol. 4369, pp. 88–101. Springer,
Heidelberg (2006)
 Simple and Scalable Time-Table Filtering
for the Cumulative Constraint

Steven Gay(B) , Renaud Hartert, and Pierre Schaus

ICTEAM, UCLouvain, Place Sainte Barbe 2, 1348 Louvain-la-neuve, Belgium

{steven.gay,renaud.hartert,pierre.schaus}@uclouvain.be

Abstract. Cumulative is an essential constraint in the CP framework,

and is present in scheduling and packing applications. The lightest filter-
ing for the cumulative constraint is time-tabling. It has been improved
several times over the last decade. The best known theoretical time com-
plexity for time-table is O(n log n)introduced by Ouellet and Quimper.
We show a new algorithm able to run in O(n), by relying on range min
query algorithms. This approach is more of theoretical rather than prac-
tical interest, because of the generally larger number of iterations needed
to reach the fixed point. On the practical side, the recent synchronized
sweep algorithm of Letort et al, with a time-complexity of O(n2 ), requires
fewer iterations to reach the fix-point and is considered as the most scal-
able approach. Unfortunately this algorithm is not trivial to implement.
In this work we present a O(n2 ) simple two step alternative approach:
first building the mandatory profile, then updating all the bounds of
the activities. Our experimental results show that our algorithm out-
performs synchronized sweep and the time-tabling implementations of
other open-source solvers on large scale scheduling instances, sometimes
significantly.

Keywords: Constraint programming · Large-scale · Scheduling ·

Cumulative constraint · Time-table

1 Preliminaries
In this paper, we focus on a single cumulative resource with a discrete finite
capacity C ∈ N and a set of n tasks Ω = {1, . . . , n}. Each task i has a start time
si ∈ Z, a fixed duration di ∈ N, and an end time ei ∈ Z such that the equality
si + di = ei holds. Moreover, each task i consumes a fixed amount of resource
ci ∈ N during its processing time. Tasks are non-preemptive, i.e., they cannot be
interrupted during their processing time. In the following, we denote by si and
si the earliest and the latest start time of task i and by ei and ei the earliest and
latest end time of task i (see Fig. 1). The cumulative constraint [1] ensures that
the accumulated resource consumptiondoes not exceed the maximum capacity
C at any time t (see Fig. 2): ∀t ∈ Z : i∈Ω : si ≤t<ei ci ≤ C.
Even tasks that are not fixed convey some information that can be used by
filtering rules. For instance, tasks with a tight execution window must consume
some resource during a specific time interval known as mandatory part.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 149–157, 2015.
DOI: 10.1007/978-3-319-23219-5 11
150 S. Gay et al.

Fig. 1. Task i is characterized by its start Fig. 2. Accumulated resource consump-

time si , its duration di , its end time ei , tion over time. The cumulative constraint
and its resource consumption ci . ensures that the maximum capacity C is
not exceeded.

Definition 1 (Mandatory part). Let us consider a task i ∈ Ω. The manda-

tory part of i is the time interval [si , ei [. Task i has a mandatory part only if its
latest start time is smaller than its earliest end time.
If task i has a mandatory part, we know that task i will consume ci units of
resource during all its mandatory part no matter its start time. Fig. 3 illustrates
the mandatory part of an arbitrary task i.

Fig. 3. Task i has a mandatory part [si , ei [ if its latest start time si is smaller than its
earliest end time ei : si < ei . Task i always consumes the resource during its mandatory
part no matter its start time.

By aggregation, mandatory parts allow us to have an optimistic view of the

resource consumption over time. This aggregation is known as the time-table.
Definition 2 (Time-Table). The time-table TTΩ is the aggregation of the
mandatory parts of all the tasks in Ω. It is defined as the following step function:

TTΩ = t ∈ Z −→ ci . (1)
i∈Ω | si ≤t<ei

The capacity of the resource is exceeded if ∃t ∈ Z : TTΩ (t) > C.

The time-table of a resource can be computed in O(n) by a sweep algorithm
given the tasks sorted by latest start time and earliest end time [2,11,15].
The time-table filtering rule is formalized as follows:

(t < ei ) ∧ (ci + TTΩ\i (t) > C) ⇒ t < si . (2)

 Simple and Scalable Time-Table Filtering for the Cumulative Constraint 151

Observe that this filtering rule only describes how to update the start time of a
task. End times are updated in a symmetrical way.
Let j be a rectangle denoted aj , bj , hj
with aj ∈ Z (resp. bj ∈ Z) its start
(resp. end) time, hj ∈ N its height, and bj − aj its duration (length). The time-
table TTΩ can be represented as a contiguous sequence of rectangles

T TΩ = −∞, a1 , 0
, a1 , b1 , h1
, . . . , am , bm , hm
, bm , ∞, 0
(3)

such that bi = ai+1 and that the following holds:

∀ aj , bj , hj
∈ T TΩ , ∀t ∈ [aj , bj [ : T TΩ (t) = hj . (4)

We assume that the sequence is minimal, i.e., no consecutive rectangles have the
same height. The maximum number of rectangles is thus limited to 2n + 1.

Definition 3 (Conflicting Rectangle). For a task i, a left-conflicting rectan-

gle is a rectangle aj , bj , hj
∈ TTΩ\i such that (aj < e)∧(bj ≥ si )∧(hj > C−ci ).
We say that the task is in left-conflict with rectangle j. Right-conflicting rectan-
gles are defined symmetrically.

The time-table filtering rule can thus be rephrased as follows:

∀i ∈ Ω, ∀ aj , bj , hj
∈ T TΩ\i : j is in left-conflict with i ⇒ bj ≤ si . (5)

Definition 4 (Time-Table Consistency). A cumulative constraint is left

(resp. right) time-table consistent if no task has a left (resp. right) conflicting
rectangle. It is time-table consistent if it is left and right time-table consistent.

2 Existing Algorithms for Time-Tabling

Using the notion of conflicting rectangles, one can design a naive time-tabling
algorithm by confronting every task to every rectangle of the profile. The fol-
lowing algorithms improve on this, mainly by avoiding fruitless confrontations
of rectangles and tasks.

Sweep-line Algorithm. The sweep-line algorithm introduced by Beldiceanu et al.

[2] introduces tasks from left to right, and builds the mandatory profile on-the-
fly. This allows to confront tasks and rectangles only if they can overlap in time.
It can factorize confrontations of a rectangle against several tasks, by organizing
tasks in a heap. It pushes tasks to the right until they have no left-conflicting
rectangle, as pictured in Figure 4(c). This algorithm runs in O(n2 ).

Idempotent Sweep-line Algorithm. The sweep-line algorithm by Letort et. al [11]

improves on building the profile on-the-fly, by taking in consideration mandatory
parts that appear dynamically as tasks are pushed. It reaches left-time-table
consistency in O(n2 ), or O(n2 log n) for its faster practical implementation.
152 S. Gay et al.

Interval Tree Algorithm. The algorithm of Ouellet and Quimper [14] first builds
the profile, then introduces rectangles and tasks in an interval tree. Rectangles
are introduced by decreasing height, tasks by increasing height. This allows tasks
and rectangles to be confronted only when their heights do conflict. For each
task introduction, the tree structure decides in log n if its time domain conflicts
with some rectangle. Its filtering is weaker, since it pushes a task i only after
left-conflicting rectangles that overlap [si , ei [, as pictured in Figure 4(b). The
algorithm has time complexity O(n log n).

Fig. 4. Filtering obtained for (a) our linear time-tabling (b) Ouellet et al [14] and (c)
Beldiceanu et al [2].

New Algorithms. In this paper, we introduce two new algorithms for time-
tabling. The first one is of theoretical interest and runs in O(n). It uses range-
max-query algorithms to determine whether a task has a conflicting rectangle.
As the algorithm of Ouellet et al [14], it confronts task i only with rectangles
overlapping [si , ei [, but only chooses the one with the largest height instead of
the largest end. Thus, it prunes even less, as depicted in Figure 4(a).
The second algorithm is practical, and runs in O(n2 ). It separates profile
building from task sweeping. To locate tasks on the profile, it exploits residues
from previous computations, and incrementally removes fixed tasks that cannot
lead to any more pruning. It uses sweeping, thus pruning as much as [2] per call,
but it updates both the earliest start and latest end times of the tasks in a single
execution.

3 A Linear Time-Table Filtering Algorithm

In order to obtain linear time complexity, we will confront task i to only one
well-chosen rectangle, for every task i.
Proposition 1. Suppose the mandatory profile does not overload the resource. Let
i be a task, and j ∗ be a highest rectangle of the profile overlapping [si , min(ei , si )
[: j ∗ = argmaxj {hj | aj , bj , hj
∈ TTΩ and [aj , bj [∩[si , min(ei , si )[= ∅} .
Then j ∗ is in left-conflict iff hj ∗ + ci > C; otherwise i has no rectangles in
left-conflict.
 Simple and Scalable Time-Table Filtering for the Cumulative Constraint 153

Proof. If i has a mandatory part, we only need to look for conflict rectangles
overlapping [si , si [, since the profile already includes the mandatory part of i.
Otherwise, we need to look at [si , ei [. If rectangle j ∗ is not in left-conflict with
i, then no other rectangle can, since it would need to be higher than j ∗ .
To retrieve the index j ∗ , we must answer the question: given a vector of values
and two indices on this vector, what is the index between those two indices that
has the highest value? This kind of query corresponds to the range max query
problem1 , it can be done in constant time, given a linear time preprocessing [7].
Example 1. Assume the vector is values = [5, 4, 2, 1, 4, 3, 0, 8, 2, 3]. The range
max query between index 4 and index 7 is 5, denoted rmq(values, 4, 7) = 5.
This is indeed at index 5 that there is the maximum value on the subvector
[1, 4, 3, 0].
In our case the vector is composed of the heights of the rectangles of the
profile heights = [h1 , h2 , . . . , hm ]. The two indices of the query are respectively:
– j1 (i) is the index j of the rectangle aj , bj , hj
s.t. si ∈ [aj , bj [.
– j2 (i) is the index j of the rectangle aj , bj , hj
s.t. min(ei , si ) − 1 ∈ [aj , bj [
The whole algorithm is given in Algorithm 1. An example of the filtering is
given in Figure 4 (a). Notice that the task is pushed after a highest conflicting
rectangle, which is not as good as the filtering of [14] (Figure 4 (b)).

Algorithm 1. MinLeftTTLinearTime(Ω,C)
Input: A set of tasks Ω, capacity C.
Output: true iff propagation failed, i.e. if the problem is infeasible.
1 initialize TTΩ // aj , bj , hj ∀i ∈ {1 . . . m}
2 if maxj∈[1;m] hj > C then return true
3 heights ← [h1 , h2 , . . . , hm ]
4 ∀i ∈ Ω compute j1 (i), j2 (i)
5 initialize rmq(heights)
6 for i ∈ Ω such that s < s do
7 j ∗ ← rmq(heights, j1 (i), j2 (i))
8 if hj ∗ + ci > C then si ← bj
9 return false

Time Complexity. As in [6], we assume that all the time points are encoded
with w − bit integers and can thus be sorted in linear time. Given the sorted
time points the profile TTΩ can be computed in linear time using a sweep line
algorithm, and all the indices j1 (i), j2 (i) can be computed in linear time as well.
The range min/max query is a well studied problem. Preprocessing in line 5
can be done in linear time, so that any subsequent query at Line 7 executes in
constant time [7]. Thus, the whole algorithm executes in O(n).
1
a straightforward variant of the well-known range min query problem.
154 S. Gay et al.

Discussion. Although the linear time complexity is an improvement over the

O(n log n)algorithm introduced in [14], we believe that this result is more of
theoretical rather than practical interest. The linear time range max query ini-
tialization hides non-negligible constants. The range max query used in Line 7
to reach this time complexity could be implemented by simply iterating on the
rectangles from j1 (i) to j2 (i). On most problems, the interval [si , min(si , e) − 1]
only overlaps a few rectangles of the profile, so the O(n) cost is not high in
practice. Another limitation of the algorithm is that (as for the one of [14]) it
may be called several times before reaching the fix-point (although it does not
suffer from the slow convergence phenomenon described in [3]) either. It may
be more efficient to continue pushing a task further to the right as in [2] rather
than limiting ourselves to only one update per task per call to the procedure.
This is precisely the objective of the algorithm introduced in the next section.

4 An Efficient O(n2 ) Time-Table Filtering

In this section, we introduce a practical algorithm for time-table filtering. It
proceeds in two main phases: first the computation of the mandatory profile,
then a per-task sweeping from left to right and from right to left. This modular
design makes the algorithm simple, and its scalability comes from being able to
exploit structures separately, for instance using sorting only on few tasks. We
review the main phases of Algorithm 2 in execution order.

Building the Profile. Line 2 computes the mandatory profile as a sequence of

rectangles. We process only those tasks in Ω that have a mandatory part. We
will try to reduce that set of tasks further ahead, reducing the work in this part.

Computing Profile Indices. Line 4 computes, for all unfixed tasks i, the pro-
file rectangle containing si . This value is saved between consecutive calls in a
residue2 ; most of the time, it is still valid and we do not have to recompute
it, if not, a dichotomic search is performed, at a cost of O(log n). Note that [2]
sorts tasks by s to locate tasks on the profile, at a theoretical cost of O(n log n).
Here we pay O(log n) only for tasks where the residue is invalid. Similarly, line
5 computes the rectangle containing the last point of i, ei − 1.

Per-Task Sweeping. The loop in line 6 looks for left and right-conflicting rect-
angles for i linearly. The main difference with the global sweeping in [2] is that
our method does not factorize sweeping according to height, wagering that the
cost of moving tasks in height-ordered heaps is higher than that of pushing every
task until no conflict remains. This main part has a worst case cost O(n2 ).

Fruitless Fixed Tasks Removal. After the main loop, line 24 removes fixed tasks
at profile extremities that can no longer contribute to pruning. This filtering is
O(n). Note that Ω is kept along the search tree using a reversible sparse-set [5].
2
this is similar to residual supports for AC algorithms [10].
 Simple and Scalable Time-Table Filtering for the Cumulative Constraint 155

Algorithm 2. ScalableTimeTable(Ω,C)
Input: A set of tasks Ω, capacity C.
Output: true iff propagation failed, i.e. if the problem is infeasible.
1 Ωu ← {i | si < si } // unfixed tasks
2 initialize TTΩ // aj , bj , hj , ∀j ∈ {1 . . . m}
3 if maxj∈[1;m] hj > C then return true
4 ∀i ∈ Ωu , compute j1 (i) such that si ∈ [aj1 (i) ; bj1 (i) [
5 ∀i ∈ Ωu , compute j2 (i) such that ei − 1 ∈ [aj2 (i) ; bj2 (i) [
6 for i ∈ Ωu do
7 j ← j1 (i)
8 s∗i ← si
9 while j ≤ m and aj < min(s∗i + di , si ) do
10 if C − ci < hj then
11 s∗i ← min(bj , si ) // j in left-conflict
12 j ←j+1
13 if s∗i > si then si ← s∗i
14
15 j ← j2 (i)
16 e∗i ← ei
17 while j ≥ 1 and bj ≥ max(e∗i − di , ei ) do
18 if C − ci < hj then
19 e∗i ← max(aj , ei ) // j in right-conflict
20 j ←j−1
21 if e∗i < ei then ei ← e∗i
u
22 smin ← mini∈Ωu si
u
23 emax ← maxi∈Ωu ei
24 Ω ← Ω \ {i ∈ Omega | ei ≤ sumin ∨ eumax ≤ si }
25 return false

5 Experiments
We have tested our ScalableTimeTable filtering against the time-table filtering
of or-tools [12], Choco3 [4] and Gecode [9] solvers. The algorithm in Choco3
and Gecode solver is the one of [2]. Similarly to our algorithm, or-tools also
builds the time-table structure before filtering. To the best of our knowledge, no
implementation of Letort et al [11] algorithm is publicly available. We have thus
implemented the heap-based variant of the algorithm, faster in practice, with the
same quality standard as our new ScalableTimeTable [13]. In the models used
for this experiment, cumulative propagators enforce only the resource constraint,
precedences are enforced by separate propagators.
We have generated randomly n (ranging from 100 to 12800) tasks with dura-
tion between 200 and 2000 and heights between 1 and 40. The capacity is fixed to
100. The search is a simple greedy heuristic selecting the current tasks with the
smallest earliest possible start, hence there is no backtrack. The search finishes
156 S. Gay et al.

when all the tasks have been placed. This simple experimental setting guarantees
that every solver has exactly the same behavior. The results are given on Figure 5.
As can be seen, the time-table implementation of Choco3 and Gecode are quickly
not able to scale well for more than 1600 tasks. The algorithm of or-tools, Letort
et al and our new algorithm are still able to handle 12800 tasks. Surprisingly, our
algorithm outperforms the one of Letort et al despite its simplicity.

298.2 s Gecode

234.8 s Choco 3
175.6 s OR-Tools
128.8 s OscaR Letort

47.8 s OscaR Scalable

0.0 s
100 200 400 800 1600 3200 6400 12800

Fig. 5. Comparison of Time-Table implementations.

6 Conclusion
We have introduced an O(n) time-table filtering using range min queries. We
believe that the usage of range min query may be useful for subsequent research
on scheduling, for instance in time-table disjunctive reasoning [8]. We introduced
simple but scalable O(n2 ) filtering for the cumulative constraint. Our results
show that despite its simplicity, it outperforms current implementations of time-
table constraints in some open-source solvers and also the recent synchronized
sweep algorithm. The resources related to this work are available here http://
bit.ly/cumulativett.

References
1. Aggoun, A., Beldiceanu, N.: Extending chip in order to solve complex scheduling
and placement problems. Mathematical and Computer Modelling 17(7), 57–73
(1993)
2. Beldiceanu, N., Carlsson, M.: A New multi-resource cumulatives constraint
with negative heights. In: Van Hentenryck, P. (ed.) CP 2002. LNCS, vol. 2470,
pp. 63–79. Springer, Heidelberg (2002)
3. Bordeaux, L., Hamadi, Y., Vardi, M.Y.: An analysis of slow convergence in inter-
val propagation. In: Bessière, C. (ed.) CP 2007. LNCS, vol. 4741, pp. 790–797.
Springer, Heidelberg (2007)
4. Charles Prud’homme, X.L., Fages, J.-G.: Choco3 documentation. TASC, INRIA
Rennes, LINA CNRS UMR 6241, COSLING S.A.S. (2014)
5. de Saint-Marcq, V.C., Schaus, P., Solnon, C., Lecoutre, C.: Sparse-sets for domain
implementation. In: CP Workshop on Techniques for Implementing Constraint
Programming Systems (TRICS), pp. 1–10 (2013)
 Simple and Scalable Time-Table Filtering for the Cumulative Constraint 157

6. Fahimi, H., Quimper, C.-G.: Linear-time filtering algorithms for the disjunctive
constraint. In: Twenty-Eighth AAAI Conference on Artificial Intelligence (2014)
7. Fischer, J., Heun, V.: Theoretical and practical improvements on the RMQ-
problem, with applications to LCA and LCE. In: Lewenstein, M., Valiente, G.
(eds.) CPM 2006. LNCS, vol. 4009, pp. 36–48. Springer, Heidelberg (2006)
8. Gay, S., Hartert, R., Schaus, P.: Time-table disjunctive reasoning for the cumula-
tive constraint. In: Michel, L. (ed.) CPAIOR 2015. LNCS, vol. 9075, pp. 157–172.
Springer, Heidelberg (2015)
9. Gecode Team. Gecode: Generic constraint development environment (2006).
http://www.gecode.org
10. Lecoutre, C., Hemery, F., et al.: A study of residual supports in arc consistency.
In: IJCAI, vol. 7, pp. 125–130 (2007)
11. Letort, A., Beldiceanu, N., Carlsson, M.: A Scalable Sweep Algorithm for the cumu-
lative Constraint. In: Milano, M. (ed.) Principles and Practice of Constraint Pro-
gramming. LNCS, pp. 439–454. Springer, Heidelberg (2012)
12. Or-tools Team. or-tools: Google optimization tools (2015). https://developers.
google.com/optimization/
13. OscaR Team. OscaR: Scala in OR (2012). https://bitbucket.org/oscarlib/oscar
14. Ouellet, P., Quimper, C.-G.: Time-Table extended-edge-finding for the cumulative
constraint. In: Schulte, C. (ed.) CP 2013. LNCS, vol. 8124, pp. 562–577. Springer,
Heidelberg (2013)
15. Vilı́m, P.: Timetable edge finding filtering algorithm for discrete cumulative
resources. In: Achterberg, T., Beck, J.C. (eds.) CPAIOR 2011. LNCS, vol. 6697,
pp. 230–245. Springer, Heidelberg (2011)
 General Bounding Mechanism for Constraint
Programs

Minh Hoàng Hà1 , Claude-Guy Quimper2(B) , and Louis-Martin Rousseau1

1
Department of Mathematics and Industrial Engineering and CIRRELT,
École Polytechnique de Montréal, C.P. 6079, Succursale Centre-ville,
Montreal, QC H3C 3A7, Canada
{minhhoang.ha,louis-martin.rousseau}@cirrelt.net
2
Département d’informatique et de génie logiciel and CIRRELT,
Université Laval, Quebec, Canada
claude-guy.quimper@ift.ulaval.ca

Abstract. Integer programming (IP) is one of the most successful

approaches for combinatorial optimization problems. Many IP solvers
make use of the linear relaxation, which removes the integrality require-
ment on the variables. The relaxed problem can then be solved using
linear programming (LP), a very efficient optimization paradigm. Con-
straint programming (CP) can solve a much wider variety of problems,
since it does not require the problem to be expressed in terms of lin-
ear equations. The cost of this versatility is that in CP there is no easy
way to automatically derive a good bound on the objective. This paper
presents an approach based on ideas from Lagrangian decomposition
(LD) that establishes a general bounding scheme for any CP. We pro-
vide an implementation for optimization problems that can be formu-
lated with knapsack and regular constraints, and we give comparisons
with pure CP approaches. Our results clearly demonstrate the benefits
of our approach on these problems.

Keywords: Constraint programming · Automatic bounding ·

Lagrangian decomposition · Knapsack constraint · Regular constraint

1 Introduction
Constraint Programming (CP) is an efficient tool for complex decision problems
arising in industry, such as vehicle routing, scheduling, and resource allocation. CP
solvers essentially solve satisfiability problems, that is, they determine whether or
not there exists a solution to a given model. When applied to optimization prob-
lems, most CP solvers solve a sequence of satisfiability problems, requiring at each
step that the solution found improves on the solution found at the previous step.
The search stops when no feasible solution can be found, proving that the pre-
vious solution was indeed optimal. The ability to compute bounds on the objec-
tive function (the optimization criterion) is crucial. It allows faster termination
of the final subproblem (by proving infeasibility more quickly), and it speeds up

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 158–172, 2015.
DOI: 10.1007/978-3-319-23219-5 12
 General Bounding Mechanism for Constraint Programs 159

the solution of all the subproblems (because filtering and pruning become possible
when a good bound on the objective is known). In all cases where CP has success-
fully solved optimization problems, the modeler had to implement good bounding
mechanisms. The lack of a good general-purpose bounding scheme is one of the
main drawbacks of CP in comparison with integer programming (IP).
In CP, the constraints are independent, and information is communicated
solely through the domain store, the set of possible remaining values for each
variable. During the search, each constraint looks at the domain store, determines
whether a solution still exists, and filters out pairs of variables and values that
no longer lead to feasible solutions. This decoupling of the problem in terms of
the independent constraints allows us to use a simple and efficient combinatorial
algorithm for each constraint. A similar concept is used in linear programming
(LP): Lagrangian decomposition (LD) relaxes the link between difficult sets of
constraints and introduces in the objective a penalty vector that acts as the glue
linking the remaining relatively simple subproblems. In CP, we observe that
there is no such glue for the constraints (or subproblems), making any default
bound computation very weak. If we look at CP through LD glasses, we see that
the relaxed constraints are actually a set of implicit constraints stating that
any given variable should have the same value in all the different constraints in
which it appears. Therefore, to apply LD techniques to a CP model, we must
penalize that different constraints assume the same variable can take different
values during search. Our goal is to develop a general bounding mechanism for
CP that is compatible with every CP model and transparent to the user.
This paper is organized as follows. The next section reviews related work on
decomposition techniques and CP, and Section 3 presents our general bounding
approach. Section 4 evaluates the method on two optimization problems: a) the
knapsack problem and b) the simple shift scheduling problem that can be mod-
eled as a combination of many regular constraints. Finally, Section 5 summarizes
our conclusions.

2 Related Work
We present here some background on LD and the global CP constraints investi-
gated in this paper.

2.1 Lagrangian Decomposition

Decomposition techniques are an efficient approach for large-scale optimization

problems. One of the most popular techniques is Lagrangian relaxation (LR),
a natural approach that exploits the problem structure. It relaxes the “compli-
cated” constraints and penalizes the constraint violation in the objective func-
tion. This simplifies the solution procedure since the resulting problem is easier
to solve.
Several authors have used LR coupled with CP to solve optimization prob-
lems. Sellmann and Fahle [20] propose an algorithm for the automatic recording
160 M.H. Hà et al.

problem. They introduce two linear substructures, and they dualize the first and
propagate the second. At convergence, they use the optimal dual information to
propagate the first substructure. Ouaja and Richards [15] embed LR and CP
into a branch-and-bound algorithm for the traffic placement problem. At each
node of the search tree, CP is used to find a solution if one exists and to prove
infeasibility otherwise, and LD indicates how close the solution is to optimality.
Benoist et al. [2] propose hybrid algorithms combining LR and CP for the trav-
eling tournament problem, which includes round-robin assignment and travel
optimization. They show that LR provides not only good bounds to limit the
search but also information to guide new solutions or efficient branching. Sell-
mann [19] investigates the theoretical foundations of CP-based LR and proves
that suboptimal Lagrangian multipliers can have stronger filtering abilities than
optimal ones.
A drawback of LR is that the problem loses its original structure since the
complicated constraints are removed from the constraint set and penalized in the
objective function. Moreover, the penalized constraints must generally be linear.
This seriously restricts the application of LR in the context of CP where most
global constraints have a nonlinear structure. For example, it is not easy to dual-
ize some popular global constraints such as the all-different, global cardinality,
and regular constraints. To overcome this restriction, Fontaine et al. [8] general-
ize LR to arbitrary high-level models using the notion of satisfiability (violation)
degree. In this method, a satisfiability degree (or violation degree) is defined for
each hard constraint and penalized in the objective function instead of in the
constraint itself. The results show that LR coupled with CP can efficiently solve
some classes of graph coloring problems.
Another way to avoid the penalization of nonlinear hard constraints is to
use LD. This was first introduced by Guignard and Kim [10]; it is also called
variable splitting. It creates “copies” of a set of variables responsible for connect-
ing important substructures in the model, using one copy per substructure and
dualizing the equality of a variable and its copy. Since the substructures of the
original problem are retained, LD provides bounds that are always at least as
tight as those from LR. The method is particularly useful if there are no apparent
complicating constraints or the complicating constraints have nonlinear forms.
The research on integrating LD into CP is limited. To the best of our knowl-
edge, only Cronholm and Ajili [5] have studied this. They propose a hybrid
algorithm for the multicast network design problem. Their approach enables the
propagation of all the substructures at every dual iteration and also provides
strong cost-based filtering. We continue their theme with our attempt to use LD
to construct an automatic bounding mechanism, thus making CP tools more effi-
cient for optimization problems. To do this, we consider each global constraint
in the CP model as a substructure, and we connect them through LD to pro-
vide valid bounds. These are used to prune the nodes in the search tree. We
investigate two global constraints: knapsack and regular constraints.
 General Bounding Mechanism for Constraint Programs 161

2.2 Knapsack and Regular Constraints

In this subsection, we discuss knapsack and regular constraints and formulate

the related optimization problems.

Knapsack Constraint. Knapsack constraints are probably the most commonly

used constraints for problems in mixed integer programming and CP. These
linear constraints have the form wx ≤ W where W is scalar, x = [x1 , x2 , ..., xn ] is
a vector of n binary variables, and w = [w1 , w2 , ..., wn ] is a vector of n coefficients.
Most CP solvers filter knapsack constraints in a straightforward manner: the
domain reduction is based on interval arithmetic and simple bounding argu-
ments. This filtering is fast and effective but requires an important branching
in practice. Another knapsack filtering has been proposed by Trick [22], who
uses a dynamic programming structure to represent the constraint and achieves
hyper-arc consistency, thus determining infeasibility before all the variables are
set. The filtering is based on a layered directed graph where the rows correspond
to W values (0 through |W |), the columns represent variable indexes (0 through
n), and the arcs represent variable-value pairs. This graph has the property that
any path from (0, 0) to the nodes of the last layer corresponds to a feasible
solution to the knapsack constraint.
The filtering can reduce the branching when dealing with the satisfiability
version of the knapsack problem. This version, the market split problem, has no
objective function (see Trick [22] for more information). An effective implemen-
tation is essential to reduce the computational time. In the next sections, we
show that integrating the filtering proposed by Trick [22] with LD reduces both
the branching and the computational time for optimization problems.
Fahle and Sellmann [7] introduce an optimization version of the knap-
sack constraint. Given a lower bound B ∈ N and an vector of n coefficients
p = [p1 , p2 , ..., pn ], the constraint enforces not only wx ≤ W but also px ≥ B.
Since achieving generalized arc consistency for this constraint is NP-hard, Fahle
and Sellmann [7] introduce the notion of relaxed consistency for optimization
constraints, and they use bounds based on LP relaxations for polynomial-time
domain filtering. Using bounds with guaranteed accuracy, Sellmann [18] exploits
an existing approximation algorithm to provide an approximate consistency. Ka-
triel et al. [11] develop an efficient incremental version of the previous algorithm.
Malitsky et al. [13] generalize the method of [11] to provide a filtering algorithm
for a more general constraint, the bounded knapsack constraint, in which the
variables are integer.
The knapsack constraint is the main component of the class of knapsack
problems, which has many variants (see Kellerer et al. [12] for more details).
The simplest is the 0/1 knapsack problem. Given a set of n items, where each
item i has a weight wi and a value pi , we must determine which items to include
in the knapsack so that the total weight is less than or equal to a given limit W
and the total value is maximized. In this paper, we focus on the multidimensional
knapsack problem (MKP), which is as follows:
162 M.H. Hà et al.


n
Maximize pi xi (1)
i=1

n
subject to wij xi ≤ Cj j = 1, . . . , m (2)
i=1
xi ∈ {0, 1} (3)

Regular Constraint. The regular global constraint is generally specified by

using a (nondeterministic) finite automaton π, which is defined as a 5-tuple
π = (Q, A, τ, q0 , F ) where:
– Q is the finite set of states;
– A is the alphabet;
– τ : Q × A × Q is the transition table;
– q0 ∈ Q is the initial state;
– F ⊆ Q is a set of final states.
The constraint regular([X1 , . . . , Xn ], π) holds if the sequence of values of
the variables X1 , . . . , Xn is a member of the regular language recognized by a
deterministic finite automaton π. The recognition is confirmed if there exists a
sequence of states qi0 , . . . , qin such that qi0 = q0 , (qik , Xi , qik+1 ) is a transition
in τ , and qin ∈ F is a final state.
Pesant [16] introduced a filtering algorithm for the regular constraint. It
constructs a layered graph similar to that proposed for the knapsack constraint
above, except that its rows correspond to states of the automaton. This approach
was later extended to the optimization constraints Soft-regular [23], Cost-
regular [6], and Multicost-regular [14] to enforce bounds on the global
cost of the assignment. The underlying solution methods compute the shortest
and longest paths in an acyclic graph.
The regular global constraint is useful in modeling complex work regulations
in the shift scheduling problem (SSP), since the deterministic finite automa-
ton can specify the rules that regulate transitions in the sequence of activi-
ties [3,4,6,17]. The problem consists in scheduling a sequence of activities (work,
break, lunch, and rest activities) for a set of employees. We will investigate a
version of the SSP in which only one employee is considered and the work regula-
tions are so complex that modeling them requires many regular constraints. Let
W be a set of work activities, T a set of periods, Π a set of deterministic finite
automata expressing work regulations, and pij the profit for the assignment of
activity i ∈ W at period j ∈ T . We must assign one and only one activity for
each period such that the transition of the activities over the periods satisfies
the work regulations defined by the automata in Π and the total profit is max-
imized. To state the model for the SSP in CP, we use the element constraint
Element(Index, Table, Value), which holds if Value is equal to the Index
of the Table, i.e., value = Table[Index], to represent the assignment. Let
Xj be the variable representing the activity assigned to period j. The problem
is then:
 General Bounding Mechanism for Constraint Programs 163

Maximize Cj (4)
j∈T

subject to regular(X, πi ) ∀πi ∈ Π (5)

element(Xj , pij , Cj ) ∀i ∈ W, j ∈ T (6)
Xj ∈ W ∀j ∈ T (7)
Cj ∈ R, min pij ≤ Cj ≤ max pij ∀j ∈ T (8)
i∈W i∈W

where constraints (6) ensure that Cj is equal to pij if and only if Xj = i. This
specifies the profit of each assignment.

3 Proposed Approach
3.1 Lagrangian Decomposition
We first recall the LD approach. Consider the problem of computing max{z =
c x|C1 (x) ∧ C2 (x)}, where x is a set of variables, c is a set of coefficients, and
C1 (x) and C2 (x) are two arbitrary constraints. One can obtain a relaxation, i.e.,
an upper bound on the solution, as follows:

 ⏐   ⏐ 
⏐ ⏐
max z = c x⏐C1 (x) ∧ C2 (x) = max c x⏐C1 (x) ∧ C2 (y) ∧ x = y
 ⏐ 
⏐
= max c x + u (x − y)⏐C1 (x) ∧ C2 (y) ∧ x = y
 ⏐ 
⏐
≤ max c x + u (x − y)⏐C1 (x) ∧ C2 (y)
 ⏐   ⏐ 
⏐ ⏐
= max (c + u )x⏐C1 (x) + max −u y ⏐C2 (y)

In LD, a set of variables x is duplicated by introducing an identical set of

variables y, and the difference (x − y) is added to the objective function with a
violation penalty cost u ≥ 0, where u is the vector of Lagrangian multipliers. The
original problem is then relaxed to two subproblems: one with C1 and the other
with C2 . The two subproblems can be solved separately. Solving the resulting
programs with a given parameter u provides a valid upper bound on the original
objective function. The non-negative multipliers u that give the best bound
can be found by solving the Lagrangian dual. The decomposition can easily be
generalized to m constraints:
   
⏐ 
m ⏐ 
m m

 ⏐  1⏐ i i 1
max z = c x⏐ Ci (x) = max c x ⏐ Ci (x ) x =x
i=1 i=1 i=2
 
m
 ⏐ 
m m

 1  1 i ⏐ i i 1
= max c x + u(i) (x − x )⏐ Ci (x ) x =x
i=2 i=1 i=2
 
m
 ⏐ 
m
 1  ⏐1 ii
≤ max c x + u(i) (x − x )⏐ Ci (x )
i=2 i=1
  
m
 ⏐ m
 ⏐

  1⏐ 1  i⏐ i
= max c + u(i) x ⏐C1 (x ) + max −u(i) x ⏐Ci (x )
i=2 i=2
164 M.H. Hà et al.

This decomposition works well for numeric variables, i.e., variables whose
domains contain scalars. In CP, we often encounter domains that contain non-
numeric values. For instance, the regular constraint described in Section 2.2
applies to variables whose domains are sets of characters. In most applications,
it makes no sense to multiply these characters with a Lagrangian multiplier. In
this situation, we do not apply the LD method to the original variables Xi but
instead to a set of binary variables xi,v ∈ {0, 1} where xi,v = 1 ⇐⇒ Xi = v.
Rather than having a Lagrangian multiplier for each variable Xi , we instead
have a multiplier for each binary variable xi,v .
Finding the optimal multipliers is the main challenge in optimizing LD since
it is nondifferentiable. There are several approaches; the most common is the
subgradient method (see Shor et al. [21]). Starting from an arbitrary value of
the multipliers u0 , it solves subproblems iteratively for different values of u.
These are updated as follows:

uk+1 = uk + tk (y k − xk ) (9)

where the index k corresponds to the iteration number, y is a copy of variable

x, and tk is the step size. The step size is computed using the distance between
the objective value of the preceding iteration, Z k−1 , and the estimated optimum
Z ∗:
λ(Z k−1 − Z ∗ )
tk = 0 ≤ λ ≤ 2. (10)

y k−1 − xk−1
2
Here, λ is a scaling factor used to control the convergence; it is normally between
0 and 2.
Our approach integrates an automatic bounding mechanism into the branch-
and-bound algorithm of CP. The idea is that, at each node of the search tree, we
use LD to yield valid bounds. The approach divides the problem into many sub-
problems; each subproblem has one global constraint and an objective function
involving Lagrangian multipliers. For each subproblem, the solver must find the
assignment that satisfies the global constraint while optimizing a linear objective
function. It is generally possible to adapt the filtering algorithm of the global
constraints to obtain an optimal support. If a constraint takes as parameter a
cost for each pair for variable/value and it constrains the cumulative cost of the
assignments to be bounded, then this constraint is compatible with the bounding
mechanism we propose. Soft constraints such as Cost-GCC and Cost-Regular [23]
can therefore be used.

3.2 The Knapsack Constraint

For the MKP, the subproblems are 0/1 knapsack problems. They can be solved
via dynamic programming, which runs in pseudo-polynomial time O(nW ), where
n is the number of items and W is the size of the knapsack. We use the algorithm
by Trick [22] to filter the constraint, and we adapt it to compute the bound. Trick
constructs an acyclic graph with n + 1 layers where Lj contains the nodes of
 General Bounding Mechanism for Constraint Programs 165

layer j for 0 ≤ j ≤ n. Each node is labeled with a pair of integers: the layer
and the accumulated weight. At layer 0, we have a single node L0 = {(0, 0)}.
At layer j, we have Lj = {(j, k) | 1 ∈ dom(Xj ) ∧ (j − 1, k − wj ) ∈ Lj−1 ∨ 0 ∈
dom(Xj ) ∧ (j − 1, k) ∈ Lj−1 }. There is an edge labeled with value 1 between the
nodes (j − 1, k − wj ) and (j, k) whenever these nodes exist and an edge labeled
with value 0 between the nodes (j − 1, k) and (j, k) whenever these nodes exist.
Trick’s algorithm filters the graph by removing all the nodes and edges that do
not lie on a path connecting the node (0, 0) to a node (n, k) for 0 ≤ k ≤ C. After
the graph has been filtered, if no edges labeled with the value 1 remain between
layer j − 1 and layer j, this value is removed from the domain of xj . Similarly,
if no edges labeled with the value 0 remain between layer j − 1 and layer j, this
value is removed from the domain of xj .
We augment this algorithm by computing costs. We assign m to each edge
connecting node (j − 1, k − wj ) to node (j, k) a cost of cj + l=2 u(l, j) for the
graph representing the first constraint (i = 1), where cj is the value of item j
and m is the number of constraints; and a cost of −u(i, j) where i ≥ 2 is the
index of the constraint for the graph representing the remaining constraints. We
use dynamic programming to compute for each node (j, k) the cost of the longest
path connecting the source node (0, 0) to the node (j, k):
⎧
⎪
⎨0 if k = j = 0
M [j, k] = −∞ if k > j = 0
⎪
⎩
max(M [j − 1, k], M [j − 1, k − wj ] + addedCost) otherwise

m
where addedCost is equal to cj + l=2 u(l, j) if i = 1 and −u(i, j) if i > 1.
The optimal bound for this constraint is given by max0≤j≤W M [n, j]. Indeed,
any path in the filtered graph connecting the source node (0, 0) to a node
(n, k) for 0 ≤ k ≤ C corresponds to a valid assignment. The expression
max0≤j≤W M [n, j] returns the largest cost of such a path and therefore the
maximum cost associated with a solution of the knapsack constraint.

3.3 The Regular Constraint

We proceed similarly with the regular constraint. Pesant [16] presents an algo-
rithm also based on a layered graph. The set Lj contains the nodes at layer
j for 0 ≤ j ≤ n. Each node is labeled with an integer representing the layer
and a state of the automaton. The first layer contains a single node, and we
have L0 = {(0, q0 )}. The other layers, for 1 ≤ j ≤ n, contain the nodes
Lj = {(j, q2 ) | (j − 1, q1 ) ∈ Lj−1 ∧ a ∈ dom(Xj ) ∧ (q1 , a, q2 ) ∈ τ }. An edge
is a triplet (n1 , n2 , a) where n1 and n2 are two nodes and a ∈ A is a label. Two
nodes can be connected to each other with multiple edges having distinct labels.
The set of edges is denoted E. There is an edge connecting node (j−1, q1 ) to node
(j, q2 ) with label a whenever these nodes exist and (q1 , a, q2 ) ∈ τ ∧ a ∈ dom(Xj )
holds. As with Trick’s algorithm, Pesant filters the graph by removing all nodes
166 M.H. Hà et al.

and edges that do not lie on a path connecting the source node (0, q0 ) to a node
(n, q) where q ∈ F is a final state. If no edges with label a connecting a node in
Lj−1 to a node in Lj remain, the value a is pruned from the domain dom(Xj ).
This filtering algorithm can be augmented by associating a cost with each
edge. We assume that there is a Lagrangian multiplier u(i, j, a) for each binary
variable xja representing the assignment of variable Xj to character a ∈ A.
Here, i ≥ 1 represents the index of the constraint. An edge (q1 , a, q2 ) has a cost of
m
paj + l=2 u(l, j, a) in the graph induced by the first constraint (i = 1), and a cost
of −u(i, j, a) in the graph induced by the remaining constraints (i > 1). Using
dynamic programming, we can compute the cost of the longest path connecting
node (0, q0 ) to a node (j, q):

0 if j = 0
R[j, q] =
max((j−1,q1 ),(j,q),a)∈E (R[j − 1, q1 ] + addedCost) otherwise
m
where addedCost is equal to paj + l=2 u(l, j, a) if i = 1 and −u(i, j, a) if i > 1.
Since every path from layer 0 to layer n in the graph corresponds to a valid
assignment for the regular constraint, the optimal bound for this constraint is
given by maxq∈F R[n, q], i.e., the greatest cost for a valid assignment.

3.4 The Subgradient Procedure

We use the subgradient procedure to solve the Lagrangian dual. The estimated
optimum Z ∗ is set to the value of the best incumbent solution, and the scaling
factor λ is set to 2. The updating strategy halves λ when the objective function
does not improve in five iterations. At each node of the search tree, the number of
iterations for the subgradient procedure, which gives the best trade-off between
the running time and the number of nodes required in the search tree, is fixed
to 60 for the MKP in which the subproblem is solved by dynamic programming
and to 10 for all other cases. The subgradient procedure is terminated as soon
as the resulting Lagrangian bound is inferior to the value of the best incumbent
solution.
The initial multipliers u0 at the root node are fixed to 1. At the other nodes,
we use an inheritance mechanism, i.e., the value of u0 at a node is taken from
the multipliers of the parent node. This mechanism is better than always fixing
the initial multipliers to a given value. We have tested two strategies: we fix u0
either to the multipliers that give the best bound or to the last multipliers of the
parent node. The results show that the latter strategy generally performs better.
This is because the limited number of subgradient iterations at each node is not
sufficient to provide tight bounds for the overall problem. Therefore, the more
iterations performed at early nodes, the better the multipliers expected at later
nodes.

4 Computational Results
This section reports some experimental results. The goal is not to present state-
of-the-art results for specific problems but to show that LD could make CP tools
 General Bounding Mechanism for Constraint Programs 167

more efficient for optimization problems. The algorithms are built around CP
Optimizer 12.6 with depth-first search. All the other parameters of the solver
are set to their default values. The criteria used to evaluate the solutions are the
number of nodes in the search tree, the computational time, and the number of
instances successfully solved to optimality.
We investigate the behavior of the algorithms in two contexts: with and
without initial lower bounds. In case A, we provide the optimal solution to the
model by adding a constraint on the objective function, setting it to be at least
the value of the optimal solution; the goal is then to prove the optimality of
this value. In case B, we start the algorithms without an initial lower bound.
These two tests provide different insights, since the presence of different bound-
ing mechanisms will impact the search and branching decisions of the solver, with
unpredictable outcomes. However, once the solver has found the optimal solu-
tion, it must always perform a last search to demonstrate optimality. A better
bounding mechanism will always contribute to this phase of the overall search.

4.1 Results for MKP

The experiments in this subsection analyze the performance of LD with two

MKP solution methods: i) direct CP using only sum constraints, and ii) the
approach of Trick [22]. For i) the Lagrangian subproblems are solved by dynamic
programming, and for ii) they are solved using the filtering graph of Trick [22].
We selected two groups of small and medium MKP instances from the OR-
Library [1] for the test: Weing (8 problems, each with 2 constraints) and Weish
(30 problems, each with 5 constraints). We limited the computational time to
one hour.
Table 1 reports the results for case B and Table 2 reports the results for
case A. The instances not included in Table 1 could not be solved by any of
the algorithms; a dash indicates that the algorithm could not solve the instance
within the time limit.
As can be seen in Table 2, the LD approach processes much fewer nodes per
seconds, but the extra effort invested in bounding significantly decreases the size
of the search tree and the time to find the optimal solution, especially for the
larger instances. The most efficient method is CP + LD: it can solve 30 instances.
However, it generally requires more time to solve the small instances. Moreover,
it is quite time-consuming compared with the original method on a few instances,
e.g., Weish6, Weish7, Weish8, and Weing15. The Trick + LD approach improves
on Trick for all the criteria, even for the small instances. This is because solving
the Lagrangian subproblems based on the information available from the filtering
algorithm is computationally less costly. Moreover, the propagation performed
at each node of the search tree is limited for the direct CP formulation, and the
computation of the Lagrangian bound significantly increases this. On the other
hand, the Trick method relies on a filtering algorithm for which the complexity is
linear in the size of the underlying graph. In this case, the additional computation
for the LD bounds has less impact.
168 M.H. Hà et al.

Table 1. Results for MKP without initial lower bound

# CP CP + LD Trick Trick + LD
Instance Nodes
Vars Nodes Time Time Nodes Time Nodes
Time Nodes Time Nodes
Time Nodes Time Nodes
Time
Weing1 28 4424 1.19 3718 860 3.95 218 4424 6.19 715 1100 6.00 183
Weing2 28 5572 1.29 4319 744 3.07 242 5572 11.94 467 744 4.30 173
Weing3 28 8650 0.55 16k 270 0.97 278 8650 15.34 564 280 1.53 183
Weing4 28 4106 1.08 3802 538 2.62 205 4106 19.72 208 538 4.14 130
Weing5 28 13k 0.58 22k 262 1.00 262 12k 21.13 615 262 1.53 171
Weing6 28 9150 1.14 8026 876 3.59 244 9150 20.50 446 1012 4.83 210
Weing7 105 - - - 32k 3410.04 9 - - - - - -
Weing8 105 - - - 19k 147.98 128 - - - 19k 655.46 29
Weish1 30 35k 1.78 20k 1320 37.31 35 35k 910.82 38 1286 59.90 21
Weish2 30 40k 1.64 24k 1280 27.14 47 40k 2481.72 16 1384 57.47 24
Weish3 30 11k 0.83 13k 674 8.47 80 11k 669.01 16 760 24.35 31
Weish4 30 2342 0.80 2927 856 11.91 72 2342 186.47 13 826 29.68 28
Weish5 30 1614 0.90 1793 728 9.19 79 1614 149.78 11 644 23.74 27
Weish6 40 1.2M 12.56 96k 4286 369.13 12 - - - 3368 320.92 10
Weish7 40 901k 15.12 60k 3888 290.79 13 - - - 3482 352.67 10
Weish8 40 1.1M 19.74 56k 4004 256.00 16 - - - 3464 392.57 9
Weish9 40 144k 3.12 46k 2426 46.78 52 - - - 2212 191.83 12
Weish10 50 28M 589.20 48k 4486 264.40 17 - - - 3969 734.65 5
Weish11 50 6M 95.84 63k 3764 159.05 24 - - - 3764 595.38 6
Weish12 50 21M 355.78 59k 4738 307.37 15 - - - 3728 651.90 6
Weish13 50 20M 338.07 59k 4208 250.62 17 - - - 4374 908.53 5
Weish14 60 - - - 8424 645.04 13 - - - 8856 2472.92 4
Weish15 60 35M 720.98 49k 13k 1363.17 10 - - - 12k 3482.11 3
Weish16 60 - - - 14k 1216.45 12 - - - 13k 3324.81 4
Weish17 60 - - - 14k 1600.78 9 - - - - - -
Weish18 70 - - - - - - - - - - - -
Weish19 70 - - - 7750 667.02 12 - - - 8907 3174.07 3
Weish20 70 - - - 18k 2610.74 7 - - - - - -
Weish21 70 - - - 15k 1642.16 9 - - - - - -
Weish22 80 - - - 15k 2905.41 5 - - - - - -
Weish23 80 - - - 12k 2002.74 6 - - - - - -

Table 2 presents the results for case A. Here, the Lagrangian approaches
are even more efficient. The algorithms with LD can successfully solve all 38
instances, and the search tree is relatively small. The computational time of CP
+ LD is still worse than that of CP on a few small instances, but the difference
has decreased significantly. The main reason for this is that the optimal solutions
used to compute the step size make the subgradient procedure more stable and
cause it to converge more quickly. This observation suggests that adding good
lower bounds as soon as possible in the solution process will improve the method.
We computed at the root of the search tree the bound that the LD provides
as well as the bound returned by CPLEX using a linear relaxation (LP) and the
bound that CPLEX computes when the integer variables are not relaxed and all
cuts are added (ILP). Due to lack of space, we do not report the bounds for each
instance. In average, the bounds for LD, LP, and ILP are 1.03%, 0,79%, and 0.22%
greater than the optimal value. Even though LD does not provide the best bound,
it represents a significative improvement to CP whose bound is very weak.
 General Bounding Mechanism for Constraint Programs 169

4.2 Results for SSP

To generate the regular constraints, we use the procedure proposed by Pesant [16]
to obtain random automata. The proportion of undefined transitions in τ is set
to 30% and the proportion of final states to 50%. To control the problem size, we
create instances with only 2 regular constraints and 50 periods. The assignment
profits are selected at random from integer values between 1 and 100. Each
instance is then generated based on two parameters: the number of activities
(nva) and the number of states (nbs) in the automata. For each pair (nva,nbs),
we randomly generate three instances. The instances are labeled nbv-nbs-i, where

Table 2. Results for MKP: proving optimality

# CP CP + LD Trick Trick + LD
Instance Nodes
Vars Nodes Time Time Nodes Time Nodes
Time Nodes Time Nodes
Time Nodes Time Nodes
Time
Weing1 28 3408 0.21 16k 24 0.62 39 3408 5.25 649 24 0.78 31
Weing2 28 5070 0.21 24k 30 0.48 63 5070 16.36 310 30 0.74 41
Weing3 28 7004 0.55 13k 32 0.32 100 7004 11.78 595 34 0.42 81
Weing4 28 2344 0.14 17k 16 0.37 43 2344 10.45 224 16 0.81 20
Weing5 28 18k 0.33 55k 26 0.32 81 18k 35.30 510 26 0.56 46
Weing6 28 8038 0.22 37k 30 0.44 68 8038 27.62 291 30 0.80 38
Weing7 105 - - - 134 4 - - - 134 174.03 1
Weing8 105 - - - 168 3.54 47 - - - 520 40.95 13
Weish1 30 11k 0.27 41k 54 2.85 19 11k 447.18 25 70 5.72 12
Weish2 30 23k 0.43 53k 66 1.98 33 23k 1679.54 14 62 9.15 7
Weish3 30 8394 0.23 36k 38 2.80 14 8394 476.57 18 46 5.59 8
Weish4 30 632 0.14 4514 34 1.26 27 632 57.19 11 38 4.64 8
Weish5 30 556 0.12 4633 34 0.87 39 556 55.93 10 36 5.67 6
Weish6 40 861k 12.92 67k 56 5.54 10 - - - 72 16.58 4
Weish7 40 456k 7.07 64k 60 8.40 7 - - - 72 17.40 4
Weish8 40 707k 10.38 68k 60 9.05 7 - - - 54 19.52 3
Weish9 40 74k 1.26 59k 48 2.28 21 - - - 74 18.18 4
Weish10 50 8.6M 132.92 65k 134 29.53 5 - - - 192 51.11 4
Weish11 50 1.4M 22.75 62k 86 10.50 8 - - - 82 28.44 3
Weish12 50 9M 135.70 66k 114 24.92 5 - - - 122 38.23 3
Weish13 50 4.1M 60.90 67k 120 17.84 7 - - - 112 35.23 3
Weish14 60 - - - 104 27.97 4 - - - 330 123.89 3
Weish15 60 9.7M 173.65 56k 92 28.20 3 - - - 146 68.70 2
Weish16 60 - - - 128 37.95 3 - - - 450 192.18 2
Weish17 60 156M 3537.25 44k 90 34.36 3 - - - 176 115.41 1.5
Weish18 70 - - - 200 87.97 2 - - - 142 116.60 1.2
Weish19 70 - - - 200 86.98 2 - - - 320 228.97 1.4
Weish20 70 - - - 174 65.84 3 - - - 596 340.87 1.7
Weish21 70 - - - 134 48.17 3 - - - 354 225.53 1.6
Weish22 80 - - - 150 87.39 2 - - - 1068 1083.83 1.0
Weish23 80 - - - 158 83.39 1.9 - - - 177 350.55 0.5
Weish24 80 - - - 146 84.90 1.7 - - - 154 194.55 0.8
Weish25 80 - - - 238 117.47 2 - - - 586 539.24 1.1
Weish26 90 - - - 178 135.75 1.3 - - - 438 567.02 0.8
Weish27 90 - - - 152 94.48 1.6 - - - 258 356.33 0.7
Weish28 90 - - - 152 96.71 1.6 - - - 266 344.16 0.8
Weish29 90 - - - 152 104.87 1.4 - - - 416 601.38 0.7
Weish30 90 - - - 152 126.96 1.2 - - - 138 372.65 0.4
170 M.H. Hà et al.

Table 3. Results for SSP without initial lower bound

Pesant Pesant + LD
Instance
Nodes Time Nodes
Time Nodes Time
Nodes
Time
10-20-01 2.2M 1232.72 1785 156k 346.92 450
10-20-02 - - - 298k 741.35 402
10-20-03 1.6M 783.29 2043 158k 332.15 476

10-80-01 256k 1325.80 193 84k 1020.53 82

10-80-02 788k 3307.39 238 238k 2834.86 84
10-80-03 847k 2344.55 361 246k 2176.73 113

20-20-01 - - - 828k 1856.10 446

20-20-02 - - - 1.3M 3404.56 382
20-20-03 2.1M 2427.72 865 164k 439.97 373

20-80-01 - - - 373k 3944.18 95

20-80-02 - - - 436k 5206.81 84
20-80-03 - - - 228k 2561.64 89

Table 4. Results for SSP: proving optimality

Pesant Pesant + LD
Instance
Nodes Time Nodes Nodes
Time Nodes Time Time
10-20-01 87k 42.20 2062 2564 5.94 432
10-20-02 407k 342.43 1189 44k 108.30 406
10-20-03 71k 28.60 2483 4118 7.48 551

10-80-01 20k 118.97 168 4546 71.16 64

10-80-02 25k 81.39 307 5466 63.19 87
10-80-03 26k 85.67 303 4762 58.34 82

20-20-01 343k 176.38 1945 2651 13.41 198

20-20-02 372k 297.97 1248 10k 51.63 194
20-20-03 53k 44.28 1197 1353 7.43 182

20-80-01 105k 486.89 216 10k 147.63 68

20-80-02 216k 1648.28 131 13k 211.23 62
20-80-03 16k 128.74 124 5128 72.33 71

i is the instance number. Because SSP instances are more difficult than MKP
instances, the computational time is increased to 2 hours.
The experiment investigates the performance of LD when combined with the
method of Pesant [16] for the SSP. As for the MKP, we test two cases: case A
has initial lower bounds and case B does not. Tables 3 and 4 present the results,
which demonstrate the performance of our approach. The LD method clearly
improves the method of Pesant [16] for all three criteria. More precisely, for case
B, it can solve 6 additional instances, reducing the size of the search tree by a
factor of about 7 and the computational time by a factor of 1.5 on average. In
case A, it works even better, reducing the size of the search tree by a factor of
about 16 and the computational time by a factor of more than 4 on average.
 General Bounding Mechanism for Constraint Programs 171

This is because we can reuse the graph constructed by the filtering algorithm to
solve the Lagrangian subproblems.

5 Conclusions and Future Work

We have introduced an automatic bounding mechanism based on the LD concept
to improve the performance of CP for optimization problems. We tested the
approach on two problems, the MKP and a synthesized version of the SSP.
These rely on two well-known global constraints: knapsack and regular.
Our next step will be to apply the approach to other global constraints
to solve real problems. One candidate is the global cardinality constraint, for
which the subproblems can be solved in polynomial time with the aid of a
graph obtained from the filtering algorithm. In addition, our approach could
be improved in several ways. First, we could use the information obtained from
solving the Lagrangian dual to design a strong cost-based filtering, as proposed
by Cron- holm and Ajili [5]. Second, by checking the feasibility of the solutions of
the Lagrangian subproblems we could find good solutions that help to limit the
search and improve the convergence of the subgradient method. Moreover, we
could use the solutions of the subproblems to guide the search by branching on
the variable with the largest difference between itself and its copies. Finally, as
reported by Guignard [9], the subgradient approach can have unpredictable con-
vergence behavior, so a more suitable algorithm could improve the performance
of our approach.

Acknowledgement. This work was financed with a Google Research Award. We

would like to thank Laurent Perron for his support.

References
1. Beasley, J.E.: OR-library (2012). http://people.brunel.ac.uk/∼mastjjb/jeb/info.
html
2. Benoist, T., Laburthe, F., Rottembourg, B.: Lagrange relaxation and constraint
programming collaborative schemes for travelling tournament problems. In: CP-
AI-OR 2001, pp. 15–26. Wye College (2001)
3. Chapados, N., Joliveau, M., L’Ecuyer, P., Rousseau, L.-M.: Retail store schedul-
ing for profit. European Journal of Operational Research 239(3), 609–624 (2014).
doi:10.1016/j.ejor.2014.05.033
4. Côté, M.-C., Gendron, B., Quimper, C.-G., Rousseau, L.-M.: Formal languages for
integer programming modeling of shift scheduling problems. Constraints 16(1),
54–76 (2011). doi:10.1007/s10601-009-9083-2
5. Cronholm, W., Ajili, F.: Strong cost-based filtering for lagrange decomposition
applied to network design. In: Wallace, M. (ed.) CP 2004. LNCS, vol. 3258,
pp. 726–730. Springer, Heidelberg (2004). http://www.springerlink.com/index/
ur3uvyqbp0216btd.pdf
6. Demassey, S., Pesant, G., Rousseau, L.-M.: A cost-regular based hybrid column
generation approach. Constraints 11(4), 315–333 (2006). http://dblp.uni-trier.de/
db/journals/constraints/constraints11.html#DemasseyPR06
172 M.H. Hà et al.

7. Fahle, T., Sellmann, M.: Cost based filtering for the constrained knapsack problem.
Annals of OR 115(1–4), 73–93 (2002). doi:10.1023/A:1021193019522
8. Fontaine, D., Michel, L.D., Van Hentenryck, P.: Constraint-based lagrangian relax-
ation. In: O’Sullivan, B. (ed.) CP 2014. LNCS, vol. 8656, pp. 324–339. Springer,
Heidelberg (2014)
9. Guignard, M.: Lagrangean relaxation. Top 11(2), 151–200 (2003). doi:10.1007/
BF02579036. ISSN: 1134–5764
10. Guignard, M., Kim, S.: Lagrangean decomposition: A model yielding stronger
Lagrangean bounds. Mathematical Programming 39(2), 215–228 (1987). doi:10.
1007/BF02592954. ISSN: 00255610
11. Katriel, I., Sellmann, M., Upfal, E., Van Hentenryck, P.: Propagating knapsack
constraints in sublinear time. In: Proceedings of the Twenty-Second AAAI Confer-
ence on Artificial Intelligence, Vancouver, British Columbia, Canada, July 22–26,
pp. 231–236 (2007). http://www.aaai.org/Library/AAAI/2007/aaai07-035.php
12. Kellerer, H., Pferschy, U., Pisinger, D.: Knapsack Problems. Springer, Berlin (2004)
13. Malitsky, Y., Sellmann, M., Szymanek, R.: Filtering bounded knapsack constraints
in expected sublinear time. In: Proceedings of the Twenty-Fourth AAAI Conference
on Artificial Intelligence, AAAI 2010, Atlanta, Georgia, USA, July 11–15, 2010.
http://www.aaai.org/ocs/index.php/AAAI/AAAI10/paper/view/1855
14. Menana, J., Demassey, S.: Sequencing and counting with the multicost-regular
constraint. In: van Hoeve, W.-J., Hooker, J.N. (eds.) CPAIOR 2009. LNCS, vol.
5547, pp. 178–192. Springer, Heidelberg (2009). doi:10.1007/978-3-642-01929-6 14.
ISBN: 3642019285
15. Ouaja, W., Richards, B.: A hybrid multicommodity routing algorithm for traf-
fic engineering. Networks 43(3), 125–140 (2004). http://dblp.uni-trier.de/db/
journals/networks/networks43.html#OuajaR04
16. Pesant, G.: A regular language membership constraint for finite sequences of vari-
ables. In: Wallace, M. (ed.) CP 2004. LNCS, vol. 3258, pp. 482–495. Springer,
Heidelberg (2004). http://www.springerlink.com/content/ed24kyhg561jjthj
17. Quimper, C.-G., Rousseau, L.-M.: A large neighbourhood search approach to the
multi-activity shift scheduling problem. J. Heuristics 16(3), 373–392 (2010). doi:10.
1007/s10732-009-9106-6
18. Sellmann, M.: Approximated consistency for knapsack constraints. In: Rossi, F.
(ed.) CP 2003. LNCS, vol. 2833, pp. 679–693. Springer, Heidelberg (2003). doi:10.
1007/978-3-540-45193-8 46
19. Sellmann, M.: Theoretical foundations of CP-based lagrangian relaxation. In: Wal-
lace, M. (ed.) CP 2004. LNCS, vol. 3258, pp. 634–647. Springer, Heidelberg (2004).
http://dblp.uni-trier.de/db/conf/cp/cp/2004.html#sellmann.ISBN:3-540-23241-9
20. Sellmann, M., Fahle, T.: Constraint programming based Lagrangian relaxation
for the automatic recording problem. Annals of Operations Research 118, 17–33
(2003). doi:10.1023/A:1021845304798. ISBN: 0254-5330
21. Shor, N.Z., Kiwiel, K.C., Ruszcayński, A., Ruszcayński, A.: Minimization methods
for non-differentiable functions. Springer-Verlag New York Inc., New York (1985).
ISBN: 0-387-12763-1
22. Trick, M.A.: A dynamic programming approach for consistency and prop-
agation for knapsack constraints. Annals of OR 118(1–4), 73–84 (2003).
http://dblp.uni-trier.de/db/journals/anor/anor118.html#Trick03
23. van Hoeve, W.J., Pesant, G., Rousseau, L.-M.: On global warming: Flow-
based soft global constraints. J. Heuristics 12(4–5), 347–373 (2006). doi:10.1007/
s10732-006-6550-4
 Smallest MUS Extraction
with Minimal Hitting Set Dualization

Alexey Ignatiev1 , Alessandro Previti2 , Mark Liffiton3 ,

and Joao Marques-Silva1,2
1
INESC-ID, IST, Technical University of Lisbon, Lisboa, Portugal
2
CASL, University College Dublin, Belfield, Ireland
3
Illinois Wesleyan University, Bloomington, IL, USA

Abstract. Minimal explanations of infeasibility are of great interest

in many domains. In propositional logic, these are referred to as Mini-
mal Unsatisfiable Subsets (MUSes). An unsatisfiable formula can have
multiple MUSes, some of which provide more insights than others. Dif-
ferent criteria can be considered in order to identify a good minimal
explanation. Among these, the size of an MUS is arguably one of the
most intuitive. Moreover, computing the smallest MUS (SMUS) finds
several practical applications that include validating the quality of the
MUSes computed by MUS extractors and finding equivalent subformulae
of smallest size, among others. This paper develops a novel algorithm for
computing a smallest MUS, and we show that it outperforms all the pre-
vious alternatives pushing the state of the art in SMUS solving. Although
described in the context of propositional logic, the presented technique
can also be applied to other constraint systems.

1 Introduction

For inconsistent formulae, finding a minimal explanation of their infeasibility

is a central task in order to disclose the source of the problem. In general, an
inconsistent formula can have multiple explanations. Thus, defining a measure of
the quality of an explanation becomes necessary in order to focus on those pro-
viding more insights. For the case of a propositional formula, a natural measure
of explanation quality is the size of the computed minimal unsatisfiable subsets
(MUSes). From a query complexity perspective, extracting an MUS is in FPNP .
In contrast, deciding whether there exists an MUS of size less than or equal to
k is Σ2P -complete [8,12]. As a consequence, extracting a smallest MUS (SMUS)
P
is in FPΣ2 .
Computing an SMUS is central to a number of practical problems, includ-
ing validating MUSes computed with modern MUS extractors and finding an
equivalent subformula of minimum size.

This work is partially supported by SFI PI grant BEACON (09/IN.1/ I2618),

FCT grant POLARIS (PTDC/EIA-CCO/123051/2010) and national funds through
Fundação para a Ciência e a Tecnologia (FCT) with reference UID/CEC/50021/2013.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 173–182, 2015.
DOI: 10.1007/978-3-319-23219-5 13
174 A. Ignatiev et al.

This paper shows how recent work on iterative enumeration of hitting sets can
be adapted to computing the smallest MUS of unsatisfiable formulae, improving
upon state-of-the-art algorithms. The approach is implemented and tested on
Boolean Satisfiability instances, but the technique applies equally well to com-
puting SMUSes of any constraint system for which the satisfiability of all subsets
is well-defined.
The paper is organized as follows. In Section 2 basic definitions are provided.
Section 3 introduces an algorithm for the extraction of an SMUS. In Section 4
some optimizations are presented. Finally, Section 5 is dedicated to the presen-
tation of the experimental results, and Section 6 concludes the paper.

2 Preliminaries
This section provides common definitions that will be used throughout the paper.
Additional standard definitions are assumed (e.g. see [3]). In what follows, F
denotes a propositional formula expressed in Conjunctive Normal Form (CNF).
A formula in CNF is a conjunction of clauses, where each clause is a disjunction
of literals. A literal is either a Boolean variable or its complement. Also, we may
refer to formulae as sets of clauses and clauses as sets of literals. An assignment
to variables that satisfies formula F is said to be a model of F. A formula is
unsatisfiable when no model exists. Unless specified explicitly, F is assumed to
be unsatisfiable. The following definitions also apply.
Definition 1. A subset U ⊆ F is a minimal unsatisfiable subset (MUS) if U is
unsatisfiable and ∀U  ⊂ U, U  is satisfiable. An MUS U of F of the smallest size
is called a smallest MUS (SMUS).
Definition 2. A subset C of F is a minimal correction subset (MCS) if F \ C
is satisfiable and ∀C  ⊆ C ∧ C  = ∅, (F \ C) ∪ C  is unsatisfiable.
Definition 3. A satisfiable subset S ⊆ F is a maximal satisfiable subset (MSS)
if ∀S  ⊆ F s.t. S ⊂ S  , S  is unsatisfiable.
An MSS can also be defined as the complement of an MCS (and vice versa).
If C is an MCS, then S = F \ C represents an MSS. On the other hand, MUSes
and MCSes are related by the concept of minimal hitting set.
Definition 4. Given a collection Γ of sets from a universe U, a hitting set h
for Γ is a set such that ∀S ∈ Γ, h ∩ S = ∅.
A hitting set h is minimal if none of its subset is a hitting set. The minimal
hitting set duality between MUSes and MCSes is well-known (e.g. see [15,17]):
Proposition 1. Given a CNF formula F, let MUSes(F) and MCSes(F) be the
set of all MUSes and MCSes of F, respectively. Then the following holds:
1. A subset U of F is an MUS iff U is a minimal hitting set of MCSes(F).
2. A subset C of F is an MCS iff C is a minimal hitting set of MUSes(F).
The duality relating MUSes and MCSes is a key aspect of the algorithm
presented below. In the next section, we will describe how the smallest MUS can
be computed by exploiting this observation.
 Smallest MUS Extraction with Minimal Hitting Set Dualization 175

Algorithm 1. Basic SMUS algorithm

Input: CNF formula F
1 begin
2 H←∅
3 while true do
4 h ← MinimumHS(H)
5 F  ← {ci | ei ∈ h}
6 if not SAT(F  ) then
7 return SMUS ← F 
8 else
9 C ← grow(F  )
10 H ← H ∪ {C}
11 end

3 Basic Algorithm

This section describes the new SMUS algorithm. We start by providing the
definition of a minimum size hitting set:

Definition 5. Let Γ be a collection of sets and MHS(Γ ) the set of all minimal
hitting sets on Γ . Then a hitting set h ∈ M HS(Γ ) is said to be a minimum
hitting set if ∀ h ∈ M HS(Γ ) we have that |h| ≤ |h |.

The algorithm is based on the following observation:

Proposition 2. A set U ⊆ F is an SMUS of F if and only if U is a minimum
hitting set of MCSes(F).
By duality [17], we also have that the minimum hitting set of MUSes(F)
corresponds to a MaxSAT solution for F. This observation has already been
exploited in [6]. Algorithm 1 can be seen as the dual of the algorithm presented
in [6]. H represents a collection of sets, where each set corresponds to an MCS
on F. Thus, elements of the sets in H represent clauses of F. Let elm(H) denote
the set of elements in H. Each element e ∈ elm(H) is associated with a clause
c ∈ F. At the beginning of the algorithm H is empty (line 2). At each step, a
minimum hitting set h is computed on H (see line 4) and the induced formula F 
is tested for satisfiability at line 6. If F  is satisfiable, it is extended by the grow
procedure into an MSS containing F  , the complement of which is returned as
the MCS C. Then C is added to the collection H. If instead F  is unsatisfiable
then F  is guaranteed to be an SMUS of F as the following lemma states:

Lemma 1. Let K ⊆ MCSes(F). Then a subset U of F is an SMUS if U is a

minimum hitting set on K and U is unsatisfiable.

Proof. Since U is unsatisfiable it means that it already hits every MCS in

MCSes(F) (Proposition 1). U is also a minimum hitting set on MCSes(F), since
176 A. Ignatiev et al.

c1 = ¬s c2 = s ∨ ¬p c3 = p
c5 = ¬m ∨ n c6 = ¬n
c4 = ¬p ∨ m

c7 = ¬m ∨ l

c8 = ¬l

Fig. 1. Formula example

Table 1. Example SMUS computation

MinimumHS(H) SAT(F  ) F \ grow(F  , F ) H=H∪C

F  ← {∅} true C ← {c3 } {{c3 }}

F  ← {c3 } true C ← {c2 , c4 } {{c3 }, {c2 , c4 }}

F  ← {c2 , c3 } true C ← {c1 , c4 } {{c3 }, {c2 , c4 }, {c1 , c4 }}

F  ← {c3 , c4 } true C ← {c1 , c5 , c7 } {{c3 }, {c2 , c4 }, {c1 , c4 }, {c1 , c5 , c7 }}

F  ← {c1 , c2 , c3 } false {c1 , c2 , c3 } is an SMUS of F

it is a minimum hitting set for K ⊆ MCSes(F) and no other added MCS can
make it grow in size. Moreover, all the other hitting sets can either grow in size
or remain the same. Thus, by Proposition 2 U must be an SMUS. 

Lemma 1 states that it is not necessary to enumerate all MCSes of formula F.
Instead, it is enough to compute only those whose minimum hitting set is an
MUS. Therefore, Algorithm 1 terminates once the first MUS is computed, which
is by construction guaranteed to be of the smallest size.
It is worth noting that nothing in Algorithm 1 is specific to Boolean CNF,
and in fact the algorithm can be applied to any type of constraint system for
which the satisfiability of constraint subsets can be checked. The algorithm and
all following additions to it are constraint agnostic.
An example of a run of Algorithm 1 is shown in Table 1 (Figure 1 illustrates
the input formula). The first column contains the formula F  induced by the
minimum hitting set on H. Whenever F  is satisfiable (second column), the grow
procedure (see line 9 of Algorithm 1) returns an MCS (third column). The last
 Smallest MUS Extraction with Minimal Hitting Set Dualization 177

column shows the current H, the collected set of MCSes from which a minimum
hitting set will be found in the next iteration. In the last row, F  = {c1 , c2 , c3 }
and since the call SAT(F  ) returns false, set {c1 , c2 , c3 } represents an SMUS.
Notice that MinimumHS(H) could have returned {c3 , c4 , c5 } instead of {c1 , c2 , c3 }.
In this case, further iterations would be necessary in order to find an SMUS.

4 Additional Details

This section describes some essential details of the approach being proposed.
These include computing a minimum hitting set with at most one SAT call,
enumerating disjoint MCSes, and reporting an upper bound within a reasonably
short time.

4.1 Reducing the Number of SAT Calls

A number of different approaches can be used for computing minimum size

hitting sets [6]. This section describes a different alternative that exploits the
use of SAT solvers as well as the problem structure. The proposed approach
exploits incremental MaxSAT solving [5].
Let the current minimum hitting set (MHS) h have size k, and let C be a
new set to hit. Call a SAT solver on the resulting formula, requesting an MHS
of size k. If the formula is satisfiable, then we have a new MHS of size k that
also hits C. However, if the formula is unsatisfiable, this means there is no MHS
of size k. Thus, an MHS of size k + 1 can be constructed by adding to h any
element of C. Clearly, by construction, h does not hit any element of C. Thus,
every MHS can be obtained with a single call to a SAT solver.
Additionally, extraction of each MCS C (see line 9) requires a number of
SAT calls. Alternatively, one can avoid the computation of MCS C and use
any correction subset C  s.t. C ⊆ C  instead. Indeed, any MHS of the set of all
correction subsets of F also hits every MCS of F [17]. Moreover, observe that
a proposition similar to Lemma 1 can be proved for a partial set of correction
subsets of F. Therefore, the grow procedure can be skipped, and the complement
to F  can be directly added to H.

4.2 Disjoint MCS Enumeration

Enumeration of disjoint inconsistencies has been used in various settings in recent

years. For example, disjoint unsatisfiable core enumeration in MaxSAT was pro-
posed in [16], and nowadays it is often used in state-of-the-art MaxSAT solvers
(e.g. see [7]). Enumeration of disjoint MCSes, which act as unsatisfiable cores
over quantified constraints, has also proven its relevance for the SMUS problem
in the context of QMaxSAT [9]. Also note that disjoint MCSes are known to
have a strong impact on the performance of the branch-and-bound algorithms
for SMUS by refining the lower bound on the size of the SMUS [13].
178 A. Ignatiev et al.

Our approach based on the minimum hitting set duality [15,17] between
MCSes and MUSes of an unsatisfiable CNF formula F can also make use of dis-
joint MCSes. Indeed, since each MCS must be hit by any MUS of a CNF formula,
disjoint MCSes must be hit by the smallest MUS of the formula. Therefore, given
a set of disjoint MCSes D, one can initialize set H with D instead of the empty
set (see line 2 of Algorithm 1) in order to boost the performance of Algorithm 1.
Note that this also simplifies the computation of the first minimum hitting set,
which for a set of disjoint MCSes of size k is exactly of size k. According to the
experimental results described in Section 5, this improvement has a huge impact
on the overall performance of the proposed approach.

4.3 Finding Approximate Solutions

Although the approach being proposed performs better than the known alter-
native approaches to SMUS, this problem is computationally much harder than
extracting any MUS of a CNF formula (the decision version of the SMUS problem
is known to be Σ2P -complete, e.g. see [8,12]). One may find this complexity char-
acterization a serious obstacle for using SMUS algorithms in practice. Indeed, a
user may prefer to find an MUS close to the smallest size within a reasonable
amount of time instead of waiting until the smallest MUS is found.
In order to resolve this issue and following the ideas of [10], Algorithm 1
can be extended for computing a “good” upper bound on the exact solution of
the SMUS problem. Any MUS can be considered as an upper bound on the
smallest MUS. Therefore and since extracting one MUS is a relatively simple
task, enumerating MUSes within a given time limit and choosing the smallest
one among them can be satisfactory for a user. MUS enumeration can be done in
a way similar to the one proposed in [14]. Observe that Algorithm 1 iteratively
refines a lower bound of the SMUS, and so can serve to measure the quality of
approximate upper bounds.
This pragmatic policy of computing an upper bound before computing the
exact solution provides a user with a temporary approximate answer within a
short period of time and continues with computing the exact solution, if needed
(i.e. if the user is not satisfied with the quality of the upper bound). Otherwise,
the upper bound is enough and the user can stop the process.

5 Experimental Results

This section evaluates the approach to the smallest MUS problem proposed
in this paper. The experiments were performed in Ubuntu Linux on an Intel
Xeon E5-2630 2.60GHz processor with 64GByte of memory. The time limit was
set to 800s and the memory limit to 10GByte. The approach proposed above was
implemented in a prototype called forqes (FORmula QuintESsence extractor ).
The underlying SAT solver of forqes is Glucose 3.01 [1]. A weakened version

1
Available from http://www.labri.fr/perso/lsimon/glucose/
 Smallest MUS Extraction with Minimal Hitting Set Dualization 179

800
vbs
700 forqes
minuc
600
digger
500 forqes-w
CPU time (s)

400

300

200

100

0
300 350 400 450 500
instances

Fig. 2. Cactus plot showing the performance of forqes, minuc, and digger

of forqes, which does not use disjoint MCS enumeration and implements Algo-
rithm 1, is referred to as forqes-w.
The performance of forqes was compared to the state-of-the-art approach
to the SMUS problem that uses the quantified MaxSAT formulation of SMUS [9].
The most efficient version of the tool described in [9] performs core-guided
QMaxSAT solving; in the following it is referred to as minuc. Additionally,
a well-known branch-and-bound SMUS extractor called digger (see [13]) was
also considered. Note that the versions of minuc and digger participating in
the evaluation make use of disjoint MCS enumeration.
Several sets of benchmarks were used to assess the efficiency of the new algo-
rithm, all used in the evaluation in [9] as well. This includes a collection of
instances from automotive product configuration benchmarks [18] and two sets
of circuit diagnosis instances. Additionally, we selected instances from the com-
plete set of the MUS competitions benchmarks2 as follows. Because extracting a
smallest MUS of a CNF formula is computationally much harder than extracting
any MUS of the formula, the instances that are difficult for a state-of-the-art
MUS extractor were excluded. Instead, we considered only formulae solvable by
the known MUS extractor muser-2 (e.g. see [2]) within 10 seconds. The total
number of instances considered in the evaluation is 682.
Figure 2 shows a cactus plot illustrating the performance of the tested solvers
on the total set of instances. forqes exhibits the best performance, being able
to solve 483 instances. minuc comes second with 448 instances solved. Thus,
forqes solves 7.8% more instances than minuc. digger and forqes-w have
almost the same performance solving 371 and 367 instances, respectively.

2
See http://www.satcompetition.org/2011/
180 A. Ignatiev et al.

103 103
800 sec. timeout 800 sec. timeout

102 102

101 101
digger

minuc
100 100

10−1 10−1

800 sec. timeout

800 sec. timeout

10−2 10−2

10−3 −3 10−3 −3
10 10−2 10−1 100 101 102 103 10 10−2 10−1 100 101 102 103
forqes forqes

(a) forqes vs digger (b) forqes vs minuc

Fig. 3. Performance of forqes, minuc, and digger

More details on the solvers’ performance can be found in Figure 3a, Figure 3b,
and Figure 4a. Also note that the virtual best solver (vbs) among all the con-
sidered solvers is able to solve 485 instances, which is only 2 more than what
forqes can solve on its own. Interestingly, neither minuc nor digger contribute
to the vbs. Although the weakened version of forqes (forqes-w) performs
quite poorly, it is the only solver (besides the normal version of forqes) that
contributes to the vbs. This can be also seen in Figure 4a.
As it was described in Section 4.3, the approach being proposed can report
an upper bound, which can be used as an approximation of the SMUS if finding
the exact solution is not efficient and requires too much time. The following
evaluates the quality of the upper bound reported by this pragmatic solving
strategy. Given an upper bound U B computed within some time limit and the
optimal value opt ≤ U B, the closer value Uopt B
is to 1 the better the quality of
U B is. Since it is often hard to find the exact optimal value, one can consider
a lower bound on the exact solution instead of the optimal value in order to
estimate the quality of the upper bound. Indeed, any set of disjoint MCSes found
within a given timeout can be seen as a lower bound on the exact solution.
Given a CNF formula, an upper bound on its smallest MUS reported by
forqes within a time limit is computed with an external call to a known MUS
enumerator called marco [14]. Several timeout values were tested, namely 5,
10, and 20 seconds. Figure 4b shows a cactus plot illustrating the quality of the
upper bounds computed this way. It is not surprising that generally the more
time is given to the solver, the better upper bound is computed. For about 400
instances, value U B
LB is extremely close to 1 meaning that the upper bound is
usually almost equal to the lower bound. As one can see, the upper bound is
less than an order of magnitude larger than the lower bound for about 550, 585,
and 620 instances if computed within 5, 10, and 20 seconds, respectively. Also
 Smallest MUS Extraction with Minimal Hitting Set Dualization 181

103 103
800 sec. timeout
t = 20s
2
t = 10s
10
t = 5s

Upper bound quality (U B/LB)

101 102
forqes-w

100

10−1 101

800 sec. timeout

10−2

10−3 −3 100
10 10−2 10−1 100 101 102 103 400 450 500 550 600 650
forqes instances

(a) forqes vs forqes-w (b) UB quality in 5, 10, and 20 sec

Fig. 4. Performance of forqes-w and UB quality

note that both lower and upper bounds are relatively easy to compute, e.g. both
of them can be computed almost for 650 instances (out of 682). Moreover, they
give an idea of how large the interval is between them. Given this information,
a user can decide how critical it is to compute the exact solution.
In summary, the experimental results indicate that the proposed approach
pushes the state of the art in SMUS solving, outperforming all the previous
approaches in terms of the number of solved instances. Moreover, the scatter
plots indicate that in most of the cases the proposed approach is also the fastest
in comparison to others. Moreover, a pragmatic policy to report an upper bound
within a short period of time can be helpful when it is hard to compute the exact
solution and provide a user with an approximate solution of the SMUS problem
of reasonable size.

6 Conclusion

This paper adapts recent algorithms for implicit set problems [4,6,11,19] to the
case of computing the smallest MUS. A number of enhancements are developed
and added to the new algorithm. Experimental results, obtained on represen-
tative problems instances, show clear gains over what currently represents the
state of the art. A natural line of research is to apply the novel SMUS algorithm
in concrete practical applications of SMUSes, e.g. finding smallest equivalent
subformulae of smallest size.
182 A. Ignatiev et al.

References
1. Audemard, G., Lagniez, J.-M., Simon, L.: Improving glucose for incremental
SAT solving with assumptions: application to MUS extraction. In: Järvisalo, M.,
Van Gelder, A. (eds.) SAT 2013. LNCS, vol. 7962, pp. 309–317. Springer,
Heidelberg (2013)
2. Belov, A., Lynce, I., Marques-Silva, J.: Towards efficient MUS extraction. AI Com-
mun. 25(2), 97–116 (2012)
3. Biere, A., Heule, M., van Maaren, H., Walsh, T. (eds.): Handbook of Satisfiability.
Frontiers in Artificial Intelligence and Applications, vol. 185. IOS Press (2009)
4. Chandrasekaran, K., Karp, R.M., Moreno-Centeno, E., Vempala, S.: Algorithms
for implicit hitting set problems. In: SODA, pp. 614–629 (2011)
5. Cimatti, A., Griggio, A., Schaafsma, B.J., Sebastiani, R.: A modular approach
to MaxSAT modulo theories. In: Järvisalo, M., Van Gelder, A. (eds.) SAT 2013.
LNCS, vol. 7962, pp. 150–165. Springer, Heidelberg (2013)
6. Davies, J., Bacchus, F.: Solving MAXSAT by solving a sequence of simpler SAT
instances. In: Lee, J. (ed.) CP 2011. LNCS, vol. 6876, pp. 225–239. Springer,
Heidelberg (2011)
7. Davies, J., Bacchus, F.: Exploiting the power of mip solvers in maxsat. In:
Järvisalo, M., Van Gelder, A. (eds.) SAT 2013. LNCS, vol. 7962, pp. 166–181.
Springer, Heidelberg (2013)
8. Gupta, A.: Learning Abstractions for Model Checking. PhD thesis, Carnegie Mel-
lon University, June 2006
9. Ignatiev, A., Janota, M., Marques-Silva, J.: Quantified maximum satisfiability: a
core-guided approach. In: Järvisalo, M., Van Gelder, A. (eds.) SAT 2013. LNCS,
vol. 7962, pp. 250–266. Springer, Heidelberg (2013)
10. Ignatiev, A., Janota, M., Marques-Silva, J.: Towards efficient optimization in pack-
age management systems. In: ICSE, pp. 745–755 (2014)
11. Karp, R.M.: Implicit hitting set problems and multi-genome alignment. In: Amir,
A., Parida, L. (eds.) CPM 2010. LNCS, vol. 6129, pp. 151–151. Springer, Heidelberg
(2010)
12. Liberatore, P.: Redundancy in logic I: CNF propositional formulae. Artif. Intell.
163(2), 203–232 (2005)
13. Liffiton, M.H., Mneimneh, M.N., Lynce, I., Andraus, Z.S., Marques-Silva, J.,
Sakallah, K.A.: A branch and bound algorithm for extracting smallest minimal
unsatisfiable subformulas. Constraints 14(4), 415–442 (2009)
14. Liffiton, M.H., Previti, A., Malik, A., Marques-Silva, J.: Fast, flexible MUS enu-
meration. Constraints (2015). http://dx.doi.org/10.1007/s10601-015-9183-0
15. Liffiton, M.H., Sakallah, K.A.: Algorithms for computing minimal unsatisfiable
subsets of constraints. J. Autom. Reasoning 40(1), 1–33 (2008)
16. Marques-Silva, J., Planes, J.: Algorithms for maximum satisfiability using unsat-
isfiable cores. In: DATE, pp. 408–413 (2008)
17. Reiter, R.: A theory of diagnosis from first principles. Artif. Intell. 32(1), 57–95
(1987)
18. Sinz, C., Kaiser, A., Küchlin, W.: Formal methods for the validation of automotive
product configuration data. AI EDAM 17(1), 75–97 (2003)
19. Stern, R.T., Kalech, M., Feldman, A., Provan, G.M.: Exploring the duality in
conflict-directed model-based diagnosis. In: AAAI, pp. 828–834 (2012)
 Upper and Lower Bounds on the Time
Complexity of Infinite-Domain CSPs

Peter Jonsson and Victor Lagerkvist(B)

Department of Computer and Information Science, Linköping University,

Linköping, Sweden
{peter.jonsson,victor.lagerkvist}@liu.se

Abstract. The constraint satisfaction problem (CSP) is a widely stud-

ied problem with numerous applications in computer science. For infinite-
domain CSPs, there are many results separating tractable and NP-hard
cases while upper bounds on the time complexity of hard cases are virtu-
ally unexplored. Hence, we initiate a study of the worst-case time cmplex-
ity of such CSPs. We analyse backtracking algorithms and show that they
can be improved by exploiting sparsification. We present even faster algo-
rithms based on enumerating finite structures. Last, we prove non-trivial
lower bounds applicable to many interesting CSPs, under the assumption
that the strong exponential-time hypothesis is true.

1 Introduction

The constraint satisfaction problem over a constraint language Γ (CSP(Γ )) is

the problem of finding a variable assignment which satisfies a set of constraints,
where each constraint is constructed from a relation in Γ . This problem is a
widely studied computational problem and it can be used to model many classi-
cal problems such as k-colouring and the Boolean satisfiability problem. In the
context of artificial intelligence, CSPs have been used for formalizing a wide
range of problems, cf. Rossi et al. [30]. Efficient algorithms for CSP problems
are hence of great practical interest. If the domain D is finite, then a CSP(Γ )
instance I with variable set V can be solved in O(|D||V | · poly(||I||)) time by
enumerating all possible assignments. Hence, we have an obvious upper bound
on the time complexity. This bound can, in many cases, be improved if addi-
tional information about Γ is known, cf. the survey by Woeginger [36] or the
textbook by Gaspers [14]. There is also a growing body of literature concerning
lower bounds [16,20,21,33].
When it comes to CSPs over infinite domains, there is a large number of
results that identify polynomial-time solvable cases, cf. Ligozat [23] or Rossi
et al. [30]. However, almost nothing is known about the time complexity of
solving NP-hard CSP problems. One may conjecture that a large number of
practically relevant CSP problems do not fall into the tractable cases, and this
motivates a closer study of the time complexity of hard problems. Thus, we
initiate such a study in this paper. Throughout the paper, we measure time

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 183–199, 2015.
DOI: 10.1007/978-3-319-23219-5 14
184 P. Jonsson and V. Lagerkvist

complexity in the number of variables. Historically, this has been the most com-
mon way of measuring time complexity. One reason is that an instance may be
massively larger than the number of variables — a SAT instance with n variables
may contain up to 22n distinct clauses if repeated literals are disallowed — and
measuring in the instance size may give far too optimistic figures, especially since
naturally appearing test examples tend to contain a moderate number of con-
straints. Another reason is that in the finite-domain case, the size of the search
space is very closely related to the number of variables. We show that one can
reason in a similar way when it comes to the complexity of many infinite-domain
CSPs.
The relations in finite-domain CSPs are easy to represent by simply listing
the allowed tuples. When considering infinite-domain CSPs, the relations need to
be implicitly represented. A natural way is to consider disjunctive formulas over
a finite set of basic relations. Let B denote some finite set of basic relations such
that CSP(B) is tractable. Let B ∨∞ denote the closure of B under disjunctions,
and let B ∨k be the subset of B ∨∞ containing only disjunctions of length at
most k. Consider the following example: let D = {true, f alse} and let B =
{B1 , B2 } where B1 = {true} and B2 = {f alse}. It is easy to see that CSP(B ∨∞ )
corresponds to the Boolean SAT problem while CSP(B ∨k ) corresponds to the
k-SAT problem.
CSPs in certain applications such as AI are often based on binary basic rela-
tions and unions of them (instead of free disjunctive formulas). Clearly, such
relations are a subset of the relations in B ∨k and we let B ∨= denote this set
of relations. We do not explicitly bound the length of disjunctions since they
are bounded by |B|. The literature on such CSPs is voluminous and we refer
the reader to Renz and Nebel [29] for an introduction. The languages B ∨∞ and
B ∨k have been studied to a smaller extent in the literature. There are both
works studying disjunctive constraints from a general point of view [9,11] and
application-oriented studies; examples include temporal reasoning [19,31], inter-
active graphics [27], rule-based reasoning [25], and set constraints (with applica-
tions in descriptive logics) [4]. We also note (see Section 2.2 for details) that there
is a connection to constraint languages containing first-order definable relations.
Assume Γ is a finite constraint language containing relations that are first-order
definable in B, and that the first order theory of B admits quantifier elimina-
tion. Then, upper bounds on CSP(Γ ) can be inferred from results such as those
that will be presented in Sections 3 and 4. This indicates that studying the time
complexity of CSP(B ∨∞ ) is worthwhile, especially since our understanding of
first-order definable constraint languages is rapidly increasing [3].
To solve infinite-domain CSPs, backtracking algorithms are usually employed.
Unfortunately, such algorithms can be highly inefficient in the worst case. Let
p denote the maximum arity of the relations in B, let m = |B|, and let |V |
denote the number of variables in a given CSP instance. We show (in Section 3.1)
m·|V |p
·log(m·|V |p )
that the time complexity ranges from O(22 · poly(||I||)) (which
is doubly exponential with respect to the number of variables) for CSP(B ∨∞ )
m p
to O(22 ·|V | ·log m · poly(||I||)) time for B ∨= (and the markedly better bound
 Upper and Lower Bounds on the Time Complexity of Infinite-Domain CSPs 185

p
of O(2|V | log m · poly(||I||)) if B consists of pairwise disjoint relations.) The use
of heuristics can probably improve these figures in some cases, but we have
not been able to find such results in the literature and it is not obvious how
to analyse backtracking combined with heuristics. At this stage, we are mostly
interested in obtaining a baseline: we need to know the performance of simple
algorithms before we start studying more sophisticated ones. However, some of
these bounds can be improved by combining backtracking search with methods
for reducing the number of constraints. We demonstrate this with sparsifica-
tion [18] in Section 3.2.
In Section 4 we switch strategy and show that disjunctive CSP problems can
be solved significantly more efficiently via a method we call structure enumera-
tion. This method is inspired by the enumerative method for solving finite-domain
p
CSPs. With this algorithm, we obtain the upper bound O(2|V | ·m · poly(||I||)) for
∨∞
CSP(B ). If we additionally assume that B is jointly exhaustive and pairwise dis-
p
joint then the running time is improved further to O(2|V | ·log m · poly(||I||)). This
bound beats or equals every bound presented in Section 3. We then proceed to
show even better bounds for certain choices of B. In Section 4.2 we consider equal-
ity constraint languages over a countably infinite domain and show that such CSP
problems are solvable in O(|V |B|V | · poly(||I||) time, where B|V | is the |V |-th Bell
number. In Section 4.3 we focus on three well-known temporal reasoning problems
and obtain significantly improved running times.
We tackle the problem of determining lower bounds for CSP(B ∨∞ ) in
Section 5, i.e. identifying functions f such that no algorithm for CSP(B ∨∞ )
has a better running time than O(f (|V |)). We accomplish this by relating CSP
problems and certain complexity-theoretical conjectures, and obtain strong lower
bounds for the majority of the problems considered in Section 4. As an exam-
ple, we show that the temporal CSP({<, >, =}∨∞ ) problem is solvable in time
O(2|V | log |V | ·poly(||I||)) but, assuming a conjecture known as the strong exponen-
tial time hypothesis (SETH), not solvable in O(c|V | ) time for any c > 1. Hence,
even though the algorithms we present are rather straightforward, there is, in
many cases, very little room for improvement, unless the SETH fails.

2 Preliminaries
We begin by defining the constraint satisfaction problem and continue by dis-
cussing first-order definable relations.

2.1 Constraint Satisfaction

Definition 1. Let Γ be a set of finitary relations over some set D of values.
The constraint satisfaction problem over Γ (CSP(Γ )) is defined as follows:
Instance: A set V of variables and a set C of constraints of the form
R(v1 , . . . , vk ), where k is the arity of R, v1 , . . . , vk ∈ V and R ∈ Γ .
Question: Is there a function f : V → D such that (f (v1 ), . . . , f (vk )) ∈ R for
every R(v1 , . . . , vk ) ∈ C?
186 P. Jonsson and V. Lagerkvist

The set Γ is referred to as the constraint language. Observe that we do

not require Γ or D to be finite. Given an instance I of CSP(Γ ) we write ||I||
for the number of bits required to represent I. We now turn our attention
to constraint languages based on disjunctions. Let D be a set of values and
let B = {B1 , . . . , Bm } denote a finite set of relations over D, i.e. Bi ⊆ Dj
for some j ≥ 1. Let the set B ∨∞ denote the set of relations defined by dis-
junctions over B. That is, B ∨∞ contains every relation R(x1 , . . . , xp ) such that
R(x1 , . . . , xp ) if and only if Bi1 (x1 )∨· · ·∨Bit (xt ) where x1 , . . . , xt are sequences
of variables from {x1 , . . . , xp } such that the length of xj equals the arity of Bij .
We refer to Bi1 (x1 ), . . . , Bit (xt ) as the disjuncts of R. We assume, without loss
of generality, that a disjunct occurs at most once in a disjunction. We define
B ∨k , k ≥ 1, as the subset of B ∨∞ where each relation is defined by a disjunction
of length at most k. It is common, especially in qualitative temporal and spa-
tial constraint reasoning, to study a restricted variant of B ∨k when all relations
in B has the same arity p. Define B ∨= to contain every relation R such that
R(x) if and only if Bi1 (x) ∨ · · · ∨ Bit (x), where x = (x1 , . . . , xp ). For examples
of basic relations, we refer the reader to Sections 4.2 and 4.3.
We adopt a simple representation of relations in B ∨∞ : every relation R in
∨∞
B is represented by its defining disjunctive formula. Note that two objects
R, R ∈ B ∨∞ may denote the same relation. Hence, B ∨∞ is not a constraint lan-
guage in the sense of Definition 1. We avoid tedious technicalities by ignoring this
issue and view constraint languages as multisets. Given an instance I = (V, C)
of CSP(B ∨∞ ) under this representation, we let Disj(I) = {Bi1 (x1 ), . . . , Bit (xt ) |
Bi1 (x1 ) ∨ · · · ∨ Bit (xt ) ∈ C} denote the set of all disjuncts appearing in C.
We close this section by recapitulating some terminology. Let B =
{B1 , . . . , Bm } be a set of relations (over a domain D) such  that all B1 , . . . , Bm
have arity p. We say that B is jointly exhaustive (JE) if B = Dp and that B is
pairwise disjoint (PD) if Bi ∩ Bj = ∅ whenever i = j. If B is both JE and PD we
say that it is JEPD. Observe that if B1 , . . . , Bm have different arity then these
properties are clearly not relevant since the intersection between two such rela-
tions is always empty. These assumptions are common in for example qualitative
spatial and temporal reasoning, cf. [24]. Let Γ be an arbitrary set of relations
with arity p ≥ 1. We say that Γ is closed under intersection if R1 ∩ R2 ∈ Γ for
all choices of R1 , R2 ∈ Γ . Let R be an arbitrary binary relation. We define the
converse R of R such that R = {(y, x) | (x, y) ∈ R}. If Γ is a set of binary
relations, then we say that Γ is closed under converse if R ∈ Γ for all R ∈ Γ .

2.2 First-Order Definable Relations

Languages of the form B ∨∞ have a close connection with languages defined

over first-order structures admitting quantifier elimination, i.e. every first-order
definable relation can be defined by an equivalent formula without quantifiers.
We have the following lemma.

Lemma 2. Let Γ be a finite constraint language first-order definable over

a relational structure (D, R1 , . . . , Rm ) admitting quantifier elimination, where
 Upper and Lower Bounds on the Time Complexity of Infinite-Domain CSPs 187

R1 , . . . , Rm are JEPD. Then there exists a k such that (1) CSP(Γ ) is polynomial-
time reducible to CSP({R1 , . . . , Rm }∨k ) and (2) if CSP({R1 , . . . , Rm }∨k ) is solvable
in O(f (|V |)·poly(||I||)) time then CSP(Γ ) is solvable in O(f (|V |)·poly(||I||)) time.

Proof. Assume that every relation R ∈ Γ is definable through a quantifier-free

first-order formula φi over R1 , . . . , Rm . Let ψi be φi rewritten in conjunctive nor-
mal form. We need to show that every disjunction in ψi can be expressed as a dis-
junction over R1 , . . . , Rm . Clearly, if ψi only contains positive literals, then this is
trivial. Hence, assume there is at least one negative literal. Since R1 , . . . , Rm are
JEPD it is easy to see that for any negated relation in {R1 , . . . , Rm } there exists
Γ ⊆ {R1 , . . . , Rm } such that the union of Γ equals the complemented relation.
We can then reduce CSP(Γ ) to CSP({R1 , . . . , Rm }∨k ) by replacing every con-
straint by its conjunctive normal formula over R1 , . . . , Rm . This reduction can
be done in polynomial time with respect to ||I|| since each such definition can be
stored in a table of fixed size. Moreover, since this reduction does not increase the
number of variables, it follows that CSP(Γ ) is solvable in O(f (|V |) · poly(||I||))
time whenever CSP(B ∨k ) is solvable in O(f (|V |) · poly(||I||)) time. 

As we will see in Section 4, this result is useful since we can use upper
bounds for CSP(B ∨k ) to derive upper bounds for CSP(Γ ), where Γ consists
of first-order definable relations over B. There is a large number of structures
admitting quantifier elimination and interesting examples are presented in every
standard textbook on model theory, cf. Hodges [15]. A selection of problems that
are highly relevant for computer science and AI are discussed in Bodirsky [3].

3 Fundamental Algorithms
In this section we investigate the complexity of algorithms for CSP(B ∨∞ ) and
CSP(B ∨k ) based on branching on the disjuncts in constraints (Section 3.1) and
the sparsification method (Section 3.2.) Throughout this section we assume that
B is a set of basic relations such that CSP(B) is in P.

3.1 Branching on Disjuncts

Let B = {B1 , . . . , Bm } be a set of basic relations with maximum arity p ≥
1. It is easy to see that CSP(B ∨∞ ) is in NP. Assume we have an instance I
p
of CSP(B ∨∞ ) with variable set V . Such an instance contains at most 2m·|V |
distinct constraints. Each such constraint contains at most m · |V |p disjuncts so
the instance I can be solved in
m·|V |p m·|V |p
·log(m·|V |p )
O((m · |V |p )2 · poly(||I||)) = O(22 · poly(||I||))

time by enumerating all possible choices of one disjunct out of every disjunctive
constraint. The satisfiability of the resulting sets of constraints can be checked
in polynomial time due to our initial assumptions. How does such an enumer-
ative approach compare to a branching search algorithm? In the worst case,
188 P. Jonsson and V. Lagerkvist

a branching algorithm without heuristic aid will go through all of these cases so
the bound above is valid for such algorithms. Analyzing the time complexity of
branching algorithms equipped with powerful heuristics is a very different (and
presumably very difficult) problem.
Assume instead that we have an instance I of CSP(B ∨k ) with variable
set V . There are at most m · |V |p different disjuncts which leads to at most
k p i p k
i=0 (m|V | ) ≤ k · (m|V | ) distinct constraints. We can thus solve instances
k·(m|V |p )k p k
with |V | variables in O(k · poly(||I||)) = O(2k·log k·(m|V | ) · poly(||I||))
time.
Finally, let I be an instance of CSP(B ∨= ) with variable set V . It is not hard to
see that I contains at most 2m · |V |p distinct constraints, where each constraint
has length at most m. Non-deterministic guessing gives that instances of this
kind can be solved in
m
·|V |p m
·|V |p ·log m
O(m2 · poly(||I||)) = O(22 · poly(||I||))

time. This may appear to be surprisingly slow but this is mainly due to the
fact that we have not imposed any additional restrictions on the set B of basic
relations. Hence, assume that the relations in B are PD. Given two relations
R1 , R2 ∈ B ∨= , it is now clear that R1 ∩ R2 is a relation in B ∨= , i.e. B ∨=
is closed under intersection. Let I = (V, C) be an instance of CSP(B ∨= ). For
any sequence of variables (x1 , . . . , xp ), we can assume that there is at most
one constraint R(x1 , . . . , xp ) in C. This implies that we can solve CSP(B ∨= ) in
p p
O(m|V | · poly(||I||)) = O(2|V | log m · poly(||I||)) time. Combining everything so
far we obtain the following upper bounds.

Lemma 3. Let B be a set of basic relations with maximum arity p and let m =
|B|. Then
m·|V |p p
– CSP(B ∨∞ ) is solvable in O(22 ·log(m·|V | )
· poly(||I||)) time,
∨k k·log k·(m|V |p )k
– CSP(B ) is solvable in O(2 · poly(||I||)) time,
m p
– CSP(B ∨= ) is solvable in O(22 ·|V | ·log m · poly(||I||)) time, and
∨= |V |p log m
– CSP(B ) is solvable in O(2 · poly(||I||)) time if B is PD.

A bit of fine-tuning is often needed when applying highly general results like
Lemma 3 to concrete problems. For instance, Renz and Nebel [29] show that
|V |2
the RCC-8 problem can be solved in O(c 2 ) for some (unknown) c > 1. This
problem can be viewed as CSP(B ∨= ) where B contains JEPD binary relations
2
and |B| = 8. Lemma 3 implies that CSP(B ∨= ) can be solved in O(23|V | ) which
is significantly slower if c < 82 . However, it is well known that B is closed under
converse. Let I = ({x1 , . . . , xn }, C) be an instance of CSP(B ∨= ). Since B is
closed under converse, we can always assume that if R(xi , xj ) ∈ C, then i ≤ j.
|V |2 |V |2
Thus, we can solve CSP(B ∨= ) in O(m 2 ·poly(||I||)) = O(2 2 log m ·poly(||I||))
time. This figure matches the bound by Renz and Nebel better when c is small.
 Upper and Lower Bounds on the Time Complexity of Infinite-Domain CSPs 189

3.2 Sparsification
The complexity of the algorithms proposed in Section 3 is dominated by the
number of constraints. An idea for improving these running times is therefore
to reduce the number of constraints within instances. One way of accomplishing
this is by using sparsification [18]. Before presenting this method, we need a
few additional definitions. An instance of the k-Hitting Set problem consists of
a finite set U (the universe) and a collection C = {S1 , . . . , Sm } where Si ⊆ U
and |Si | ≤ k, 1 ≤ i ≤ m. A hitting set for C is a set C ⊆ U such that C ∩ Si = ∅
for each Si ∈ C. Let σ(C) be the set of all hitting sets of C. The k-Hitting Set
problem is to find a minimal size hitting set. T is a restriction of C if for each
S ∈ C there is a T ∈ T with T ⊆ S. If T is a restriction of C, then σ(T ) ⊆ σ(C).
We then have the following result1 .

Theorem 4 (Impagliazzo et al. [18]). For all ε > 0 and positive k, there
is a constant K and an algorithm that, given an instance C of k-Hitting Set on
a universe ofsize n, produces a list of t ≤ 2ε·n restrictions T1 , . . . , Tt of C so
t
that σ(C) = i=1 σ(Ti ) and so that for each Ti , |Ti | ≤ Kn. Furthermore, the
algorithm runs in time poly(n) · 2ε·n .

Lemma 5. Let B be a set of basic relations with maximum arity p and let m =
p
|B|. Then CSP(B ∨k ) is solvable in O(2(ε+K log k)·|V | ·m ·poly(||I||)) time for every
ε > 0, where K is a constant depending only on ε and k.

Proof. Let I = (V, C) be an instance of CSP(B ∨k ). We can easily reduce

CSP(B ∨k ) to k-Hitting set by letting U = Disj(I) and C be the set correspond-
ing to all disjunctions in C. Then choose some ε > 0 and let {T1 , . . . , Tt } be
the resulting sparsification. Let {T1 , . . . , Tt } be the corresponding instances of
CSP(B ∨k ). Each instance Ti contains at most K · |U | ≤ K · |V |p · m distinct
constraints, where K is a constant depending on ε and k, and can therefore be
p
solved in time O(poly(||I||) · k K·|V | ·m ) by exhaustive search à la Section 3.1.
Last, answer yes if and only if some Ti is satisfiable. This gives a total running
time of
p p p
poly(|V |p · m) · 2ε·|V | ·m
+ 2ε·|V | ·m
· k K·|V | ·m
· poly(||I||) ∈
p p p
O(2ε·|V | ·m
· 2K·|V | ·m·log k
· poly(||I||)) = O(2(ε+K log k)·|V | ·m
· poly(||I||))

since t ≤ 2ε·n . 

This procedure can be implemented using only polynomial space, just as the
enumerative methods presented in Section 3.1. This follows from the fact that the
restrictions T1 , . . . , Tt of C can be computed one after another with polynomial
delay [10, Theorem5.15]. Although this running time still might seem excessively
p k
slow observe that it is significantly more efficient than the 2k·log k·(m|V | ) algo-
∨k
rithm for CSP(B ) in Lemma 3.
1
We remark that Impagliazzo et al. [18] instead refer to the k-Hitting set problem as
the k-Set cover problem.
190 P. Jonsson and V. Lagerkvist

4 Improved Upper Bounds

In this section, we show that it is possible to obtain markedly better upper
bounds than the ones presented in Section 3. In Section 4.1 we first consider
general algorithms for CSP(B ∨∞ ) based on structure enumeration, and in Sec-
tions 4.2 and 4.3, based on the same idea, we construct even better algorithms
for equality constraint languages and temporal reasoning problems.

4.1 Structure Enumeration

We begin by presenting a general algorithm for CSP(B ∨∞ ) based on the idea of

enumerating all variable assignments that are implicitly described in instances.
As in the case of Section 3 we assume that B is a set of basic relations such that
CSP(B) is solvable in O(poly(||I||)) time.

Theorem 6. Let B be a set of basic relations with maximum arity p and let
p
m = |B|. Then CSP(B∨∞ ) is solvable in O(2m|V | · poly(||I||)) time.

Proof. Let I = (V, C) be an instance of CSP(B ∨∞ ). Let S = Disj(I) and note

that |S| ≤ m|V |p . For each subset Si of S first determine whether Si is satisfiable.
Due to the initial assumption this can be done in O(poly(||I||)) time since this
set of disjuncts can be viewed as an instance of CSP(B). Next, check whether Si
satisfies I by, for each constraint in C, determine whether at least one disjunct is
included in Si . Each such step can determined in time O(poly(||I||)) time. The
p
total time for this algorithm is therefore O(2m|V | · poly(||I||)). 

The advantage of this approach compared to the branching algorithm in

Section 3 is that enumeration of variable assignments is much less sensitive
to instances with a large number of constraints. We can speed up this result
even further by making additional assumptions on the set B. This allows us to
enumerate smaller sets of constraints than in Theorem 6.

Theorem 7. Let B be a set of basic relations with maximum arity p and let
m = |B|. Then
p
1. CSP(B∨∞ ) solvable in O(2|V | ·log m · poly(||I||))) time if B is JEPD, and
p
2. CSP(B ∨∞ ) is solvable in O(2|V | ·log(m+1) · poly(||I||))) time if B is PD.

Proof. First assume that B is JEPD and let I = (V, C) be an instance of

CSP(B ∨∞ ). Observe that every basic relation has the same arity p since B is
p
JEPD. Let F be the set of functions from |V |p to B. Clearly |F | ≤ 2|V | log m .
For every fi ∈ F let Sfi = {Bj (xj ) | xj ∈ V p , fi (xj ) = Bj }. For a set Sfi
one can then determine in O(poly(||I||)) time whether it satisfies I by, for every
constraint in C, check if at least one disjunct in every constraint is included in
Sfi . Hence, the algorithm is sound. To prove completeness, assume that g is a
satisfying assignment of I and let Sg be the set of disjuncts in C which are true
in this assignment. For every Bi (xi ) ∈ Sg define the function f as f (xi ) = Bi .
 Upper and Lower Bounds on the Time Complexity of Infinite-Domain CSPs 191

Since B is PD it cannot be the case that f (xi ) = Bi = Bj for some Bj ∈ B

distinct from Bi . Next assume that there exists xi ∈ V p but no Bi ∈ B such that
Bi (xi ) ∈ Sg . Let B = {B1 , . . . , Bm } and let f1 , . . . , fm be functions agreeing
with f for every value for which it is defined and such that fi (xi ) = Bi . Since B
is JE it holds that f satisfies I if and only if some fi satisfies I.
Next assume that B is PD but not JE. In this case we use the same construc-
tion but instead consider the set of functions F  from V p to B ∪ {Dp }. There are
p p
2|V | ·log(m+1) ) such functions, which gives the desired bound O(2|V | ·log(m+1) ·
p
poly(||I||)). The reason for adding the additional element D to the domains of
the functions is that if f ∈ F  , and if f (x) = Dp for some x ∈ V p , then this
constraint does not enforce any particular values on x. 

4.2 Equality Constraint Languages

Let E = {=, =} over some countably infinite domain D. The language E ∨∞ is a

particular case of an equality constraint language [5], i.e. sets of relations definable
through first-order formulas over the structure (D, =). Such languages are of
fundamental interest in complexity classifications for infinite domain CSPs, since
a classification of CSP problems based on first-order definable relations over some
fixed structure, always includes the classification of equality constraint language
2
CSPs. We show that the O(2|V | · poly(||I||)) time algorithm in Theorem 7 can
be improved upon quite easily. But first we need some additional machinery.
mX with n elements is a pairwise disjoint set {X1 , . . . , Xm },
A partition of a set
m ≤ n such that i=1 Xi = X. A set X with n elements has Bn partitions, where
0.792n
Bn is the n-th Bell number. Let L(n) = ln(n+1) . It is known that Bn < L(n)n [1]
and that all partitions can be enumerated in O(nBn ) time [13,32].

Theorem 8. CSP(E ∨∞ ) is solvable in O(|V |2|V |·log L(|V |) · poly(||I||)) time.

Proof. Let I = (V, C) be an instance of CSP(E ∨∞ ). For every partition S1 ∪. . .∪

Sn of V we interpret the variables in Si as being equal and having the value i, i.e.
a constraint (x = y) holds if and only if x and y belong to the same set and (x =
y) holds if and only if x and y belong to different sets. Then check in poly(||I||)
time if this partition satisfies I using the above interpretation. The complexity of
this algorithm is therefore O(|V |B|V | ·poly(||I||)) ⊆ O(|V |L(|V |)|V | ·poly(||I||)) =
O(|V |2|V |·log L(|V |) · poly(||I||)). 

2
Observe that this algorithm is much more efficient than the O(2|V | ·
poly(||I||)) algorithm in Theorem 7. It is well known that equality constraint
languages admit quantifier elimination [5]. Hence, we can use Lemma 2 to
extend Theorem 8 to cover arbitrary equality constraint languages.

Corollary 9. Let Γ be a finite set of relations first-order definable over (D, =).
Then CSP(Γ ) is solvable in O(|V |2|V |·log L(|V |) · poly(||I||)) time.
192 P. Jonsson and V. Lagerkvist

4.3 Temporal Constraint Reasoning

Let T = {<, >, =} denote the JEPD order relations on Q and recall that CSP(T )
2
is tractable [34]. Theorem 7 implies that CSP(T ∨∞ ) can be solved in O(2|V | ·log 3 ·
poly(||I||)) time. We improve this as follows.
Theorem 10. CSP(T ∨∞ ) is solvable in O(2|V | log |V | · poly(||I||)) time.
Proof. Let I = (V, C) be an instance of CSP(T ∨∞ ). Assume f : V → Q sat-
isfies this instance. It is straightforward to see that there exists some g : V →
{1, . . . , |V |} which satisfies I, too. Hence, enumerate all 2|V | log |V | functions from
V to {1, . . . , |V |} and answer yes if any of these satisfy the instance. 

It is well known that the first-order theory of (Q, <) admits quantifier elimi-
nation [6,15]. Hence, we can exploit Lemma 2 to obtain the following corollary.
Corollary 11. Let Γ be a finite temporal constraint language over (Q, <). If
CSP(Γ ) is NP-complete, then it is solvable in O(2|V | log |V | · poly(||I||)) time.
We can also obtain strong bounds for Allen’s interval algebra, which is a well-
known formalism for temporal reasoning. Here, one considers relations between
intervals of the form [x, y], where x, y ∈ R is the starting and ending point,
respectively. Let Allen be the 213 = 8192 possible unions of the set of the thirteen
relations in Table 1. For convenience we write constraints such as (p ∨ m)(x, y)
as x{p, m}y, using infix notation and omitting explicit disjunction signs. The
problem CSP(Allen) is NP-complete and all tractable fragments have been iden-
tified [22].
Given an instance I = (V, C) of CSP(Allen) we first create two fresh variables
xsi and xei for every x ∈ V , intended to represent the startpoint and endpoint of
the interval x. Then observe that a constraint x{r1 , . . . , rm }y ∈ C, where each
ri is a basic relation, can be represented as a disjunction of temporal constraints
over xs , xe , y s and y e by using the definitions of each basic relation in Table 1.
Applying Theorem 10 to the resulting instance gives the following result.

Table 1. The thirteen basic relations in Allen’s interval algebra. The endpoint relations
xs < xe and y s < y e that are valid for all relations have been omitted.

Basic relation Example Endpoints

x precedes y p xxx xe < y s
y preceded by x p−1 yyy
x meets y m xxxx xe = y s
y met-by x m−1 yyyy
x overlaps y o xxxx y s < xe ,
y overl.-by x o−1 yyyy xe < y e
x during y d xxx ys ,
y includes x d−1 yyyyyyy xe < y e
x starts y s xxx ys ,
y started by x s−1 yyyyyyy xe < y e
x finishes y f xxx xe = y e ,
y finished by x f−1 yyyyyyy ys
x equals y ≡ xxxx ys ,
yyyy xe = y e
 Upper and Lower Bounds on the Time Complexity of Infinite-Domain CSPs 193

Corollary 12. CSP(Allen) is solvable in O(22|V |(1+log |V |) · poly(||I||)) time.

Finally, we consider branching time. We define the following relations on

the set of all points in the forest containing all oriented, finite trees where the
in-degree of each node is at most one.

1. x = y if and only if there is a path from x to y and a path from y to x,

2. x < y if and only if and there is a path from x to y but no path from y to x,
3. x > y if and only if there is a path from y to x but no path from x to y,
4. x||y if and only if there is no path from x to y and no path from y to x,

These four basic relations are known as the point algebra for branching time.
We let P = {||, <, >, =}. The problem CSP(P ∨∞ ) is NP-complete and many
tractable fragments have been identified [8].

Theorem 13. CSP(P ∨∞ ) is solvable in O(2|V |+2|V | log |V | · poly(||I||)) time.

Proof. Let I = (V, C) be an instance of CSP(P ∨∞ ). We use the following algo-

rithm.

1. enumerate all directed forests over V where the in-degree of each node is at
most one,
2. for every forest F , if at least one disjunct in every constraint in C is satisfied
by F , answer yes,
3. answer no.

It is readily seen that this algorithm is sound and complete for CSP(P ∨∞ ). As
for the time complexity, recall that the number of directed labelled trees with
|V | vertices is equal to |V ||V |−2 by Cayley’s formula. These can be efficiently
enumerated by e.g. enumerating all Prüfer sequences [28] of length |V | − 2. To
enumerate all forests instead of trees, we can enumerate all labelled trees with
|V | + 1 vertices and only consider the trees where the extra vertex is connected
to all other vertices. By removing this vertex we obtain a forest with |V | vertices.
Hence, there are at most 2|V | |V ||V |−1 directed forests over V . The factor 2|V |
stems from the observation that each forest contains at most |V | edges, where
each edge has two possible directions. We then filter out the directed forests
containing a tree where the in degree of any vertex is more than one. Last, for
each forest, we enumerate all |V ||V | functions from V to the forest, and check
in poly(||I||) time whether it satisfies I. Put together this gives a complexity of
O(2|V | |V ||V |−1 |V ||V | · poly(||I||)) ⊆ O(2|V |+2|V | log |V | · poly(||I||)). 

Branching time does not admit quantifier elimination [3, Section 4.2] so
Lemma 2 is not applicable. However, there are closely connected constraint lan-
guages on trees that have this property. Examples include the triple consistency
problem with important applications in bioinformatics [7].
194 P. Jonsson and V. Lagerkvist

5 Lower Bounds
The algorithms presented in Section 4 give new upper bounds for the complexity
of CSP(B ∨∞ ). It is natural to also ask, given reasonable complexity theoretical
assumptions, how much room there is for improvement. This section is divided
into Section 5.1, where we obtain lower bounds for CSP(B ∨∞ ) and CSP(B ∨k ) for
B that are JEPD, and in Section 5.2, where we obtain lower bounds for Allen’s
interval algebra.

5.1 Lower Bounds for JEPD Languages

One of the most well-known methods for obtaining lower bounds is to exploit the
exponential-time hypothesis (ETH). The ETH states that there exists a δ > 0
such that 3-SAT is not solvable in O(2δ|V | ) time by any deterministic algorithm,
i.e. it is not solvable in subexponential time [16]. If the ETH holds, then there
is an increasing sequence s3 , s4 , . . . of reals such that k-SAT cannot be solved
in time 2sk |V | but it can be solved in 2(sk +)|V | time for arbitrary  > 0. The
strong exponential-time hypothesis (SETH) is the conjecture that the limit of the
sequence s3 , s4 , . . . equals 1, and, as a consequence, that SAT is not solvable in
time O(2δ|V | ) for any δ < 1 [16]. These conjectures have in recent years success-
fully been used for proving lower bounds of many NP-complete problems [26].
Theorem 14. Let B = {R1 , R2 , . . . , Rm } be a JEPD set of nonempty basic
relations. If the SETH holds then CSP(B ∨∞ ) cannot be solved in O(2δ|V | ) time
for any δ < 1.
Proof. If the SETH holds then SAT cannot be solved in O(2δ|V | ) time for any δ <
1. We provide a polynomial-time many-one reduction from SAT to CSP(B ∨∞ )
which only increases the number of variables by a constant — hence, if CSP(B ∨∞ )
is solvable in O(2δ|V | ) time for some δ < 1 then SAT is also solvable in O(2δ|V | )
time, contradicting the original assumption.
Let I = (V, C) be an instance of SAT, where V is a set of variables and C
a set of clauses. First observe that since m ≥ 2 and since B is JEPD, B must
be defined over a domain with two or more elements. Also note that the require-
ment that B is JEPD implies that complement of R1 (x) can be expressed as
R2 (x) ∨ . . . ∨ Rm (x). Now, let p denote the arity of the relations in B. We intro-
duce p−1 fresh variables T1 , . . . , Tp−1 and then for every clause ( 1 ∨. . .∨ k ) ∈ C
create the constraint (φ1 (x1 , T1 , . . . , Tp−1 ) ∨ . . . ∨ φk (xk , T1 , . . . , Tp−1 )), where
φi (xi , T1 , . . . , Tp−1 ) = R1 (xi , T1 , . . . , Tp−1 ) if i = xi and φi (xi , T1 , . . . , Tp−1 ) =
R2 (xi , T1 , . . . , Tp−1 ) ∨ . . . ∨ Rm (xi , T1 , . . . , Tp−1 ) if i = ¬xi . Hence, the result-
ing instance is satisfiable if and only if I is satisfiable. Since the reduction
only introduces p − 1 fresh variables it follows that SAT is solvable in time
O(2δ(|V |+p−1) ) = O(2δ|V | ). 

Even though this theorem does not rule out the possibility that CSP(B ∨k )
can be solved significantly faster for some k it is easy to see that CSP(B ∨k )
cannot be solved in subexponential time for any k ≥ 3(|B| − 1). First assume
 Upper and Lower Bounds on the Time Complexity of Infinite-Domain CSPs 195

that the ETH holds. By following the proof of Theorem 14 we can reduce 3-SAT
to CSP(B ∨3(|B|−1) ), which implies that CSP(B ∨3(|B|−1) ) cannot be solved in 2δn
time either. The bound k = 3(|B| − 1) might obviously feel a bit unsatisfactory
and one might wonder if this can be improved. We can in fact make this much
more precise by adding further restrictions to the set B. As in the case of the
equality constraint languages in Section 4.2 we let = denote the equality relation
on a given countably infinite domain.

Theorem 15. Let B = {=, R1 , . . . , Rm } be a set of binary PD, nonempty rela-

tions. If the ETH holds then CSP(B ∨k ) cannot be solved in O(2sk |V | ) time.

Proof. We prove this result by reducing k-SAT to CSP(B ∨k ) in such a way

that we at most introduce one fresh variable. Let I = (V, C) be an instance
of k-SAT, where V is a set of variables and C a set of clauses. We know that
R1 ⊆ {(a, b) | a, b ∈ D and a = b} since B is PD. Introduce one fresh variable T .
For every clause ( 1 ∨ . . . ∨ k ) ∈ C create the constraint (φ1 ∨ . . . ∨ φk ), where
φi := xj = T if i = xj and φi = R1 (xj , T ) if i = ¬xj . Let (V  , C  ) be the
resulting instance of CSP(B ∨k ). We show that (V  , C  ) has a solution if and only
if (V, C) has a solution.
Assume first that (V, C) has a solution f : V → {0, 1}. Arbitrarily choose
a tuple (a, b) ∈ R1 . We construct a solution f  : V  → {a, b} for (V  , C  ). Let
f  (T ) = b, and for all v ∈ V let f  (v) = b if f (v) = 1 and let f  (v) = a if
f (v) = 0. Arbitrarily choose a clause ( 1 ∨ . . . ∨ k ) ∈ C and assume for instance
that 1 evaluates to 1 under the solution f . If 1 = xi , then f (xi ) = 1 and
the corresponding disjunct in the corresponding disjunctive constraint in C  is
xi = T . By definition, (f  (xi ), f  (T )) = (b, b). If 1 = ¬xi , then f (xi ) = 0 and
the corresponding disjunct in the corresponding disjunctive constraint in C  is
R1 (xi , T ). By definition, (f  (xi ), f  (T )) = (a, b) and (a, b) ∈ R1 .
Assume instead that f  : V  → D is a solution to (V  , C  ), and that f  (T ) = c.
We construct a solution f : V → {0, 1} to (V, C) as follows. Arbitrarily choose a
disjunctive constraint (d1 ∨· · ·∨dk ) ∈ C  and let ( 1 ∨· · ·∨ k ) be the corresponding
clause in C  . Assume that 1 = xi . If d1 is true under f  , then let f (xi ) = 1
and, otherwise, f (xi ) = 0. If 1 = ¬xi , then do the opposite: f (xi ) = 0 if d1
is true and f (xi ) = 1 otherwise. If the function f is well-defined, then f is
obviously a solution to (V, C). We need to prove that there is no variable that
is simultaneously assigned 0 and 1. Assume this is the case. Then there is some
variable xi such that the constraints xi = T and R1 (xi , T ) are simultaneously
satisfied by f  . This is of course impossible due to the fact that R1 contains no
tuple of the form (a, a). 

If we in addition require that B is JE we obtain substantially better lower

bounds for CSP(B ∨∞ ).

Theorem 16. Let B = {=, R1 , . . . , Rm } be a set of binary JEPD relations over

a countably infinite domain. If the SETH holds then CSP(B ∨∞ ) cannot be solved
in O(c|V | ) time for any c > 1.
196 P. Jonsson and V. Lagerkvist

m First observe that the binary inequality relation = over D can be defined
Proof.
as i=1 Ri since m B is JEPD. In the the proof we therefore use = as an abbre-
viation for i=1 Ri . Let I = (V, C) be an instance of SAT with variables
V = {x1 , . . . , xn } and the set of clauses C. Let K be an integer such that
K > log c. Assume without loss of generality that n is a multiple of K. We
n n
will construct an instance of CSP(B ∨∞ ) with K + 2K = K + O(1) variables.
K
First, introduce 2 fresh variables v1 , . . . , v2K and make them different by impos-
n
ing = constraints. Second, introduce K fresh variables y1 , . . . , y Kn , and for each
n
i ∈ {1, . . . , K } impose the constraint (yi = v1 ∨ yi = v2 ∨ · · · ∨ yi = v2k ). Let
V1 , . . . , V Kn be a partition of V such that each |Vi | = K. We will represent each set
Vi of Boolean variables by one yi variable over D. To do this we will interpret each
auxiliary variable zi as a K-ary Boolean tuple. Let h : {v1 , . . . , v2K } → {0, 1}K
be an injective function which assigns a Boolean K-tuple for every variable
vi . Let g+ be a function from {1, . . . , K} to subsets of {v1 , . . . , v2K } such that
vi ∈ g(j) if and only if the j-th element in h(vi ) is equal to 1. Define g− in the
analogous way. Observe that |g+ (j)| = |g− (j)| = 2K−1 for each j ∈ {1, . . . , K}.
For the reduction, let ( i1 ∨. . .∨ in ), ij = xij or ij = ¬xij , be a clause in C.
We assume that n ≤ n since the clause contains repeated literals otherwise. For
each literal ij let Vi ⊆  V be the set of variables such that xij ∈Vi . Each literal
ij is then replaced by z∈g+ (ij ) yi = z if ij = xij , and with z∈g− (ij ) yi = z
if ij = ¬xij . This reduction can be done in polynomial time since a clause with
n literals is replaced by a disjunctive constraint with n 2K−1 disjuncts (since K
is a constant depending only on c). It follows that SAT can be solved in
n n
O(c K +O(1) · poly(||I||)) = O(2( K +O(1))·log c · poly(||I||)) = O(2δ·n · poly(||I||))

for some δ < 1, since K > log c . Thus, the SETH does not hold. 

As an illustrative use of the theorem we see that the temporal problem

CSP(T ∨∞ ) is solvable in O(2|V | log |V | · poly(||I||)) time but not in O(c|V | ) time
for any c > 1 if the SETH holds. Lower bounds can also be obtained for
the branching time problem CSP(P ∨∞ ) since there is a trivial reduction from
CSP(T )∨∞ which does not increase the number of variables: simply add a con-
straint (x < y ∨ x > y ∨ x = y) for every pair of variables in the instance.
Similarly, the equality constraint satisfaction problem CSP(E ∨∞ ) is not solvable
in O(c|V | ) time for any c > 1 either, unless the SETH fails. Hence, even though
the algorithms in Sections 4.2 and 4.3 might appear to be quite simple, there is
very little room for improvement.

5.2 Lower Bounds for Allen’s Interval Algebra

Theorems 14, 15 and 16 gives lower bounds for all the problems considered in
Sections 4.2 and 4.3 except for CSP(Allen) since unlimited use of disjunction is
not allowed in this language. It is however possible to relate the complexity of
CSP(Allen) to the Chromatic Number problem, i.e. the problem of computing
the number of colours needed to colour a given graph.
 Upper and Lower Bounds on the Time Complexity of Infinite-Domain CSPs 197

√ |V |
Theorem 17. If CSP(Allen) can be solved in O( c ) time for some c < 2,
then Chromatic Number can be solved in O((c+)|V | ) time for arbitrary  > 0.
Proof. We first present a polynomial-time many-one reduction from k-
Colourability to CSP(Allen) which introduces k fresh variables. Given an
undirected graph G = ({v1 , . . . , vn }, E), introduce the variables z1 , . . . , zk and
v1 , . . . , vn , and:

1. impose the constraints z1 {m}z2 {m} . . . {m}zk ,

2. for each vi , 1 ≤ i ≤ n, add the constraints vi {≡, s−1 }z1 , vi {p, m, f −1 , d−1 }zj
(2 ≤ j ≤ k − 1), and vi {p, m, f −1 }zk ,
3. for each (vi , vj ) ∈ E, add the constraint vi {s, s−1 }vj .

Consulting Table 1, we see that for each vi , it holds that its right endpoint
must equal the right endpoint of some zi , and its left endpoint must equal the
left endpoint of z1 . It is now obvious that the resulting instance has a solution if
and only if G is k-colourable. The result then follows since there is a polynomial-
time Turing reduction from Chromatic Number to CSP(Allen) by combining
binary search (that will evaluate log n Allen instances) with the reduction above
(recall that O(log n · cn ) ⊆ O((c + )n ) for every  > 0) . Observe that if k =√n
then the reduction introduces n fresh variables, which is where the constant c
√ |V |
in the expression O( c ) stems from. CSP(Allen). 

The exact complexity of Chromatic Number has been analysed and dis-
cussed in the literature. Björklund et al. [2] have shown that the problem is
solvable in 2|V | · poly(||I||) time. Impagliazzo and Paturi [17] poses the following
question: “Assuming SETH, can we prove a 2n−o(n) lower bound for Chromatic
√ |V |
Number?”. Hence, an O( c ), c < 2, algorithm for CSP(Allen) would also be
a major breakthrough for Chromatic Number.

6 Discussion
We have investigated several novel algorithms for solving disjunctive CSP prob-
lems, which, with respect to worst-case time complexity, are much more efficient
than e.g. backtracking algorithms without heuristics. These bounds can likely
be improved, but, due to the lower bounds in Section 5, probably not to a great
degree. Despite this, algorithms for solving infinite domain constraint satisfaction
problems are in practice used in many non-trivial applications. In light of this the
following research direction is particularly interesting: how to formally analyse
the time complexity of branching algorithms equipped with (powerful) heuristics?
In the case of finite-domain CSPs and, in particular, DPLL-like algorithms for
the k-SAT problem there are numerous results to be found in the literature,
cf. the survey by Vsemirnov et al. [35]. This is not the case for infinite-domain
CSPs, even though there is a considerable amount of empirical evidence that
infinite-domain CSPs can be efficiently solved by such algorithms, so one ought
to be optimistic about the chances of actually obtaining non-trivial bounds. Yet,
sharp formal analyses appear to be virtually nonexistent in the literature.
198 P. Jonsson and V. Lagerkvist

Another research direction is to strengthen the lower bounds in Section 5 even

further. It would be interesting to prove stronger lower bounds for CSP(B ∨k ) for
some concrete choices of B and k. As an example, consider the temporal problem
CSP(T ∨4 ). From Theorem 15 we see that CSP(T ∨4 ) is not solvable in O(2s4 |V | )
time for some s4 < log 1.6, assuming the ETH holds, since the currently best
deterministic algorithm for 4-SAT runs in O(1.6|V | ) time [12]. On the other hand,
√ |V |
if CSP(T ∨4 ) is solvable in O( c ) time for some c < 2, then Chromatic
Number can be solved in O((c + )|V | ) time for arbitrary  > 0. This can be
proven similar to the reduction in Theorem 17 but by making use of temporal
constraints instead of interval constraints. Hence, for certain choices of B and k
it might be possible to improve upon the general bounds given in Section 5.

References
1. Berend, D., Tassa, T.: Improved bounds on Bell numbers and on moments of
sums of random variables. Probability and Mathematical Statistics 30(2), 185–205
(2010)
2. Björklund, A., Husfeldt, T., Koivisto, M.: Set partitioning via inclusion-exclusion.
SIAM Journal on Computing 39(2), 546–563 (2009)
3. Bodirsky, M.: Complexity classification in infinite-domain constraint satisfaction.
Mémoire d’habilitation à diriger des recherches, Université Diderot - Paris 7.
arXiv:1201.0856 (2012)
4. Bodirsky, M., Hils, M.: Tractable set constraints. Journal of Artificial Intelligence
Research 45, 731–759 (2012)
5. Bodirsky, M., Kára, J.: The complexity of equality constraint languages. Theory
of Computing Systems 43(2), 136–158 (2008)
6. Bodirsky, M., Kára, J.: The complexity of temporal constraint satisfaction prob-
lems. Journal of the ACM 57(2), 9:1–9:41 (2010)
7. Bodirsky, M., Mueller, J.K.: The complexity of rooted phylogeny problems. Logical
Methods in Computer Science 7(4) (2011)
8. Broxvall, M., Jonsson, P.: Point algebras for temporal reasoning: Algorithms and
complexity. Artificial Intelligence 149(2), 179–220 (2003)
9. Broxvall, M., Jonsson, P., Renz, J.: Disjunctions, independence, refinements. Arti-
ficial Intelligence 140(1–2), 153–173 (2002)
10. Calabro, C.: The Exponential Complexity of Satisfiability Problems. PhD thesis,
University of California, San Diego, CA, USA (2009)
11. Cohen, D., Jeavons, P., Jonsson, P., Koubarakis, M.: Building tractable disjunctive
constraints. Journal of the ACM 47(5), 826–853 (2000)
12. Dantsin, E., Goerdt, A., Hirsch, E.A., Kannan, R., Kleinberg, J.M., Papadimitriou,
C.H., Raghavan, P., Schöning, U.: A deterministic (2 − 2/(k + 1))n algorithm for
k-SAT based on local search. Theoretical Computer Science 289(1), 69–83 (2002)
13. Djokic, B., Miyakawa, M., Sekiguchi, S., Semba, I., Stojmenovic, I.: A fast iterative
algorithm for generating set partitions. The Computer Journal 32(3), 281–282
(1989)
14. Gaspers, S.: Exponential Time Algorithms - Structures, Measures, and Bounds.
VDM (2010)
15. Hodges, W.: A Shorter Model Theory. Cambridge University Press, New York
(1997)
16. Impagliazzo, R., Paturi, R.: On the complexity of k-SAT. Journal of Computer
and System Sciences 62(2), 367–375 (2001)
 Upper and Lower Bounds on the Time Complexity of Infinite-Domain CSPs 199

17. Impagliazzo, R., Paturi, R.: Exact complexity and satisfiability. In: Gutin, G.,
Szeider, S. (eds.) IPEC 2013. LNCS, vol. 8246, pp. 1–3. Springer, heidelberg (2013)
18. Impagliazzo, R., Paturi, R., Zane, F.: Which problems have strongly exponential
complexity? Journal of Computer and System Sciences 63(4), 512–530 (2001)
19. Jonsson, P., Bäckström, C.: A unifying approach to temporal constraint reasoning.
Artificial Intelligence 102(1), 143–155 (1998)
20. Jonsson, P., Lagerkvist, V., Nordh, G., Zanuttini, B.: Complexity of SAT prob-
lems, clone theory and the exponential time hypothesis. In: Proceedings of
the 24th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA-2013),
pp. 1264–1277 (2013)
21. Kanj, I., Szeider, S.: On the subexponential time complexity of CSP. In: Proceed-
ings of the Twenty-Seventh AAAI Conference on Artificial Intelligence (AAAI-
2013) (2013)
22. Krokhin, A., Jeavons, P., Jonsson, P.: Reasoning about temporal relations: The
tractable subalgebras of Allen’s interval algebra. Journal of the ACM 50(5),
591–640 (2003)
23. Ligozat, G.: Qualitative Spatial and Temporal Reasoning. Wiley-ISTE (2011)
24. Ligozat, G., Renz, J.: What Is a qualitative calculus? a general framework. In:
Zhang, C., W. Guesgen, H., Yeap, W.-K. (eds.) PRICAI 2004. LNCS (LNAI), vol.
3157, pp. 53–64. Springer, Heidelberg (2004)
25. Liu, B., Jaffar, J.: Using constraints to model disjunctions in rule-based reasoning.
In: Proceedings of the Thirteenth National Conference on Artificial Intelligence,
AAAI 1996, Portland, Oregon, vol. 2, pp. 1248–1255 (1996)
26. Lokshtanov, D., Marx, D., Saurabh, S.: Lower bounds based on the exponential
time hypothesis. Bulletin of EATCS 3(105) (2013)
27. Marriott, K., Moulder, P., Stuckey, P.J.: Solving disjunctive constraints for interac-
tive graphical applications. In: Walsh, T. (ed.) CP 2001. LNCS, vol. 2239, p. 361.
Springer, Heidelberg (2001)
28. Prüfer, H.: Neuer beweis eines satzes über permutationen. Archiv der Mathematik
und Physik 27, 742–744 (1918)
29. Renz, J., Nebel, B.: Efficient methods for qualitative spatial reasoning. Journal of
Artificial Intelligence Research 15(1), 289–318 (2001)
30. Rossi, F., van Beek, P., Walsh, T. (eds.) Handbook of Constraint Programming.
Elsevier (2006)
31. Stergiou, K., Koubarakis, M.: Backtracking algorithms for disjunctions of temporal
constraints. Artificial Intelligence 120(1), 81–117 (2000)
32. Stojmenović, I.: An optimal algorithm for generating equivalence relations on a
linear array of processors. BIT Numerical Mathematics 30(3), 424–436 (1990)
33. Traxler, P.: The time complexity of constraint satisfaction. In: Grohe, M.,
Niedermeier, R. (eds.) IWPEC 2008. LNCS, vol. 5018, pp. 190–201. Springer,
Heidelberg (2008)
34. Vilain, M.B., Kautz, H.A.: Constraint propagation algorithms for temporal rea-
soning. In: Proceedings of the 5th National Conference on Artificial Intelligence
(AAAI 1986), pp. 377–382 (1986)
35. Vsemirnov, M., Hirsch, E., Dantsin, E., Ivanov, S.: Algorithms for SAT and upper
bounds on their complexity. Journal of Mathematical Sciences 118(2), 4948–4962
(2003)
36. Woeginger, G.: Exact algorithms for NP-hard problems: a survey. In: Juenger, M.,
Reinelt, G., Rinaldi, G. (eds.) Combinatorial Optimization - Eureka! You Shrink!.
LNCS, vol. 2570, pp. 185–207. Springer, Heidelberg (2000)
 Generalized Totalizer Encoding
for Pseudo-Boolean Constraints

Saurabh Joshi1(B) , Ruben Martins1 , and Vasco Manquinho2

1
Department of Computer Science, University of Oxford, Oxford, UK
{saurabh.joshi,ruben.martins}@cs.ox.ac.uk
2
INESC-ID/Instituto Superior Técnico, Universidade de Lisboa, Lisboa, Portugal
vasco.manquinho@inesc-id.pt

Abstract. Pseudo-Boolean constraints, also known as 0-1 Integer Lin-

ear Constraints, are used to model many real-world problems. A common
approach to solve these constraints is to encode them into a SAT for-
mula. The runtime of the SAT solver on such formula is sensitive to
the manner in which the given pseudo-Boolean constraints are encoded.
In this paper, we propose generalized Totalizer encoding (GTE), which
is an arc-consistency preserving extension of the Totalizer encoding to
pseudo-Boolean constraints. Unlike some other encodings, the number of
auxiliary variables required for GTE does not depend on the magnitudes
of the coefficients. Instead, it depends on the number of distinct combina-
tions of these coefficients. We show the superiority of GTE with respect
to other encodings when large pseudo-Boolean constraints have low num-
ber of distinct coefficients. Our experimental results also show that GTE
remains competitive even when the pseudo-Boolean constraints do not
have this characteristic.

1 Introduction

Pseudo-Boolean constraints (PBCs) or 0-1 Integer Linear constraints have been

used to model a plethora of real world problems such as computational biol-
ogy [13,24], upgradeability problems [3,15,16], resource allocation [27], schedul-
ing [26] and automated test pattern generation [22]. Due to its importance and a
plethora of applications, a lot of research has been done to efficiently solve PBCs.
One of the popular approaches is to convert PBCs into a SAT formula [7,11,21]
thus making them amenable to off-the-shelf SAT solvers. We start by formally
introducing PBC, followed by a discussion on how to convert a PBC into a SAT
formula.
A PBC is defined over a finite set of Boolean variables x1 , . . . , xn which can be
assigned a value 0 (f alse) or 1 (true). A literal li is either a Boolean variable xi
(positive literal) or its negation ¬xi (negative literal). A positive (resp. negative)
literal li is said to be assigned 1 if and only if the corresponding variable xi is
assigned 1 (resp. 0). Without a loss of generality, PBC can be defined as a linear
inequality of the following normal form:


c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 200–209, 2015.
DOI: 10.1007/978-3-319-23219-5 15
 Generalized Totalizer Encoding for Pseudo-Boolean Constraints 201

(O : o2 , o3 , o5 , o6 : 6)

(A : a2 , a3 , a5 : 5) (B : b3 , b6 : 6)

(C : l1 : 2) (D : l2 : 3) (E : l3 : 3) (F : l4 : 3)

Fig. 1. Generalized Totalizer Encoding for 2l1 + 3l2 + 3l3 + 3l4 ≤ 5


wi li ≤ k (1)

Here, wi ∈ N+ are called coefficients or weights, li are input literals and

k ∈ N+ is called the bound. Linear inequalities in other forms (e.g. other inequal-
ity, equalities or negative coefficients) can be converted into this normal form
in linear time [8]. Cardinality constraint is a special case of PBC when all the
weights have the value 1. Many different encodings have been proposed to encode
cardinality constraints[4,5,25,28]. Linear pseudo-Boolean solving (PBS) is a gen-
eralization of the SAT formulation where constraints are not restricted to clauses
and can be PBCs. A related problem to PBS is the linear pseudo-Boolean opti-
mization (PBO) problem, where all the constraints must be satisfied and the
value of a linear cost function is optimized. PBO usually requires an iterative
algorithm which solves a PBS in every iteration [11,18,19,21]. Considering that
the focus of the paper is on encodings rather than algorithms, we restrict our-
selves to the decision problem (PBS).
This paper makes the following contributions.

– We propose an arc-consistency [12] preserving extension of Totalizer encod-

ing [5] called Generalized Totalizer encoding (GTE) in Section 2.
– We compare various PBC encoding schemes that were implemented in a com-
mon framework, thus providing a fair comparison. After discussing related
work in Section 3, we show GTE as a promising encoding through its com-
petitive performance in Section 4.

2 Generalized Totalizer Encoding

The Totalizer encoding [5] is an encoding to convert cardinality constraints into a

SAT formula. In this section, the generalized Totalizer encoding (GTE) to encode
PBC into SAT is presented. GTE can be better visualized as a binary tree, as
shown in Fig. 1. With the exception of the leaves, every node is represented as
(node name : node vars : node sum). The node sum for every node represents
the maximum possible weighted sum of the subtree rooted at that node. For
any node A, a node variable aw represents a weighted sum w of the underlying
subtree. In other words, whenever the weighted sum of some of the input literals
in the subtree becomes w, aw must be set to 1. Note that for any node A, we
202 S. Joshi et al.

would need one variable corresponding to every distinct weighted sum that the
input literals under A can produce. Input literals are at the leaves, represented
as (node name : literal name : literal weight) with each of the terms being self
explanatory.
For any node P with children Q and R, to ensure that weighted sum is
propagated from Q and R to P , the following formula is built for P :
⎛ ⎞
⎛ ⎞
⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎜ q (¬qw1 ∨ ¬rw2 ∨ pw3 )⎟
⎜ w1 ∈ Q.node vars ⎟ ∧ ⎝ sw ∈ (Q.node vars ∪ R.node vars) (¬sw ∨ pw )⎠ (2)
⎝ rw2 ∈ R.node vars ⎠ w=w
w3 = w1 + w2 pw ∈ P.node vars
pw3 ∈ P.node vars

The left part of Eqn. (2) ensures that, if node Q has witnessed a weighted
sum of w1 and R has witnessed a weighted sum of w2 , then P must be considered
to have witnessed the weighted sum of w3 = w1 + w2 . The right part of Eqn. (2)
just takes care of the boundary condition where weighted sums from Q and R
are propagated to P without combining it with their siblings. This represents
that Q (resp. R) has witnessed a weighted sum of w but R (resp. Q) may not
have witnessed any positive weighted sum.
Note that node O in Fig. 1 does not have variables for the weighted sums
larger than 6. Once the weighted sum goes above the threshold of k, we represent
it with k + 1. Since all the weighted sums above k would result in the constraint
being not satisfied, it is sound to represent all such sums as k + 1. This is in some
sense a generalization of k-simplification described in [9,17]. For k-simplification,
w3 in Eqn. (2) would change to w3 = min(w1 + w2 , k + 1).
Finally, to enforce that the weighted sum does not exceed the given threshold
k, we add the following constraint at the root node O :

¬ok+1 (3)

Encoding Properties: Let AIw represent the multiset of weights of all the
input literals in the
subtree rooted at node A. For any given multiset S of weights,
let W eight(S) = e∈S e. For a given multiset S, let unique(S) denote the set
with all the multiplicity removed from S. Let |S| denote the cardinality of the
set S. Hence, the total number of node variables required at node A is:

|unique ({W eight(S)|S ⊆ AIw ∧ S = ∅})| (4)

Note that unlike some other encodings [7,14] the number of auxiliary vari-
ables required for GTE does not depend on the magnitudes of the weights.
Instead, it depends on how many unique weighted sums can be generated. Thus,
we claim that for pseudo-Boolean constraints where the distinct weighted sum
combinations are low, GTE should perform better. We corroborate our claim in
Section 4 through experiments.
Nevertheless, in the worst case, GTE can generate exponentially many auxil-
iary variables and clauses. For example, if the weights of input literals l1 , . . . , ln
 Generalized Totalizer Encoding for Pseudo-Boolean Constraints 203

are respectively 20 , . . . , 2n−1 , then every possible weighted sum combination

would be unique. In this case, GTE would generate exponentially many aux-
iliary variables. Since every variable is used in at least one clause, it will also
generate exponentially many clauses.
Though GTE does not depend on the magnitudes of the weights, one can
use the magnitude of the largest weight to categorize a class of PBCs for which
GTE is guaranteed to be of polynomial size. If there are n input literals and
the largest weight is a polynomial P (n), then GTE is guaranteed to produce a
polynomial size formula. If the largest weight is P (n), then the total number
of distinct weight combinations (Eqn. (4)) is bounded by nP (n), resulting in a
polynomial size formula.
The best case for GTE occurs when all of the weights are equal, in which
case the number of auxiliary variables and clauses is, respectively, O(n log2 n)
and O(n2 ). Notice that for this best case with k-simplification, we have O(nk)
variables and clauses, since it will behave exactly as the Totalizer encoding [5].
Note also that the generalized arc consistency (GAC) [12] property of Total-
izer encoding holds for GTE as well. GAC is a property of an encoding which
allows the solver to infer maximal possible information through propagation,
thus helping the solver to prune the search space earlier. The original proof [5]
makes an inductive argument using the left subtree and the right subtree of a
node. It makes use of the fact that, if there are q input variables set to 1 in
the left child Q and r input variables are set to 1 in the right child R, then
the encoding ensures that in the parent node P , the variable pq+r is set to 1.
Similarly, GTE ensures that if the left child Q contributes w1 to the weighted
sum (qw1 is set to 1) and the right child R contributes w2 to the weighted sum
(rw2 is set to 1), then the parent node P registers the weighted sum to be at
least w3 = w2 + w1 (pw3 is set to 1). Hence, the GAC proof still holds for GTE.

3 Related Work

The idea of encoding a PBC into a SAT formula is not new. One of the first such
encoding is described in [11,30] which uses binary adder circuit like formulation
to compute the weighted sum and then compare it against the threshold k.
This encoding creates O(n log2 k) auxiliary clauses, but it is not arc-consistent.
Another approach to encode PBCs into SAT is to use sorting networks [11].
This encoding produces O(N log22 N ) auxiliary clauses, where N is bounded by
log2 w1
+ . . . + log2 wn
. This encoding is also not arc-consistent for PBCs, but
it preserves more implications than the adder encoding, and it maintains GAC
for cardinality constraints.
The Watchdog encoding [7] scheme uses the Totalizer encoding, but in a
completely different manner than GTE. It uses multiple Totalizers, one for each
bit of the binary representation of the weights. The Watchdog encoding was the
first polynomial sized encoding that maintains GAC for PBCs and it only gen-
erates O(n3 log2 n log2 wmax ) auxiliary clauses. Recently, the Watchdog encoding
has been generalized to a more abstract framework with the Binary Merger
204 S. Joshi et al.

encoding [20]. Using a different translation of the components of the Watchdog

encoding allows the Binary Merger encoding to further reduce the number of
auxiliary clauses to O(n2 log22 n log2 wmax ). The Binary Merger is also polyno-
mial and maintains GAC.
Other encodings that maintain GAC can be exponential in the worst case
scenario, such as BDD based encodings [1,6,11]. These encodings share quite a
lot of similarity to GTE, such as GAC and independence from the magnitude of
the weight. One of the differences is that GTE always has a tree like structure
amongst auxiliary variables and input literals. However, the crucial difference
lies in the manner in which auxiliary variables are generated, and what they
represent. In BDD based approaches, an auxiliary variable Di attempts to reason
about the weighted sum of the input literals either li , . . . , ln or l1 , . . . , li . On the
other hand, an auxiliary variable aw at a node A in GTE attempts to only
reason about the weighted sum of the input literals that are descendants of A.
Therefore, two auxiliary variables in two disjoint subtrees in GTE are guaranteed
to reason about disjoint sets of input literals. We believe that such a localized
reasoning could be a cause of relatively better performance of GTE as reported in
Section 4. It is worth noting that the worst case scenario for GTE, when weights
are of the form ai , where a ≥ 2, would generate a polynomial size formula for
BDD based approaches [1,6,11].
As GTE generalizes the Totalizer encoding, the Sequential Weighted Counter
(SWC) encoding [14] generalizes sequential encoding [28] for PBCs. Like BDD
based approaches and GTE, SWC can be exponential in the worst case.

4 Implementation and Evaluation

All experiments were performed on two AMD 6276 processors (2.3 GHz) run-
ning Fedora 18 with a timeout of 1,800 seconds and a memory limit of 16
GB. Similar resource limitations were used during the last pseudo-Boolean (PB)
evaluation of 20121 . For a fair comparison, we implemented GTE (gte) in the
PBLib [29] (version 1.2) open source library which contains a plethora of encod-
ings, namely, Adder Networks (adder) [11,30], Sorting Networks (sorter) [11],
watchdog (watchdog) [7], Binary Merger (bin-merger) [20], Sequential Weighted
Counter (swc) [14], and BDDs (bdd) [1]. A new encoding in PBLib can be added
by implementing encode method of the base class Encoder. Thus, all the encod-
ings mentioned above, including GTE, only differ in how encode is implemented
while they share the rest of the whole environment. PBLib provides parsing
and normalization [11] routines for PBC and uses Minisat 2.2.0 [10] as a back-
end SAT solver. When the constraint to be encoded into CNF is a cardinality
constraint, we use the default setting of PBLib that dynamically selects a car-
dinality encoding based on the number of auxiliary clauses. When the constraint
to be encoded into CNF is a PBC, we specify one of the above encodings.
Benchmarks: Out of all 355 instances from the DEC-SMALLINT-LIN cate-
gory in the last PB evaluation of 2012 (PB’12), we only considered those 214
1
http://www.cril.univ-artois.fr/PB12/
 Generalized Totalizer Encoding for Pseudo-Boolean Constraints 205

Table 1. Characteristics of pseudo-Boolean benchmarks


Benchmark #PB #lits k max wi wi #diff wi
PB’12 164.31 32.25 27.94 12.55 167.14 6.72
pedigree 1.00 10,794.13 11,106.69 456.28 4,665,237.38 2.00

Table 2. Number of solved instances

Benchmark Result sorter swc adder watchdog bin-merger bdd gte

PB’12 SAT 72 74 73 79 79 81 81
(214) UNSAT 74 77 83 85 85 84 84
pedigree SAT 2 7 6 25 43 82 83
(172) UNSAT 0 7 6 23 35 72 75
Total SAT/UNSAT 146 165 172 212 242 319 323

instances2 that contain at least 1 PBC. We also consider an additional set of

pedigree benchmarks from computational biology [13]. These benchmarks were
originally encoded in Maximum Satisfiability (MaxSAT) and were used in the
last MaxSAT Evaluation of 20143 . Any MaxSAT problem can be converted to a
corresponding equivalent pseudo-Boolean problem [2]. We generate two pseudo-
Boolean decision problems (one satisfiable, another unsatisfiable) from the opti-
mization version of each of these benchmarks. The optimization function is trans-
formed into a PBC with the value of the bound k set to a specific value. Let
the optimum value for the optimization function be kopt . The satisfiable decision
problem uses kopt as the value for the bound k, whereas the unsatisfiable deci-
sion problem uses kopt − 1 as the value for the bound k. Out of 200 generated
instances4 , 172 had at least 1 PBC and were selected for further evaluation.
Tab. 1 shows the characteristics of the benchmarks used in this evaluation.

#PB denotes the average number of PBCs per instance. #lits, k, max wi , wi
and #diff wi denote the per constraint per instance average of input literals,
bound, the largest weight, maximum possible weighted sum and the number
of distinct weights. PB’12 benchmarks are a mix of crafted as well as industrial
benchmarks, whereas all of the pedigree benchmarks are from the same biological
problem [13]. The PB’12 benchmarks have on average several PBCs, however,
they are relatively small in magnitude. In contrast, the pedigree benchmarks
contain one large PB constraint with very large total weighted sum. pedigree
benchmarks have only two distinct values of weights, thus making them good
candidates for using GTE.
Results: Tab. 2 shows the number of instances solved using different encodings.
sorter, adder and swc perform worse than the remaining encodings for both sets of
benchmarks. The first two are not arc-consistent therefore the SAT solver is not
2
Available at http://sat.inesc-id.pt/∼ruben/benchmarks/pb12-subset.zip
3
http://www.maxsat.udl.cat/14/
4
Available at http://sat.inesc-id.pt/∼ruben/benchmarks/pedigrees.zip
206 S. Joshi et al.

·107 ·107
1.2
swc
sorter
watchdog 1 swc
3 watchdog
gte
bin-merger sorter
0.8
bdd bin-merger
adder bdd
#Vars

#Vars
2 0.6 gte
adder

0.4
1
0.2

0 0
170 180 190 200 210 0 20 40 60 80 100 120 140 160
instances instances

(a) # Variables on PB’12 benchmarks (b) # Variables on pedigree benchmarks

·107 ·107
2
swc
gte swc
adder
3 watchdog
sorter 1.5 gte
watchdog sorter
bin-merger bin-merger
#Clauses

#Clauses

bdd bdd
2 1 adder

1 0.5

0 0
150 160 170 180 190 200 210 0 20 40 60 80 100 120 140 160
instances instances

(c) # Clauses on PB’12 benchmarks (d) # Clauses on pedigree benchmarks

sorter sorter
swc swc
adder adder
1,500 bin-merger 1,500
watchdog
watchdog bin-merger
bdd bdd
seconds

seconds

gte gte
1,000 1,000

500 500

0 0
100 120 140 160 0 20 40 60 80 100 120 140 160
instances instances

(e) Runtime on PB’12 benchmarks (f) Runtime on pedigree benchmarks

Fig. 2. Cactus plots of number of variables, number of clauses and runtimes

 Generalized Totalizer Encoding for Pseudo-Boolean Constraints 207

able to infer as much information as with arc-consistent encodings. swc, though

arc-consistent, generates a large number of auxiliary variables and clauses, which
deteriorates the performance of the SAT solver.
gte provides a competitive performance to bdd, bin-merger and watchdog for
PB’12. However, only the gte and bdd encodings are able to tackle pedigree
benchmarks, which contain a large number of literals and only two different
coefficients. Unlike other encodings, gte and bdd are able to exploit the charac-
teristics of these benchmarks.
swc requires significantly large number of variables as the value of k increases,
whereas bdd and gte keep the variable explosion in check due to reuse of variables
on similar combinations (Figs. 2a and 2b). This reuse of auxiliary variables
is even more evident on pedigree benchmarks (Fig. 2b) as these benchmarks
have only two different coefficients resulting in low number of combinations. k-
simplification also helps gte in keeping the number of variables low as all the
combinations weighing more than k + 1 are mapped to k + 1.
Number of clauses required for gte is quite large as compared to some other
encodings (Figs. 2c and 2d). gte requires clauses to be generated for all the
combinations even though most of them produce the same value for the weighted
sum, thus reusing the same variable. Though bdd has an exponential worst case,
in practice it appears to generate smaller formulas (Figs. 2c and 2d).
Fig. 2e shows that gte provides a competitive performance with respect to
bin-merger, watchdog and bdd. Runtime on pedigree benchmarks as shown in
Fig. 2f establishes gte as the clear winner with bdd performing a close second.
The properties that gte and bdd share help them perform better on pedigree
benchmarks as they are not affected by large magnitude of weights in the PBCs.

5 Conclusion

Many real-world problems can be formulated using pseudo-Boolean constraints

(PBC). Given the advances in SAT technology, it becomes crucial how to encode
PBC into SAT, such that SAT solvers can efficiently solve the resulting formula.
In this paper, an arc-consistency preserving generalization of the Totalizer
encoding is proposed for encoding PBC into SAT. Although the proposed encod-
ing is exponential in the worst case, the new Generalized Totalizer encoding
(GTE) is very competitive in relation with other PBC encodings. Moreover,
experimental results show that when the number of different weights in PBC
is small, it clearly outperforms all other encodings. As a result, we believe the
impact of GTE can be extensive, since one can further extend it into incremental
settings [23].

Acknowledgments. This work is partially supported by the ERC project

280053, FCT grants AMOS (CMUP-EPB/TIC/0049/2013), POLARIS (PTDC/EIA-
CCO/123051/2010), and INESC-ID’s multiannual PIDDAC UID/CEC/50021/2013.
208 S. Joshi et al.

References
1. Abı́o, I., Nieuwenhuis, R., Oliveras, A., Rodrı́guez-Carbonell, E., Mayer-Eichberger,
V.: A New Look at BDDs for Pseudo-Boolean Constraints. Journal of Artificial
Intelligence Research 45, 443–480 (2012)
2. Aloul, F., Ramani, A., Markov, I., Sakallah, K.: Generic ILP versus specialized
0–1 ILP: an update. In: International Conference on Computer-Aided Design,
pp. 450–457. IEEE Press (2002)
3. Argelich, J., Berre, D.L., Lynce, I., Marques-Silva, J., Rapicault, P.: Solving linux
upgradeability problems using boolean optimization. In: Workshop on Logics for
Component Configuration, pp. 11–22 (2010)
4. Ası́n, R., Nieuwenhuis, R., Oliveras, A., Rodrı́guez-Carbonell, E.: Cardinality Net-
works: a theoretical and empirical study. Constraints 16(2), 195–221 (2011)
5. Bailleux, O., Boufkhad, Y.: Efficient CNF encoding of boolean cardinality con-
straints. In: Rossi, F. (ed.) CP 2003. LNCS, vol. 2833, pp. 108–122. Springer,
Heidelberg (2003)
6. Bailleux, O., Boufkhad, Y., Roussel, O.: A Translation of Pseudo Boolean Con-
straints to SAT. Journal on Satisfiability, Boolean Modeling and Computation
2(1–4), 191–200 (2006)
7. Bailleux, O., Boufkhad, Y., Roussel, O.: New encodings of pseudo-boolean con-
straints into CNF. In: Kullmann, O. (ed.) SAT 2009. LNCS, vol. 5584, pp. 181–194.
Springer, Heidelberg (2009)
8. Barth, P.: Logic-based 0–1 Constraint Programming. Kluwer Academic Publishers
(1996)
9. Büttner, M., Rintanen, J.: Satisfiability planning with constraints on the number
of actions. In: International Conference on Automated Planning and Scheduling,
pp. 292–299. AAAI Press (2005)
10. Eén, N., Sörensson, N.: An extensible SAT-solver. In: Giunchiglia, E., Tacchella, A.
(eds.) SAT 2003. LNCS, vol. 2919, pp. 502–518. Springer, Heidelberg (2004)
11. Eén, N., Sörensson, N.: Translating Pseudo-Boolean Constraints into SAT. Journal
on Satisfiability, Boolean Modeling and Computation 2(1–4), 1–26 (2006)
12. Gent, I.P.: Arc consistency in SAT. In: European Conference on Artificial Intelli-
gence, pp. 121–125. IOS Press (2002)
13. Graça, A., Lynce, I., Marques-Silva, J., Oliveira, A.L.: Efficient and accurate haplo-
type inference by combining parsimony and pedigree information. In: Horimoto, K.,
Nakatsui, M., Popov, N. (eds.) ANB 2010. LNCS, vol. 6479, pp. 38–56. Springer,
Heidelberg (2012)
14. Hölldobler, S., Manthey, N., Steinke, P.: A compact encoding of pseudo-boolean
constraints into SAT. In: Glimm, B., Krüger, A. (eds.) KI 2012. LNCS, vol. 7526,
pp. 107–118. Springer, Heidelberg (2012)
15. Ignatiev, A., Janota, M., Marques-Silva, J.: Towards efficient optimization in pack-
age management systems. In: International Conference on Software Engineering,
pp. 745–755. ACM (2014)
16. Janota, M., Lynce, I., Manquinho, V.M., Marques-Silva, J.: PackUp: Tools for
Package Upgradability Solving. JSAT 8(1/2), 89–94 (2012)
17. Koshimura, M., Zhang, T., Fujita, H., Hasegawa, R.: QMaxSAT: A Partial Max-
SAT Solver. Journal on Satisfiability, Boolean Modeling and Computation 8,
95–100 (2012)
18. Le Berre, D., Parrain, A.: The Sat4j library, release 2.2. Journal on Satisfiability,
Boolean Modeling and Computation 7(2–3), 59–64 (2010)
 Generalized Totalizer Encoding for Pseudo-Boolean Constraints 209

19. Manquinho, V., Marques-Silva, J., Planes, J.: Algorithms for weighted boolean
optimization. In: Kullmann, O. (ed.) SAT 2009. LNCS, vol. 5584, pp. 495–508.
Springer, Heidelberg (2009)
20. Manthey, N., Philipp, T., Steinke, P.: A more compact translation of pseudo-
boolean constraints into cnf such that generalized arc consistency is maintained.
In: Lutz, C., Thielscher, M. (eds.) KI 2014. LNCS, vol. 8736, pp. 123–134. Springer,
Heidelberg (2014)
21. Manthey, N., Steinke, P.: npSolver - A SAT based solver for optimization problems.
In: Pragmatics of SAT (2012)
22. Marques-Silva, J.: Integer programming models for optimization problems in test
generation. In: Asia and South Pacific Design Automation, pp. 481–487. IEEE
Press (1998)
23. Martins, R., Joshi, S., Manquinho, V., Lynce, I.: Incremental cardinality con-
straints for MaxSAT. In: O’Sullivan, B. (ed.) CP 2014. LNCS, vol. 8656,
pp. 531–548. Springer, Heidelberg (2014)
24. Miranda, M., Lynce, I., Manquinho, V.: Inferring phylogenetic trees using pseudo-
Boolean optimization. AI Communications 27(3), 229–243 (2014)
25. Ogawa, T., Liu, Y., Hasegawa, R., Koshimura, M., Fujita, H.: Modulo based CNF
encoding of cardinality constraints and its application to MaxSAT solvers. In:
International Conference on Tools with Artificial Intelligence, pp. 9–17. IEEE Press
(2013)
26. Prestwich, S., Quirke, C.: Boolean and pseudo-boolean models for scheduling. In:
International Workshop on Modelling and Reformulating Constraint Satisfaction
Problems (2003)
27. Ribas, B.C., Suguimoto, R.M., Montaño, R.A.N.R., Silva, F., de Bona, L., Castilho,
M.A.: On modelling virtual machine consolidation to pseudo-boolean constraints.
In: Pavón, J., Duque-Méndez, N.D., Fuentes-Fernández, R. (eds.) IBERAMIA
2012. LNCS, vol. 7637, pp. 361–370. Springer, Heidelberg (2012)
28. Sinz, C.: Towards an optimal CNF encoding of boolean cardinality constraints. In:
van Beek, P. (ed.) CP 2005. LNCS, vol. 3709, pp. 827–831. Springer, Heidelberg
(2005)
29. Steinke, P., Manthey, N.: PBLib-A C++ Toolkit for Encoding Pseudo-Boolean
Constraints into CNF. Tech. rep., Technische Universität Dresden (2014). http://
tools.computational-logic.org/content/pblib.php
30. Warners, J.: A Linear-Time Transformation of Linear Inequalities into Conjunctive
Normal Form. Information Processing Letters 68(2), 63–69 (1998)
 Smaller Selection Networks for Cardinality
Constraints Encoding

Michal Karpiński(B) and Marek Piotrów

Institute of Computer Science, University of Wroclaw,

Joliot-Curie 15, 50-383 WroclAw, Poland
{karp,mpi}@cs.uni.wroc.pl

Abstract. Selection comparator networks have been studied for many

years. Recently, they have been successfully applied to encode cardinal-
ity constraints for SAT-solvers. To decrease the size of generated formula
there is a need for constructions of selection networks that can be effi-
ciently generated and produce networks of small sizes for the practical
range of their two parameters: n – the number of inputs (Boolean vari-
ables) and k – the number of selected items (a cardinality bound). In
this paper we give and analyze a new construction of smaller selection
networks that are based on the pairwise selection networks introduced
by Codish and Zazon-Ivry. We prove also that standard encodings of
cardinality constraints with selection networks preserve arc-consistency.

1 Introduction
Comparator networks are probably the simplest data-oblivious model for sorting-
related algorithms. The most popular construction is due to Batcher [4] and
it’s called odd-even sorting network. For all practical values, this is the best
known sorting network. However, in 1992 Parberry [10] introduced the serious
competitor to Batcher’s construction, called pairwise sorting network. In context
of sorting, pairwise network is not better than odd-even network, in fact it has
been proven that they have exactly the same size and depth. As Parberry said
himself: ”It is the first sorting network to be competitive with the odd-even sort
for all values of n“. There is a more sophisticated relation between both types
of network and their close resemblance. For overview of sorting networks, see
Knuth [8] or Parberry [9].
In recent years new applications for sorting networks have been found, for
example in encoding of pseudo Boolean constraints and cardinality constraints
for SAT-solvers. Cardinality constraints take the form x1 + x2 + . . . + xn ∼ k,
where x1 , x2 , . . . , xn are Boolean variables, k is a natural number, and ∼ is a
relation from the set {=, <, ≤, >, ≥}. Cardinality constraints are used in many
applications, the significant ones worth mentioning arise in SAT-solvers. Using
cardinality constraints with cooperation of SAT-solvers we can handle many
practical problems, for example, cumulative scheduling [11] and timetabling [3],
that are proven to be hard. Works of Ası́n et al. [1,2] describe how to use odd-
even sorting network to encode cardinality constraints into Boolean formulas. In
[7] authors do the same with pseudo Boolean constraints.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 210–225, 2015.
DOI: 10.1007/978-3-319-23219-5 16
 Smaller Selection Networks for Cardinality Constraints Encoding 211

It has already been observed that using selection networks instead of sort-
ing networks is more efficient for the encoding of cardinality constraints. Codish
and Zazon-Ivry [6] introduced pairwise cardinality networks, which are networks
derived from pairwise sorting networks that express cardinality constraints. Two
years later, same authors [5] reformulated the definition of pairwise selection net-
works and proved that their sizes are never worse than the sizes of corresponding
odd-even selection networks. To show the difference they plotted it for selected
values of n and k.
In this paper we give a new construction of smaller selection networks that
are based on the pairwise selection ones and we prove that the construction is
correct. We also estimate the size of our networks and compute the difference in
sizes between our selection networks and the corresponding pairwise ones. The
difference can be as big as n log n/2 for k = n/2. Finally, we analyze the standard
3(6)-clause encoding of a comparator and prove that such CNF encoding of
any selection network preserves arc-consistency with respect to a corresponding
cardinality constraint.
The rest of the paper is organized in the following way: in Section 2 we give
definitions and notations used in this paper. In Section 3 we recall the definition
of pairwise selection networks and define auxiliary bitonic selection networks
that we will use to estimate the sizes of our networks. In Section 4 we present
the construction of our selection networks and prove its correctness. In Section
5 we analyze the sizes of the networks and, finally, in Section 6 we examine the
arc-consistency of selection networks.

2 Preliminaries
In this section we will introduce definitions and notations used in the rest of the
paper.
Definition 1 (Input Sequence). Input sequence of length n is a sequence of
natural numbers x̄ = x1 , . . . , xn , where xi ∈ N (for all i = 1..n). We say that
x̄ ∈ Nn is sorted if xi ≥ xi+1 (for each i = 1..n − 1). Given x̄ = x1 , . . . , xn ,
ȳ = y1 , . . . , yn  we define concatenation as x̄ :: ȳ = x1 , . . . , xn , y1 , . . . , yn .
We will use the following functions from Nn to Nn/2 :

lef t(x̄) = x1 , . . . , xn/2 , right(x̄) = xn/2+1 , . . . , xn 

Let n, m ∈ N. We define a relation ’’ on Nn × Nm . Let x̄ = x1 , . . . , xn  and
ȳ = y1 , . . . , ym , then:

x̄  ȳ ⇐⇒ ∀i∈{1,...,n} ∀j∈{1,...,m} xi ≥ yj
Definition 2 (Comparator). Let x̄ ∈ Nn and let i, j ∈ N, where 1 ≤ i < j ≤ n.
A comparator is a function ci,j defined as:

ci,j (x̄) = ȳ ⇐⇒ yi = max{xi , xj } ∧ yj = min{xi , xj } ∧ ∀k=i,j xk = yk

212 M. Karpiński and M. Piotrów

Definition 3 (Comparator Network). We say that f n : Nn → Nn is a com-

parator network of order n, if it can be represented as the composition of finite
number of comparators, namely, f n = ci1 ,j1 ◦ · · · ◦ cik ,jk . The size of comparator
network (number of comparators) is denoted by |f n |. Comparator network of size
0 is denoted by idn .

Definition 4 (V-shaped and Bitonic Sequences). A sequence x̄ ∈ Nn is

called v-shaped if x1 ≥ . . . ≥ xi ≤ . . . ≤ xn for some i, where 1 ≤ i ≤ n. A
v-shaped sequence or its circular shift is traditionally called bitonic.

Definition 5 (Sorting Network). A comparator network f n is a sorting net-

work, if for each x̄ ∈ Nn , f n (x̄) is sorted.

Two types of sorting networks are of interest to us: odd-even and pairwise.
Based on their ideas, Knuth [8] (for odd-even network) and Codish and Zazon-
Ivry [5,6] (for pairwise network) showed how to transform them into selection
networks (we name them oe selkn and pw selkn respectively).

Definition 6 (Top k Sorted Sequence). A sequence x̄ ∈ Nn is top k sorted,

with k ≤ n, if x1 , . . . , xk  is sorted and x1 , . . . , xk   xk+1 , . . . , xn .

Definition 7 (Selection Network). A comparator network fkn (where k ≤ n)

is a selection network, if for each x̄ ∈ Nn , fkn (x̄) is top k sorted.

To simplify the presentation we assume that n and k are powers of 2.

A clause is a disjunction of literals (Boolean variables x or their negation
¬x). A CNF formula is a conjunction of one or more clauses.
A unit propagation (UP) is a process, that for given CNF formula, clauses are
sought in which all literals but one are false (say l) and l is undefined (initially
only clauses of size one satisfy this condition). This literal l is set to true and
the process is iterated until reaching a fix point.
Cardinality constraints are of the form x1 + . . . + xn ∼ k, where k ∈ N
and ∼ belongs to {<, ≤, =, ≥, >}. We will focus on cardinality constraints with
less-than relation, i.e. x1 + . . . + xn < k. An encoding (a CNF formula) of such
constraint preserves arc-consistency, if as soon as k − 1 variables among the xi ’s
become true, the unit propagation sets all other xi ’s to false.
In [7] authors are using sorting networks for an encoding of cardinality con-
straints, where inputs and outputs of a comparator are Boolean variables and
comparators are encoded as a CNF formula. In addition, the k-th greatest out-
put variable yk of the network is forced to be 0 by adding ¬yk as a clause to
the formula that encodes x1 + . . . + xn < k. They showed that the encoding
preserves arc-consistency.
A single comparator can be translated to a CNF formula in the following way:
let a and b be variables denoting upper and lower inputs of the comparator, and
c and d be variables denoting upper and lower outputs of a comparator, then:

f comp(a, b, c, d) ⇔ (c ⇔ a ∨ b) ∧ (d ⇔ a ∧ b)
 Smaller Selection Networks for Cardinality Constraints Encoding 213

is the full encoding of a comparator. Notice that it consists of 6 clauses. Let f

be a comparator network. Full encoding φ of f is a conjunction of full encoding
of every comparator of f .
In [2] authors observe that in case of ∼ being < or ≤, it is sufficient to use
only 3 clauses for a single comparator, namely:

hcomp(a, b, c, d) ⇔ (a ⇒ c) ∧ (b ⇒ c) ∧ (a ∧ b ⇒ d) (1)
        
(c1) (c2) (c3)

We call it: a half encoding. In [2] it is used to translate an odd-even sorting

network to an encoding that preserves arc-consistency. We show a more general
result (with respect to both [7] and [2]), that the half encoding of any selection
network preserves arc-consistency for the “<” and “≤” relations. Similar results
can be proved for the “=” relation using the full encoding of comparators and
for the “>” or “≥” relations using an encoding symmetric to hcomp(a, b, c, d),
namely: (d ⇒ a) ∧ (d ⇒ b) ∧ (c ⇒ a ∨ b).

3 Pairwise and Bitonic Selection Networks

Now we present two constructions for selection networks. First, we recall the def-
inition of pairwise selection networks by Codish and Zazon-Ivry [5,6]. Secondly,
we give the auxiliary construction of a bitonic selection network bit selkn , that
we will use to estimate the sizes of our improved pairwise selection network in
Section 5.

Definition 8 (Domination). x̄ ∈ Nn dominates ȳ ∈ Nn if xi ≥ yi (for i =

1..n).

Definition 9 (Splitter). A comparator network f n is a splitter if for any

sequence x̄ ∈ Nn , if ȳ = f n (x̄), then lef t(ȳ) dominates right(ȳ).

Observation 1. We can construct splitter splitn by joining inputs i, n/2 + i,
for i = 1..n/2, with a comparator. The size of a splitter is |splitn | = n/2.

Lemma 1. If b̄ ∈ Nn is bitonic and ȳ = splitn (b̄), then lef t(ȳ) and right(ȳ)
are bitonic and lef t(ȳ)  right(ȳ).

Proof. See Appendix B of [4].

The construction of a pairwise selection network is presented in Network 1.

Notice that since a splitter is used as the third step, in the recursive calls we need
to select k top elements from the first half of ȳ, but only k/2 top elements from
the second half. The reason is that rk/2+1 cannot be one of the first k largest
elements of ¯l :: r̄. First, rk/2+1 is not greater than any one of r1 , . . . , rk/2  (by
the definition of top k sorted sequence), and second, l1 , . . . , lk/2  dominates
r1 , . . . , rk/2 , so rk/2+1 is not greater than any one of l1 , . . . , lk/2 . Based on
these arguments we can make the following observation:
214 M. Karpiński and M. Piotrów

Network 1. pw selkn ; see [5,6]

Input: any x̄ ∈ Nn
1: if k = 1 then return maxn (x̄)
2: if k = n then return oe sortn (x̄)
3: ȳ ← split(x̄)
n/2 n/2
4: ¯
l ← pw selk (lef t(ȳ)) and r̄ ← pw selk/2 (right(ȳ))
n ¯
5: return pw mergek (l :: r̄)

Observation 2. If ¯l ∈ Nn/2 is top k sorted, r̄ ∈ Nn/2 is top k/2 sorted and

l1 , . . . , lk/2  dominates r1 , . . . , rk/2 , then k largest elements of ¯l :: r̄ are in
l1 , . . . , lk  :: r1 , . . . , rk/2 .
The last step of Network 1 merges k top elements from ¯l and k/2 top elements
from r̄ with so called pairwise merger. We will omit the construction of this
merger, because it is not relevant to our work. We would only like to note, that
its size is: |pw mergenk | = k log k − k + 1. Construction of the merger as well as
the detailed proof of correctness of network pw selkn can be found in Section 6
of [5].

Definition 10 (Bitonic Splitter). A comparator network f n is a bitonic split-

ter if for any two sorted sequences x̄, ȳ ∈ Nn/2 , if z̄ = f n (x̄ :: ȳ), then (1)
lef t(z̄)  right(z̄) and (2) lef t(z̄) and right(z̄) are bitonic.

Observation 3. We can construct bitonic splitter bit splitn by joining inputs

i, n − i + 1, for i = 1..n/2, with a comparator. The size of a bitonic splitter is
|bit splitn | = n/2.
We now present the procedure for construction of the bitonic selection net-
work. We use the odd-even sorting network oe sort and the network bit merge
(also by Batcher [4]) for sorting bitonic sequences as black-boxes. As a reminder:
bit mergen consists of two steps, first we use ȳ = splitn (x̄), then recursively com-
pute bit mergen/2 for lef t(ȳ) and right(ȳ) (base case, n = 2, consists of a single
comparator). The size of this network is: |bit mergen | = n log n/2. A bitonic
selection network bit selkn is constructed by the procedure Network 2.

Theorem 1. A comparator network bit selkn constructed by the procedure Net-

work 2 is a selection network.

Proof. Let x̄ ∈ Nn be the input to bit selkn . After step one we get sorted sequences
B1 , . . . , Bl , where l = n/k. Let lm be the value of l after m iterations. Let
B1m , . . . , Blmm
be the blocks after m iterations. We will prove by induction that:
P (m): if B1 , . . . , Bl are sorted contain k largest elements of x̄, then after m-
th iteration of the second step: lm = l/2m , B1m , . . . , Blmm
are sorted and contain
k largest elements of x̄.
If m = 0, then l0 = l, so P (0) holds. We show that ∀m≥0 (P (m) ⇒ P (m + 1)).
Consider (m+1)-th iteration of step two. By the induction hypothesis lm = l/2m ,
 Smaller Selection Networks for Cardinality Constraints Encoding 215

Network 2. bit selkn

Input: any x̄ ∈ Nn
1: l ← n/k and partition input x̄ into l consecutive blocks, each of size k, then sort
each block with oe sortk , obtaining sorted blocks B1 , . . . , Bl
2: while l > 1 do
3: Collect blocks into pairs B1 , B2 , . . . , Bl−1 , Bl 
4: for all i ∈ {1, 3, . . . , l − 1} do y¯i ← bit split2k (Bi :: Bi+1 )

5: for all i ∈ {1, 3, . . . , l − 1} do Bi/2 ← bit mergek (lef t(y¯i ))
6: l ← l/2 and relabel Bi to Bi , for 1 ≤ i ≤ l

B1m , . . . , Blm
m
are sorted and contain k largest elements of x̄. We will show that
(m + 1)-th iteration does not remove any element from k largest elements of x̄.
To see this, notice that if y¯i = bit split2k (Bim :: Bi+1 m
) (for i ∈ {1, 3, . . . , lm − 1}),
then lef t(y¯i )  right(y¯i ) and that lef t(y¯i ) is bitonic (by Definition 10). Because
of those two facts, right(y¯i ) is discarded and lef t(y¯i ) is sorted using bit mergek .
After this, lm+1 = lm /2 = l/2m+1 and blocks B1m+1 , . . . , Blm+1 m+1
are sorted. Thus
P (m + 1) is true.
Since l = n/k, then by P (m) we see that the second step will terminate after
m = log nk iterations and that B1 is sorted and contains k largest elements of
x̄.

The size of bitonic selection network is:

n n 
|bit selkn | =|oe sortk | + − 1 (|bit split2k | + |bit mergek |)
k k
1 1 1 n
= n log k + n log k + 2n − k log k − k −
2
(2)
4 4 2 k

4 New Smaller Selection Networks

As mentioned in the previous section, only the first k/2 elements from the second
half of the input are relevant when we get to the merging step in pw selkn . We
will exploit this fact to create a new, smaller merger. We will use the concept
of bitonic sequences, therefore the new merger will be called pw bit mergenk and
the new selection network: pw bit selkn . The network pw bit selkn is generated by
substituting the last step of pw selkn with pw bit mergenk . The new merger is
constructed by the procedure Network 3.

Theorem 2. The output of Network 3 consists of sorted k largest elements from

input ¯l :: r̄, assuming that ¯l ∈ Nn/2 is top k sorted and r̄ ∈ Nn/2 is top k/2 sorted
and l1 , . . . , lk/2  dominates r1 , . . . , rk/2 .

Proof. We have to prove two things: (1) b̄ is bitonic and (2) b̄ consists of k largest
elements from ¯l :: r̄.
216 M. Karpiński and M. Piotrów

Network 3. pw bit mergenk

Input: ¯ l :: r̄, where ¯ l ∈ Nn/2 is top k sorted and r̄ ∈ Nn/2 is top k/2 sorted and
l1 , . . . , lk/2  dominates r1 , . . . , rk/2 
1: ȳ ← bit splitk (lk/2+1 , . . . , lk , r1 , . . . , rk/2 ) and b̄ ← l1 , . . . , lk/2  :: y1 , . . . , yk/2 
2: return bit mergek (b̄)

(1) Let j be the last index in the sequence k/2 + 1, . . . , k, for which lj >
rk−j+1 . If such j does not exist, then y1 , . . . , yk/2  is nondecreasing, hence b̄
is bitonic (nondecreasing). Assume that j exists, then yj−k/2+1 , . . . , yk/2  is
nondecreasing and y1 , . . . , yk−j  is nonincreasing. Adding the fact that lk/2 ≥
lk/2+1 = y1 proves, that b̄ is bitonic (v-shaped).
(2) By Observation 2, it is sufficient to prove that b̄  yk/2+1 , . . . , yk .
Since ∀k/2<j≤k lk/2 ≥ lj ≥ min{lj , rk−j+1 } = y3k/2−j+1 , then l1 , . . . , lk/2  
yk/2+1 , . . . , yk  and by Definition 10: y1 , . . . , yk/2   yk/2+1 , . . . , yk . Therefore
b̄ consists of k largest elements from ¯l :: r̄.
The bitonic merger in step 2 receives a bitonic sequence, so it outputs a
sorted sequence, which completes the proof.
The first step of improved pairwise merger is illustrated in Figure 1. We
use k/2 comparators in the first step and k log k/2 comparators in the second
step. We get a merger of size k log k/2 + k/2, which is better than the previous
approach. In the following it is shown that we can do even better and eliminate
k/2 term.

Fig. 1. Making the bitonic sequence. Arrows on the right picture show directions of
inequalities. Sequence on the right is v-shaped s-dominating at point i.

The main observation is that the result of the first step of pw bit merge oper-
ation: b1 , b2 , . . . , bk  is not only bitonic, but what we call v-shaped s-dominating.

Definition 11 (S-domination). A sequence b̄ = b1 , b2 , . . . , bk  is s-dominating

if ∀1≤j≤k/2 bj ≥ bk−j+1 .

Lemma 2. If b̄ = b1 , b2 , . . . , bk  is v-shaped and s-dominating, then b̄ is non-

increasing or ∃k/2<i<k bi < bi+1 .
 Smaller Selection Networks for Cardinality Constraints Encoding 217

Proof. Assume that b̄ is not nonincreasing. Then ∃1≤j<k bj < bj+1 . Assume that
j ≤ k/2. Since b̄ is v-shaped, bj+1 must be in nondecreasing part of b̄. If follows
that bj < bj+1 ≤ . . . ≤ bk/2 ≤ . . . ≤ bk−j+1 . That means that bj < bk−j+1 . On
the other hand, b̄ is s-dominating, thus bj ≥ bk−j+1 – a contradiction.

We will say that a sequence b̄ is v-shaped s-dominating at point i if i is the

smallest index greater than k/2 such that bi < bi+1 or i = k for a nonincreasing
sequence.

Lemma 3. Let b̄ = b1 , b2 , . . . , bk  be v-shaped s-dominating at point i, then

b1 , . . . , bk/4   bk/2+1 , . . . , b3k/4 .

Proof. If b̄ is nonincreasing, then the lemma holds. Otherwise, from Lemma 2:

k/2 < i < k. If i > 3k/4, then by Definition 4: b1 ≥ . . . ≥ b3k/4 ≥ . . . ≥ bi ,
so the lemma holds. If k/2 < i ≤ 3k/4, then by Definition 4: b1 ≥ . . . ≥ bi , so
b1 , . . . , bk/4   bk/2+1 , . . . , bi . Since bi < bi+1 ≤ . . . ≤ b3k/4 , it suffices to prove
that bk/4 ≥ b3k/4 . By Definition 11 and 4: bk/4 ≥ b3k/4+1 ≥ b3k/4 .

Definition 12 (Half Splitter). A half splitter is a comparator network con-

structed by comparing inputs k/4 + 1, 3k/4 + 1, . . . , k/2, k (normal splitter
with first k/4 comparators removed). We will call it half splitk .

Lemma 4. If b̄ is v-shaped s-dominating, then half splitk (b̄) = splitk (b̄).

Proof. Directly from Lemma 3.

Lemma 5. Let b̄ be v-shaped s-dominating. The following statements are true:

(1) lef t(half splitk (b̄)) is v-shaped s-dominating; (2) right(half splitk (b̄)) is
bitonic; (3) lef t(half splitk (b̄))  right(half splitk (b̄)).

Proof. (1) Let ȳ = lef t(half splitk (b̄)). First we show that ȳ is v-shaped. If ȳ is
nonincreasing, then it is v-shaped. Otherwise, let j be the first index from the
range {1, . . . , k/2}, where yj−1 < yj . Since yj = max{bj , bj+k/2 } and yj−1 ≥
bj−1 ≥ bj , thus bj < bj+k/2 . Since b̄ is v-shaped, element bj+k/2 must be in
nondecreasing part of b̄. It follows that bj ≥ . . . ≥ bk/2 and bj+k/2 ≤ . . . ≤ bk .
From this we can see that ∀j≤j  ≤k/2 yj  = max{bj  , bj  +k/2 } = bj  +k/2 , so yj ≤
. . . ≤ yk/2 . Therefore ȳ is v-shaped.
Next we show that ȳ is s-dominating. Consider any j, where 1 ≤ j ≤ k/4.
By Definition 4 and 11: bj ≥ bk/2−j+1 and bj ≥ bk−j+1 , therefore yj = bj ≥
max{bk/2−j+1 , bk−j+1 } = yk/2−j+1 , thus proving that ȳ is s-dominating. Con-
cluding: ȳ is v-shaped s-dominating.
(2) Let z̄ = right(half splitk (b̄)). By Lemma 4: z̄ = right(splitk (b̄)). We
know that b̄ is a special case of bitonic sequence, therefore using Lemma 1 we
get that z̄ is bitonic.
(3) Let w̄ = half splitk (b̄). By Lemma 4: w̄ = splitk (b̄). We know that b̄ is
a special case of bitonic sequence, therefore using Lemma 1 we get lef t(w̄) 
right(w̄).
218 M. Karpiński and M. Piotrów

Network 4. pw hbit mergenk

Input: ¯ l :: r̄, where ¯ l ∈ Nn/2 is top k sorted, r̄ ∈ Nn/2 is top k/2 sorted and
l1 , . . . , lk/2  dominates r1 , . . . , rk/2 
1: ȳ ← bit splitk (lk/2+1 , . . . , lk , r1 , . . . , rk/2 ) and b̄ ← l1 , . . . , lk/2  :: y1 , . . . , yk/2 
2: return half bit mergek (b̄), where
3: function half bit mergek (b̄)
4: if k = 2 then return (b1 , b2 )
5: b¯ ← half split(b1 , . . . , bk )
6: return half bit mergek/2 (lef t(b¯ )) :: bit mergek/2 (right(b¯ ))

Using both half split and Batcher’s bit merge and successively applying
Lemma 5 to the resulting v-shaped s-dominating half of the output, we have
all the tools needed to construct the improved pairwise merger pw hbit mergenk
(Network 4) using half splitters and then to prove that the construction is correct.
Theorem 3. The output of Network 4 consists of sorted k largest elements from
input ¯l :: r̄, assuming that ¯l ∈ Nn/2 is top k sorted and r̄ ∈ Nn/2 is top k/2
sorted and l1 , . . . , lk/2  dominates r1 , . . . , rk/2 . Moreover, |pw hbit mergenk | =
k log k/2.
Proof. Since step 1 in Network 4 is the same as in Network 3, we can reuse
the proof of Theorem 2 to deduce, that b̄ is v-shaped and contains k largest
elements from ¯l :: r̄. Also, since ∀1≤j≤k/2 lj ≥ lk−j+1 and lj ≥ rj , then bj = lj ≥
max{lk−j+1 , rj } = bk−j+1 , so b̄ is s-dominating.
We prove by the induction on k, that if b̄ is v-shaped s-dominating, then the
sequence half bit mergek (b̄) is sorted. For the base case, consider k = 2 and a v-
shaped s-dominating sequence b1 , b2 . By Definition 11 this sequence is already
sorted and we are done. For the induction step, consider b¯ = half splitk (b̄). By
Lemma 5 we get that lef t(b¯ ) is v-shaped s-dominating and right(b¯ ) is bitonic.
Using the induction hypothesis we sort lef t(b¯ ) and using bitonic merger we
sort right(b¯ ). By Lemma 5: lef t(b¯ )  right(b¯ ), which completes the proof of
correctness.
As mentioned in Definition 12: half splitk is just splitk with the first k/4
comparators removed. So half bit mergek is just bit mergek with some com-
parators removed. Let’s count them: in each level of recursion step we take half
of comparators from splitk and additional one comparator from the base case
(k = 2). We sum them together to get:

log k−1
 


log k−2
k k  1
i
2 k 2 2 k
1+ =1+ − =1+ 2− − =
i=0
2i+2 4 i=0
2 k 4 k k 2

Therefore we have:
|pw hbit mergenk | = k/2 + k log k/2 − k/2 = k log k/2
 Smaller Selection Networks for Cardinality Constraints Encoding 219

The only difference between pw sel and our pw hbit sel is the use of improved
merger pw hbit merge rather than pw merge. By Theorem 3, we conclude that
|pw mergenk | ≥ |pw hbit mergenk |, so it follows that:
Remark 1. |pw hbit selkn | ≤ |pw selkn |

5 Sizes of New Selection Networks

In this section we estimate the size of pw hbit selkn . To this end we show that the
size of pw hbit selkn is upper-bounded by the size of bit selkn and use this fact in
our estimation. We also compute the exact difference between sizes of pw selkn
and pw hbit selkn and show that it can be as big as n log n/2. Finally we show
graphically how much smaller is our selection network on practical values of n
and k.
We have the recursive formula for the number of comparators of pw hbit selkn :

⎧ n/2 n/2
⎪
⎪ |pw hbit selk | + |pw hbit selk/2 |+
⎨
|pw hbit selkn | = +|split k| + |pw hbit merge |
n k
if k < n (3)
⎪
⎩ |oe sort
⎪ | if k = n
|maxn | if k = 1
Lemma 6. |pw hbit selkn | ≤ |bit selkn |.
Proof. Let aux selkn be the comparator network that is generated by substituting
recursive calls in pw hbit selkn by calls to bit selkn . Size of this network (for 1 <
k < n) is:

n/2 n/2
|aux selkn | = |bit selk | + |bit selk/2 | + |splitn | + |pw hbit mergek | (4)
Lemma 6 follows from Lemma 7 and Lemma 8 below, where we show that:

|pw hbit selkn | ≤ |aux selkn | ≤ |bit selkn |

Lemma 7. For 1 < k < n (both powers of 2), |aux selkn | ≤ |bit selkn |.
Proof. We compute both values from Equation 2 and 4:

1 5 1 5 3n
|aux selkn | = n log2 k + n − k log k − k −
4 2 4 4 2k
1 1 1 n
|bit selk | = n log k + n log k + 2n − k log k − k −
n 2
4 4 2 k
We simplify both sides to get the following inequality:
1 n 1
n − k − ≤ (n − k) log k
2 k 2
which can be easily proved by induction.
220 M. Karpiński and M. Piotrów

Lemma 8. For 1 ≤ k < n (both powers of 2), |pw hbit selkn | ≤ |aux selkn |.

Proof. By induction. For the base case, consider 1 = k < n. If follows by defi-
nitions that |pw hbit selkn | = |aux selkn | = n − 1. For the induction step assume
that for each (n , k  ) ≺ (n, k) (in lexicographical order) the lemma holds, we get:

|pw hbit selkn |

n/2 n/2
= |pw hbit selk/2 | + |pw hbit selk | + |splitn | + |pw hbit mergek |
(by the definition of pw hbit sel)
n/2 n/2
≤ |aux selk/2 | + |aux selk | + |splitn | + |pw hbit mergek |
(by the induction hypothesis)
n/2 n/2
≤ |bit selk/2 | + |bit selk | + |splitn | + |pw hbit mergek |
(by Lemma 7)
= |aux selkn |
(by the definition of aux sel)

Let N = 2n and K = 2k . We will compute upper bound for P (n, k) =

|pw hbit selK
N
| using B(n, k) = |bit selK
N
|.

Lemma 9. Let:

 i
 m

m−1
i    m
P (n, k, m) = (k − j)2k−j−1
+2n−i−1
+ P (n − m, k − i).
i=0 j=0
j i=0
i

Then ∀0≤m≤min(k,n−k) P (n, k, m) = P (n, k).

Proof. The lemma can be easily proved by induction on m.

 2     
k 3 m k
Lemma 10. P (n, k, m) ≤ 2n−2 k − m 2 + k + 7m
4 + 8 +2 2 2 −
m
6 −
 
n−k 3 m
2 (k + 1) − 2
k
2 .

Proof. Due to space restriction we only present schema of the proof. The lemma
can be proven directly from the inequality below, which is a consequence of
Lemma 9 and 6.

 i
 m

m−1
i    m
P (n, k, m) ≤ (k − j)2k−j−1
+2n−i−1
+ B(n − m, k − i)
i=0 j=0
j i=0
i

  
7 2
Theorem 4. For m = min(k, n−k), P (n, k) ≤ 2n−2 k− m
2 − 4 + 9k
2 + 79
16
 m  k m   
n−k 3 m
+2k 32 2 − 6 − 2 (k + 1) − 2
k
2 .
 Smaller Selection Networks for Cardinality Constraints Encoding 221

Proof. Directly from Lemma 9 and 10.

We will now present the size difference SD(n, k) between pairwise selection
network and our network. Merging step in pw selK N
costs 2k k − 2k + 1 and in
N k−1
pw hbit selK : 2 k, so the difference is given by the following equation:

⎧
⎪
⎪ 0 if n = k
⎨
0 if k = 0
SD(n, k) = (5)
⎪
⎪ 2k−1
k − 2 k
+ 1+
⎩
+SD(n − 1, k) + SD(n − 1, k − 1) if 0 < k < n
k   k−j
Theorem 5. Let Sn,k = j=0 n−k+j j 2 . Then:


n n+1 n − 2k + 1
SD(n, k) = − Sn,k − 2k (k − 1) − 1
k 2 2
Proof. By straightforward calculation
  one can verify that Sn,0 = 1, Sn,n =
2n+1 − 1, Sn−1,k−1 = 12 (Sn,k − nk ) and Sn−1,k−1 + Sn−1,k = Sn,k . It follows that
the theorem is true for k = 0 and k = n. We prove the theorem by induction
on pairs (k, n). Take any (k, n), 0 < k < n, and assume that theorem holds for
every (k  , n ) ≺ (k, n) (in lexicographical order). Then we have:

SD(n, k) = 2k−1 k − 2k + 1 + SD(n − 1, k) + SD(n − 1, k − 1)



n−1 n n−1 n
=2 k−1
k−2 +1+
k
+ − 2k (k − 1) − 1
k 2 k−1 2
n − 2k n − 2k + 2
− 2k−1 (k − 2) − 1 − (Sn−1,k + Sn−1,k−1 )

 2 2
n n n − 2k
= − Sn,k − Sn−1,k−1 − 2k (k − 1) − 1
k 2 2


n n+1 n − 2k + 1
= − Sn,k − 2k (k − 1) − 1
k 2 2

−4
Corollary 1. |pw selN/2
N
|−|pw hbit selN/2
N
| = N log N
2 +log N +2, for N = 2n .
Plots in Figure 2 show how much pw sel and the upper bound from The-
orem 4 are worse than pw hbit sel. Lines labeled pw sel are plotted from
(|pw selKN
| − |pw hbit selK
N
|)/|pw hbit selK
N
| and the ones labeled upper are plot-
ted from the formula (|upperK |−|pw hbit selK
N N
|)/|pw hbit selK
N
|, where |upperKN
|
is the upper bound from Theorem 4. Both |pw selK | and |pw hbit selK | were
N N

computed directly from recursive formulas. We can see that we save the most
number of comparators when k is larger than n/2, nevertheless for small values
of n superiority of our network is apparent for any k. As for the upper bound,
it gives a good approximation of |pw hbit selK N
| when n is small, but for larger
values of n it becomes less satisfactory.
222 M. Karpiński and M. Piotrów

100 % 100 % 100 %

pw_sel pw_sel pw_sel
upper upper upper
80 % 80 % 80 %

60 % 60 % 60 %

40 % 40 % 40 %

20 % 20 % 20 %

0% 0% 0%
0 1 2 3 4 5 6 7 0 2 4 6 8 10 12 14 0 5 10 15 20 25 30
log(K) log(K) log(K)

(a) N = 27 (b) N = 215 (c) N = 231

N
Fig. 2. The relative percentage change of the size of pw selK and the upper bound
N
given in Thm 4 with respect to the size of pw hbit selK .

6 Arc-Consistency of Selection Networks

In this section we prove that half encoding of any selection network preserves
arc-consistency with respect to “less-than” cardinality constraints. The proof
can be generalized to other types of cardinality constraints.
We introduce the convention, that x1 , . . . , xn  will denote the input and
y1 , . . . , yn  will denote the output of some order n comparator network. We
would also like to view them as sequences of Boolean variables, that can be set
to either true (1), false (0) or undefined (X).
From now on we assume that every network f is half encoded and when we
say “comparator” or “network”, we view it in terms of CNF formulas. We denote
V [φ(f )] to be the set of variables in encoding φ(f ).

Observation 4. A single comparator hcomp(a, b, c, d) has the following propa-

gation properties:

1. If a = 1 or b = 1, then UP sets c = 1 (by 1.c1 or 1.c2).

2. If a = b = 1, then UP sets c = d = 1 (by 1.c1 and 1.c3).
3. If c = 0, then UP sets a = b = 0 (by 1.c1 and 1.c2).
4. If b = 1 and d = 0, then UP sets a = 0 (by 1.c3).
5. If a = 1 and d = 0, then UP sets b = 0 (by 1.c3).

Lemma 11. Let fkn be a selection network. Assume that k − 1 inputs are set
to 1, and rest of the variables are undefined. Unit propagation will set variables
y1 , . . . , yk−1 to 1.

Proof. From propagation properties of hcomp(a, b, c, d) we can see that if com-

parator receives two 1s, then it outputs two 1s, when it receives 1 on one input
and X on the other, then it outputs 1 on the upper output and X on the lower
output. From this we conclude that a single comparator will sort its inputs, as
long as one of the inputs is set to 1. No 1 is lost, so they must all reach the out-
puts. Because the comparators comprise a selection network, the 1s will appear
at outputs y1 , . . . , yk−1 .
 Smaller Selection Networks for Cardinality Constraints Encoding 223

The process of propagating 1s we call a forward propagation. For the remain-

der of this section assume that: fkn is a selection network; k − 1 inputs are set
to 1, and the rest of the variables are undefined; forward propagation has been
performed resulting in y1 , . . . , yk−1 to be set to 1.
Definition 13 (Path). A path is a sequence of Boolean variables z1 , . . . , zm 
such that ∀1≤i≤m zi ∈ V [φ(fkn )] and for all 1 ≤ i < m there exists a comparator
hcomp(a, b, c, d) in φ(fkn ) for which zi ∈ {a, b} and zi+1 ∈ {c, d}.
Definition 14 (Propagation Path). Let x be an undefined input variable. A
path z̄x = z1 , . . . , zm  (m ≥ 1) is a propagation path, if z1 ≡ x and z2 , . . . , zm 
is the sequence of variables that would be set to 1 by UP, if we would set z1 = 1.
Lemma 12. If z̄x = z1 , . . . , zm  is a propagation path for an undefined variable
x, then zm ≡ yk .
Proof. Remember that all y1 , . . . , yk−1 are set to 1. Setting any undefined input
variable x to 1 will result in UP to set yk to 1. Otherwise fkn would not be a
selection network.
The following lemma shows that propagation paths are deterministic.
Lemma 13. Let z̄x = z1 , . . . , zm  be a propagation path. For each 1 ≤ i ≤ m
and z1 ≡ zi , if z1 , . . . , zm

  is a path that would be set to 1 by UP if we would
  
set z1 = 1, then z1 , . . . , zm   = zi , . . . , zm .

Proof. By induction on l = m−i. If l = 0, then z1 ≡ zm ≡ yk (by Lemma 12), so

the lemma holds. Let l ≥ 0 and assume that the lemma is true for zl . Consider
z1 ≡ zl−1 ≡ zm−i−1 . Set zm−i−1 = 1 and use UP to set zm−i = 1. Notice that
zm−i ≡ z2 , otherwise there would exist a comparator hcomp(a, b, c, d), for which
zm−i−1 is equivalent to either a or b and zm−i ≡ c and z2 ≡ d (or vice versa).
That would mean that a single 1 on the input produces two 1s on the outputs.
This contradicts our reasoning in the proof of Lemma 11. By the induction
hypothesis z2 , . . . , zm
  
  = zm−i , . . . , zm , so z1 , . . . , zm  = zm−i−1 , . . . , zm .

For each undefined input variable x and propagation path z̄x = z1 , . . . , zm 
we define a directed graph Px = {zi , zi+1  : 1 ≤ i < m}.
Lemma 14. Let {xi1 , . . . , xit } (t > 0) be the set of undefined input variables.
Then T = Pxi1 ∪ . . . ∪ Pxit is the tree rooted at yk .
Proof. By induction on t. If t = 1, then T = Pxi1 and by Lemma 12, Pxi1 ends
in yk , so the lemma holds. Let t > 0 and assume that the lemma is true for t.
We will show that it is true for t + 1. Consider T = Pxi1 ∪ . . . ∪ Pxit ∪ Pxit+1 .
By the induction hypothesis T  = Pxi1 ∪ . . . ∪ Pxit is a tree rooted at yk . By
Lemma 12, V (Pxit+1 ) ∩ V (T  ) = ∅. Let z ∈ V (Pxit+1 ) be the first variable, such
that z ∈ V (T  ). Since z ∈ V (T  ), there exists j (1 ≤ j ≤ t) such that z ∈ Pxij .
By Lemma 13, starting from variable z, paths Pxit+1 and Pxij are identical.
224 M. Karpiński and M. Piotrów

Graph T from the above lemma will be called a propagation tree.

Theorem 6. If we set yk = 0, then unit propagation will set all undefined input
variables to 0.

Proof. Let T be the propagation tree rooted at yk . We prove by induction on

the height h of a subtree T  of T , that (*) if the root of T  is set to 0, then UP
sets all nodes of T  to 0. It follows that if yk is set to 0 then UP sets all undefined
input variables to 0. If h = 0, then V = {yk }, so (*) is trivially true. Let h > 0
and assume that (*) holds. We will show that (*) holds for height h + 1. Let
T  be the propagation tree of height h + 1 and let r = 0 be the root. Consider
children of r in T  and a comparator hcomp(a, b, c, d) for which r ∈ {c, d}:
Case 1: r has two children. The only case is when r ≡ c = 0. Unit propagation
sets a = b = 0. Nodes a and b are roots of propagation trees of height h and are
set to 0, therefore by the induction hypothesis all nodes in T  will be set to 0.
Case 2: r has one child. Consider two cases: (i) if r ≡ c = 0 and either a or b is
the child of r, then UP sets a = b = 0 and either a or b is the root of propagation
tree of height h and is set to 0, therefore by the induction hypothesis all nodes
in T  will be set to 0, (ii) r ≡ d = 0 and either a = c = 1 and b is the child of
r or b = c = 1 and a is the child of r. Both of them will be set to 0 by UP and
again we get the root of propagation tree of height h that is set to 0, therefore
by the induction hypothesis all nodes in T  will be set to 0.

7 Conclusions
We have constructed a new family of selection networks, which are based on the
pairwise selection ones, but require less comparators to merge subsequences. The
difference in sizes grows with k and is equal to n log 2n−4 + log n + 2 for k = n/2.
In addition, we have shown that any selection network encoded in a standard
way to a CNF formula preserves arc-consistency with respect to a corresponding
cardinality constraint. This property is important, as many SAT-solvers take
advantage of the unit-propagation algorithm, making the computation signifi-
cantly faster.
It’s also worth noting that using encodings based on selection networks give
an extra edge in solving optimization problems for which we need to solve a
sequence of problems that differ only in the decreasing bound of a cardinality
constraint. In this setting we only need to add one more clause ¬yk for a new
value of k, and the search can be resumed keeping all previous clauses as it is.
This works because if a comparator network is a k-selection network, then it is
also a k  -selection network for any k  < k. This property is called incremental
strengthening and most state-of-the-art SAT-solvers provide a user interface for
doing this.
We are expecting that our smaller encodings of cardinality constraints should
improve the performance of SAT solvers, but the statement needs an experimen-
tal evaluation. We start doing the evaluation and the results will be presented
in the near future.
 Smaller Selection Networks for Cardinality Constraints Encoding 225

References
1. Ası́n, R., Nieuwenhuis, R., Oliveras, A., Rodrı́guez-Carbonell, E.: Cardinality
networks and their applications. In: Kullmann, O. (ed.) SAT 2009. LNCS,
vol. 5584, pp. 167–180. Springer, Heidelberg (2009)
2. Ası́n, R., Nieuwenhuis, R., Oliveras, A., Rodrı́guez-Carbonell, E.: Cardinality
networks: a theoretical and empirical study. Constraints 16(2), 195–221 (2011)
3. Ası́n, R., Nieuwenhuis, R.: Curriculum-based course timetabling with SAT and
MaxSAT. Annals of Operations Research 218(1), 71–91 (2014)
4. Batcher, K.E.: Sorting networks and their applications. In: Proc. of the April
30-May 2, 1968, Spring Joint Computer Conference, AFIPS 1968 (Spring),
pp. 307–314. ACM, New York (1968)
5. Codish, M., Zazon-Ivry, M.: Pairwise networks are superior for selection. http://
www.cs.bgu.ac.il/∼mcodish/Papers/Sources/pairwiseSelection.pdf
6. Codish, M., Zazon-Ivry, M.: Pairwise cardinality networks. In: Clarke, E.M.,
Voronkov, A. (eds.) LPAR-16 2010. LNCS, vol. 6355, pp. 154–172. Springer,
Heidelberg (2010)
7. Eén, N., Sörensson, N.: Translating pseudo-boolean constraints into sat. Journal
on Satisfiability, Boolean Modeling and Computation 2, 1–26 (2006)
8. Knuth, D.E.: The Art of Computer Programming, Sorting and Searching, vol. 3,
2nd edn. Addison Wesley Longman Publishing Co. Inc., Redwood City (1998)
9. Parberry, I.: Parallel complexity theory. Pitman, Research notes in theoretical
computer science (1987)
10. Parberry, I.: The pairwise sorting network. Parallel Processing Letters 2, 205–211
(1992)
11. Schutt, A., Feydy, T., Stuckey, P.J., Wallace, M.G.: Why Cumulative decompo-
sition is not as bad as it sounds. In: Gent, I.P. (ed.) CP 2009. LNCS, vol. 5732,
pp. 746–761. Springer, Heidelberg (2009)
 PREFIX-PROJECTION Global Constraint
for Sequential Pattern Mining

Amina Kemmar1 , Samir Loudni2(B) , Yahia Lebbah1 ,

Patrice Boizumault2 , and Thierry Charnois3
1
LITIO, University of Oran 1, EPSECG of Oran, Oran, Algeria
kemmami@yahoo.fr, lebbah.yahia@univ-oran.dz
2
GREYC (CNRS UMR 6072), University of Caen, Caen, France
{samir.loudni,patrice.boizumault}@unicaen.fr
3
LIPN (CNRS UMR 7030), University PARIS 13, Villetaneuse, France
thierry.charnois@lipn.univ-paris13.fr

Abstract. Sequential pattern mining under constraints is a challenging

data mining task. Many efficient ad hoc methods have been developed
for mining sequential patterns, but they are all suffering from a lack
of genericity. Recent works have investigated Constraint Programming
(CP) methods, but they are not still effective because of their encoding.
In this paper, we propose a global constraint based on the projected
databases principle which remedies to this drawback. Experiments show
that our approach clearly outperforms CP approaches and competes well
with ad hoc methods on large datasets.

1 Introduction

Mining useful patterns in sequential data is a challenging task. Sequential pat-

tern mining is among the most important and popular data mining task with
many real applications such as the analysis of web click-streams, medical or
biological data and textual data. For effectiveness and efficiency considerations,
many authors have promoted the use of constraints to focus on the most promis-
ing patterns according to the interests given by the final user. In line with [15],
many efficient ad hoc methods have been developed but they suffer from a lack
of genericity to handle and to push simultaneously sophisticated combination of
various types of constraints. Indeed, new constraints have to be hand-coded and
their combinations often require new implementations.
Recently, several proposals have investigated relationships between sequen-
tial pattern mining and constraint programming (CP) to revisit data mining
tasks in a declarative and generic way [5,9,11,12]. The great advantage of these
approaches is their flexibility. The user can model a problem and express his
queries by specifying what constraints need to be satisfied. But, all these pro-
posals are not effective enough because of their CP encoding. Consequently, the
design of new efficient declarative models for mining useful patterns in sequential
data is clearly an important challenge for CP.


c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 226–243, 2015.
DOI: 10.1007/978-3-319-23219-5 17
 PREFIX-PROJECTION Global Constraint for Sequential Pattern Mining 227

To address this challenge, we investigate in this paper the other side

of the cross fertilization between data-mining and constraint programming,
namely how the CP framework can benefit from the power of candidate prun-
ing mechanisms used in sequential pattern mining. First, we introduce the
global constraint Prefix-Projection for sequential pattern mining. Prefix-
Projection uses a concise encoding and its filtering relies on the principle
of projected databases [14]. The key idea is to divide the initial database into
smaller ones projected on the frequent subsequences obtained so far, then, mine
locally frequent patterns in each projected database by growing a frequent pre-
fix. This global constraint utilizes the principle of prefix-projected database to
keep only locally frequent items alongside projected databases in order to remove
infrequent ones from the domains of variables. Second, we show how the concise
encoding allows for a straightforward implementation of the frequency constraint
(Prefix-Projection constraint) and constraints on patterns such as size, item
membership and regular expressions and the simultaneous combination of them.
Finally, experiments show that our approach clearly outperforms CP approaches
and competes well with ad hoc methods on large datasets for mining frequent
sequential patterns or patterns under various constraints. It is worth noting that
the experiments show that our approach achieves scalability while it is a major
issue of CP approaches.
The paper is organized as follows. Section 2 recalls preliminaries. Section 3
provides a critical review of ad hoc methods and CP approaches for sequential
pattern mining. Section 4 presents the global constraint Prefix-Projection.
Section 5 reports experiments we performed. Finally, we conclude and draw some
perspectives.

2 Preliminaries
This section presents background knowledge about sequential pattern mining
and constraint satisfaction problems.

2.1 Sequential Patterns

Let I be a finite set of items. The language of sequences corresponds to LI = I n
where n ∈ N+ .
Definition 1 (sequence, sequence database). A sequence s over LI is an
ordered list s1 s2 . . . sn , where si , 1 ≤ i ≤ n, is an item. n is called the length
of the sequence s. A sequence database SDB is a set of tuples (sid, s), where sid
is a sequence identifier and s a sequence.
Definition 2 (subsequence,  relation). A sequence α = α1 . . . αm  is a
subsequence of s = s1 . . . sn , denoted by (α  s), if m ≤ n and there exist
integers 1 ≤ j1 ≤ . . . ≤ jm ≤ n, such that αi = sji for all 1 ≤ i ≤ m. We
also say that α is contained in s or s is a super-sequence of α. For example,
the sequence BABC is a super-sequence of AC : AC  BABC. A tuple
(sid, s) contains a sequence α, if α  s.
228 A. Kemmar et al.

Table 1. SDB1 : a sequence database example.

sid Sequence
1 ABCBC
2 BABC
3 AB
4 BCD

The cover of a sequence p in SDB is the set of all tuples in SDB in which
p is contained. The support of a sequence p in SDB is the number of tuples in
SDB which contain p.
Definition 3 (coverage, support). Let SDB be a sequence database and
p a sequence. coverSDB (p)={(sid, s) ∈ SDB | p  s} and supSDB (p) = #
coverSDB (p).
Definition 4 (sequential pattern). Given a minimum support threshold
minsup, every sequence p such that supSDB (p) ≥ minsup is called a sequen-
tial pattern [1]. p is said to be frequent in SDB.

Example 1. Table 1 represents a sequence database of four sequences where

the set of items is I = {A, B, C, D}. Let the sequence p = AC. We have
coverSDB1 (p) = {(1, s1 ), (2, s2 )}. If we consider minsup = 2, p = AC is a
sequential pattern because supSDB1 (p) ≥ 2.
Definition 5 (sequential pattern mining (SPM)). Given a sequence
database SDB and a minimum support threshold minsup. The problem of
sequential pattern mining is to find all patterns p such that supSDB (p) ≥
minsup.

2.2 SPM under Constraints

In this section, we define the problem of mining sequential patterns in a sequence
database satisfying user-defined constraints Then, we review the most usual
constraints for the sequential mining problem [15].
Problem statement. Given a constraint C(p) on pattern p and a sequence
database SDB, the problem of constraint-based pattern mining is to find the
complete set of patterns satisfying C(p). In the following, we present different
types of constraints that we explicit in the context of sequence mining. All these
constraints will be handled by our concise encoding (see Sections 4.2 and 4.5).
- The minimum size constraint size(p, min ) states that the number of items of
p must be greater than or equal to min .
- The item constraint item(p, t) states that an item t must belong (or not) to a
pattern p.
- The regular expression constraint [7] reg(p, exp) states that a pattern p must be
accepted by the deterministic finite automata associated to the regular expres-
sion exp.
 PREFIX-PROJECTION Global Constraint for Sequential Pattern Mining 229

2.3 Projected Databases

We now present the necessary definitions related to the concept of projected

databases [14].

Definition 6 (prefix, projection, suffix). Let β = β1 . . . βn  and α =

α1 . . . αm  be two sequences, where m ≤ n.
- Sequence α is called the prefix of β iff ∀i ∈ [1..m], αi = βi .
- Sequence β = β1 . . . βn  is called the projection of some sequence s w.r.t. α,
iff (1) β  s, (2) α is a prefix of β and (3) there exists no proper super-sequence
β  of β such that β   s and β  also has α as prefix.
- Sequence γ = βm+1 . . . βn  is called the suffix of s w.r.t. α. With the standard
concatenation operator ”concat”, we have β = concat(α, γ).

Definition 7 (projected database). Let SDB be a sequence database, the α-

projected database, denoted by SDB|α , is the collection of suffixes of sequences
in SDB w.r.t. prefix α.

[14] have proposed an efficient algorithm, called PrefixSpan, for mining

sequential patterns based on the concept of projected databases. It proceeds by
dividing the initial database into smaller ones projected on the frequent subse-
quences obtained so far; only their corresponding suffixes are kept. Then, sequen-
tial patterns are mined in each projected database by exploring only locally
frequent patterns.

Example 2. Let us consider the sequence database of Table 1 with minsup = 2.

PrefixSpan starts by scanning SDB 1 to find all the frequent items, each of them
is used as a prefix to get projected databases. For SDB 1 , we get 3 disjoint subsets
w.r.t. the prefixes A, B, and C. For instance, SDB 1 |A consists of 3 suffix
sequences: {(1, BCBC), (2, BC), (3, B)}. Consider the projected database
SDB 1 |<A> , its locally frequent items are B and C. Thus, SDB 1 |<A> can be
recursively partitioned into 2 subsets w.r.t. the two prefixes AB and AC.
The AB- and AC- projected databases can be constructed and recursively
mined similarly. The processing of a α-projected database terminates when no
frequent subsequence can be generated.

Proposition 1 establishes the support count of a sequence γ in SDB|α [14]:

Proposition 1 (Support count). For any sequence γ in SDB with prefix α
and suffix β s.t. γ = concat(α, β), supSDB (γ) = supSDB|α (β).
This proposition ensures that only the sequences in SDB grown from α need
to be considered for the support count of a sequence γ. Furthermore, only those
suffixes with prefix α should be counted.
230 A. Kemmar et al.

2.4 CSP and Global Constraints

A Constraint Satisfaction Problem (CSP) consists of a set X of n variables, a

domain D mapping each variable Xi ∈ X to a finite set of values D(Xi ), and a
set of constraints C. An assignment σ is a mapping from variables in X to values
in their domains: ∀Xi ∈ X, σ(Xi ) ∈ D(Xi ). A constraint c ∈ C is a subset of the
cartesian product of the domains of the variables that are in c. The goal is to
find an assignment such that all constraints are satisfied.
Domain Consistency (DC). Constraint solvers typically use backtracking
search to explore the space of partial assignments. At each assignment, filtering
algorithms prune the search space by enforcing local consistency properties like
domain consistency. A constraint c on X is domain consistent, if and only if, for
every Xi ∈ X and for every di ∈ D(Xi ), there is an assignment σ satisfying c
such that σ(Xi ) = di . Such an assignment is called a support.
Global constraints provide shorthands to often-used combinatorial substruc-
tures. We present two global constraints. Let X = X1 , X2 , ..., Xn  be a sequence
of n variables.
Let V be a set of values, l and u be two integers s.t. 0 ≤ l ≤ u ≤ n, the constraint
Among(X, V, l, u) states that each value a ∈ V should occur at least l times and at
most u times in X [4]. Given a deterministic finite automaton A, the constraint
Regular(X, A) ensures that the sequence X is accepted by A [16].

3 Related Works

This section provides a critical review of ad hoc methods and CP approaches for
SPM.

3.1 Ad hoc Methods for SPM

GSP [17] was the first algorithm proposed to extract sequential patterns. It uses
a generate-and test approach. Later, two major classes of methods have been
proposed:
- Depth-first search based on a vertical database format e.g. cSpade incorporat-
ing contraints (max-gap, max-span, length) [21], SPADE [22] or SPAM [2].
- Projected pattern growth such as PrefixSpan [14] and its extensions, e.g.
CloSpan for mining closed sequential patterns [19] or Gap-BIDE [10] tackling the
gap constraint.
In [7], the authors proposed SPIRIT based on GSP for SPM with regular
expressions. Later, [18] introduces Sequence Mining Automata (SMA), a new
approach based on a specialized kind of Petri Net. Two variants of SMA were
proposed: SMA-1P (SMA one pass) and SMA-FC (SMA Full Check). SMA-1P pro-
cesses by means of the SMA all sequences one by one, and enters all resulting
valid patterns in a hash table for support counting, while SMA-FC allows fre-
quency based pruning during the scan of the database. Finally, [15] provides a
 PREFIX-PROJECTION Global Constraint for Sequential Pattern Mining 231

survey for other constraints such as regular expressions, length and aggregates.
But, all these proposals, though efficient, are ad hoc methods suffering from a
lack of genericity. Adding new constraints often requires to develop new imple-
mentations.

3.2 CP Methods for SPM

Following the work of [8] for itemset mining, several methods have been proposed
to mine sequential patterns using CP.
Proposals. [5] have proposed a first SAT-based model for discovering a special
class of patterns with wildcards1 in a single sequence under different types of
constraints (e.g. frequency, maximality, closedness). [11] have proposed a CSP
model for SPM. Each sequence is encoded by an automaton capturing all sub-
sequences that can occur in it. [9] have proposed a CSP model for SPM with
wildcards. They show how some constraints dealing with local patterns (e.g.
frequency, size, gap, regular expressions) and constraints defining more complex
patterns such as relevant subgroups [13] and top-k patterns can be modeled using
a CSP. [12] have proposed two CP encodings for the SPM. The first one uses
a global constraint to encode the subsequence relation (denoted global-p.f),
while the second one encodes explicitly this relation using additional variables
and constraints (denoted decomposed-p.f).
All these proposals use reified constraints to encode the database. A reified
constraint associates a boolean variable to a constraint reflecting whether the
constraint is satisfied (value 1) or not (value 0). For each sequence s of SDB,
a reified constraint, stating whether (or not) the unknown pattern p is a subse-
quence of s, is imposed: (Ss = 1) ⇔ (p  s). A great consequenceis that the
encoding of the frequency measure is straightforward: f req(p) = s∈SDB Ss .
But such an encoding has a major drawback since it requires (m = #SDB)
reified constraints to encode the whole database. This constitutes a strong limi-
tation of the size of the databases that could be managed.
Most of these proposals encode the subsequence relation (p  s) using
variables P oss,j (s ∈ SDB and 1 ≤ j ≤ ) to determine a position where p
occurs in s. Such an encoding requires a large number of additional variables
(m×) and makes the labeling computationally expensive. In order to address
this drawback, [12] have proposed a global constraint exists-embedding to
encode the subsequence relation, and used projected frequency within an ad hoc
specific branching strategy to keep only frequent items before branching over
the variables of the pattern. But, this encoding still relies on reified constraints
and requires to impose m exists-embedding global constraints.
So, we propose in the next section the Prefix-Projection global con-
straint that fully exploits the principle of projected databases to encode both
the subsequence relation and the frequency constraint. Prefix-Projection
does not require any reified constraints nor any extra variables to encode the

1
A wildcard is a special symbol that matches any item of I including itself.
232 A. Kemmar et al.

subsequence relation. As a consequence, usual SPM constraints (see Section 2.2)

can be encoded in a straightforward way using directly the (global) constraints
of the CP solver.

4 PREFIX-PROJECTION Global Constraint

This section presents the Prefix-Projection global constraint for the SPM
problem.

4.1 A Concise Encoding

Let P be the unknown pattern of size  we are looking for. The symbol 2 stands
for an empty item and denotes the end of a sequence. The unknown pattern P is
encoded with a sequence of  variables P1 , P2 , . . . , P  s.t. ∀i ∈ [1 . . . ], D(Pi ) =
I ∪ {2}. There are two basic rules on the domains:
1. To avoid the empty sequence, the first item of P must be non empty, so
(2 ∈ D1 ).
2. To allow patterns with less than  items, we impose that ∀i ∈
[1..(−1)], (Pi = 2) → (Pi+1 = 2).

4.2 Definition and Consistency Checking

The global constraint Prefix-Projection ensures both subsequence relation
and minimum frequency constraint.
Definition 8 (Prefix-Projection global constraint). Let P = P1 , P2 , . . . , P 
be a pattern of size . d1 , ..., d  ∈ D(P1 ) × . . . × D(P ) is a solution of Prefix-
Projection (P, SDB, minsup) iff supSDB (d1 , ..., d ) ≥ minsup.
Proposition 2. A Prefix-Projection (P, SDB, minsup) constraint has a
solution if and only if there exists an assignment σ = d1 , ..., d  of variables
of P s.t. SDB|σ has at least minsup suffixes of σ: #SDB|σ ≥ minsup.
Proof: This is a direct consequence of proposition 1. We have straightforwardly
supSDB (σ) = supSDB|σ () = #SDB|σ . Thus, suffixes of SDB|σ are supports
of σ in the constraint Prefix-Projection (P, SDB, minsup), provided that
#SDB|σ ≥ minsup. 2
The following proposition characterizes values in the domain of unassigned
(i.e. future) variable Pi+1 that are consistent with the current assignment of
variables P1 , ..., Pi .
Proposition 3. Let σ 2 = d1 , . . . , di  be a current assignment of variables
P1 , . . . , Pi , Pi+1 be a future variable. A value d ∈ D(Pi+1 ) appears in a solu-
tion for Prefix-Projection (P, SDB, minsup) if and only if d is a frequent
item in SDB|σ :
#{(sid, γ)|(sid, γ) ∈ SDB|σ ∧ dγ} ≥ minsup
2
We indifferently denote σ by d1 , . . . , di  or by σ(P1 ), . . . , σ(Pi ).
 PREFIX-PROJECTION Global Constraint for Sequential Pattern Mining 233

Proof: Suppose that value d ∈ D(Pi+1 ) occurs in SDB|σ more than minsup.
From proposition 1, we have supSDB (concat(σ, d)) = supSDB|σ (d). Hence,
the assignment σ ∪ d satisfies the constraint, so d ∈ D(Pi+1 ) participates in a
solution. 2
Anti-monotonicity of the Frequency Measure. If a pattern p is not fre-
quent, then any pattern p satisfying p  p is not frequent. From proposition 3
and according to the anti-monotonicity property, we can derive the following
pruning rule:

Proposition 4. Let σ = d1 , . . . , di  be a current assignment of variables

P1 , . . . , Pi . All values d ∈ D(Pi+1 ) that are locally not frequent in SDB|σ
can be pruned from the domain of variable Pi+1 . Moreover, these values d can
also be pruned from the domains of variables Pj with j ∈ [i + 2, . . . , ].

Proof: Let σ = d1 , . . . , di  be a current assignment of variables P1 , . . . , Pi . Let

d ∈ D(Pi+1 ) s.t. σ  = concat(σ, d). Suppose that d is not frequent in SDB|σ .
According to proposition 1, supSDB|σ (d) = supSDB (σ  ) < minsup, thus σ  is
not frequent. So, d can be pruned from the domain of Pi+1 .
Suppose that the assignment σ has been extended to concat(σ, α), where α
corresponds to the assignment of variables Pj (with j > i). If d ∈ D(Pi+1 ) is not
frequent, it is straightforward that supSDB|σ (concat(α, d)) ≤ supSDB|σ (d) <
minsup. Thus, if d is not frequent in SDB|σ , it will be also not frequent in
SDB|concat(σ,α) . So, d can be pruned from the domains of Pj with j ∈ [i +
2, . . . , ]. 2

Example 3. Consider the sequence database of Table 1 with minsup = 2. Let

P = P1 , P2 , P3  with D(P1 ) = I and D(P2 ) = D(P3 ) = I ∪ {2}. Suppose that
σ(P1 ) = A, Prefix-Projection(P, SDB, minsup) will remove values A and
D from D(P2 ) and D(P3 ), since the only locally frequent items in SDB 1 |<A>
are B and C.

Proposition 4 guarantees that any value (i.e. item) d ∈ D(Pi+1 ) present

but not frequent in SDB|σ does not need to be considered when extending
σ, thus avoiding searching over it. Clearly, our global constraint encodes the
anti-monotonicity of the frequency measure in a simple and elegant way, while
CP methods for SPM have difficulties to handle this property. In [12], this is
achieved by using very specific propagators and branching strategies, making
the integration quite complex (see [12]).

4.3 Building the Projected Databases

The key issue of our approach lies in the construction of the projected databases.
When projecting a prefix, instead of storing the whole suffix as a projected sub-
sequence, one can represent each suffix by a pair (sid, start) where sid is the
sequence identifier and start is the starting position of the projected suffix in
the sequence sid. For instance, let us consider the sequence database of Table 1.
234 A. Kemmar et al.

Algorithm 1. ProjectSDB(SDB, P rojSDB, α)

Data: SDB: initial database; P rojSDB: projected sequences; α: prefix
begin
1 SDB|α ← ∅ ;
2 for each pair (sid, start) ∈ P rojSDB do
3 s ← SDB[sid] ;
4 posα ← 1; poss ← start ;
5 while (posα ≤ #α ∧ poss ≤ #s) do
6 if (α[posα ] = s[poss ]) then
7 posα ← posα + 1 ;
8 poss ← poss + 1 ;
9 if (posα = #α + 1) then
10 SDB|α ← SDB|α ∪ {(sid, poss )}

11 return SDB|α ;

As shown in example 2, SDB|A consists of 3 suffix sequences: {(1, BCBC),

(2, BC), (3, B)}. By using the pseudo-projection, SDB|A can be repre-
sented by the following three pairs: {(1, 2), (2, 3), (3, 2)}. This is the principle of
pseudo-projection, adopted in , exploited during the filtering step of our Prefix-
Projection global constraint. Algorithm 1 details this principle. It takes as
input a set of projected sequences P rojSDB and a prefix α. The algorithm pro-
cesses all the pairs (sid, start) of P rojSDB one by one (line 2), and searches
for the lowest location of α in the sequence s corresponding to the sid of that
sequence in SDB (lines 6-8).
In the worst case, ProjectSDB processes all the items of all sequences. So,
the time complexity is O( × m), with m = #SDB and  is the length of the
longest sequence in SDB. The worst case space complexity of pseudo-projection
is O(m), since we need to store for each sequence only a pair (sid, start), while
for the standard projection the space complexity is O(m×). Clearly, the pseudo-
projection takes much less space than the standard projection.

4.4 Filtering

Ensuring DC on Prefix-Projection(P, SDB, minsup) is equivalent to finding

a sequential pattern of length ( − 1) and then checking whether this pattern
remains a frequent pattern when extended to any item d in D(P ). Thus, finding
such an assignment (i.e. support) is as much as difficult than the original prob-
lem of sequential pattern mining. [20] has proved that the problem of counting
the number of maximal3 frequent patterns in a database of sequences is #P-
complete, thereby proving the NP-hardness of the problem of mining maximal
frequent sequences. The difficulty is due to the exponential number of candi-
dates that should be parsed to find the frequent patterns. Thus, finding, for
every variable Pi ∈ P and for every di ∈ D(Pi ), an assignment σ satisfying
Prefix-Projection(P, SDB, minsup) s.t. σ(Pi ) = di is of exponential nature.

3
A sequential pattern p is maximal if there is no sequential pattern q such that pq.
 PREFIX-PROJECTION Global Constraint for Sequential Pattern Mining 235

Algorithm 2. Filter-Prefix-Projection(SDB, σ, i, P , minsup)

Data: SDB: initial database; σ: current prefix
σ(P1 ), . . . , σ(Pi ); minsup: the minimum
support threshold; PSDB: internal data structure of Prefix-Projection for storing
pseudo-projected databases
begin
1 if (i ≥ 2 ∧ σ(Pi ) = 2) then
2 for j ← i + 1 to  do
3 Pj ← 2;
4 return True;
else
5 PSDBi ← ProjectSDB(SDB, PSDBi−1 ,
σ(Pi ));
6 if (#PSDBi < minsup) then
7 return False ;
else
8 F I ← getFreqItems(SDB, PSDBi , minsup) ;
9 for j ← i + 1 to  do
10 foreach a ∈ D(Pj ) s.t.(a = 2 ∧ a ∈
/ F I) do
11 D(Pj ) ← D(Pj ) − {a};

12 return True;

Function getFreqItems (SDB, P rojSDB, minsup) ;

Data: SDB: the initial database; P rojSDB: pseudo-projected database; minsup: the
minimum support threshold; ExistsItem, SupCount: internal data structures using
a hash table for support counting over items;
begin
13 SupCount[] ← {0, ..., 0}; F ← ∅ ;
14 for each pair (sid, start) ∈ P rojSDB do
15 ExistsItem[] ← {f alse, ..., f alse}; s ← SDB[sid] ;
16 for i ← start to #s do
17 a ← s[i] ;
18 if (¬ExistsItem[a]) then
19 SupCount[a] ← SupCount[a] + 1; ExistsItem[a] ← true;
20 if (SupCount[a] ≥ minsup) then
21 F ← F ∪ {a};

22 return F ;

So, the filtering of the Prefix-Projection constraint maintains a consis-

tency lower than DC. This consistency is based on specific properties of the
projected databases (see Proposition 3), and anti-monotonicity of the frequency
constraint (see Proposition 4), and resembles forward-checking regarding Propo-
sition 3. Prefix-Projection is considered as a global constraint, since all vari-
ables share the same internal data structures that awake and drive the filtering.
Algorithm 2 describes the pseudo-code of the filtering algorithm of the
Prefix-Projection constraint. It is an incremental filtering algorithm that
should be run when some i first variables are assigned according to the following
lexicographic ordering P1 , P2 , . . . , P  of variables of P . It exploits internal data-
structures enabling to enhance the filtering algorithm. More precisely, it uses an
incremental data structure, denoted PSDB, that stores the intermediate pseudo-
projections of SDB, where PSDBi (i ∈ [0, . . . , ]) corresponds to the σ-projected
database of the current partial assignment σ = σ(P1 ), . . . , σ(Pi ) (also called
prefix) of variables P1 , . . . , Pi , and PSDB0 = {(sid, 1)|(sid, s) ∈ SDB} is the
236 A. Kemmar et al.

initial pseudo-projected database of SDB (case where σ = ). It also uses a

hash table indexing the items I into integers (1 . . . #I) for an efficient support
counting over items (see function getFreqItems).
Algorithm 2 takes as input the current partial assignment σ = σ(P1 ), . . . , σ(Pi )
of variables P1 , . . . , Pi , the length i of σ (i.e. position of the last assigned vari-
able in P ) and the minimum support threshold minsup. It starts by checking if
the last assigned variable Pi is instantiated to 2 (line 1). In this case, the end of
sequence is reached (since value 2 can only appear at the end) and the sequence
σ(P1 ), . . . , σ(Pi ) constitutes a frequent pattern in SDB; hence the algorithm sets
the remaining ( − i) unassigned variables to 2 and returns true (lines 2-4). Oth-
erwise, the algorithm computes incrementally PSDBi from PSDBi−1 by calling
function ProjectSDB (see Algorithm 1). Then, it checks in line 6 whether the
current assignment σ is a legal prefix for the constraint (see Proposition 2). This
is done by computing the size of PSDBi . If this size is less than minsup, we stop
growing σ and we return false. Otherwise, the algorithm computes the set of locally
frequent items FI in PSDBi by calling function getFreqItems (line 8).
Function getFreqItems processes all the entries of the pseudo-projected
database one by one, counts the number of first occurrences of items a (i.e.
SupCount[a]) in each entry (sid, start), and keeps only the frequent ones (lines
13-21). This is done by using ExistsItem data structure. After the whole pseudo-
projected database has been processed, the frequent items are returned (line 22),
and Algorithm 2 updates the current domains of variables Pj with j ≥ (i + 1)
by pruning inconsistent values, thus avoiding searching over not frequent items
(lines 9-11).
Proposition 5. In the worst case, filtering with Prefix-Projection global
constraint can be achieved in O(m ×  + m × d +  × d). The worst case space
complexity of Prefix-Projection is O(m × ).
Proof: Let  be the length of the longest sequence in SDB, m = #SDB, and
d = #I. Computing the pseudo-projected database PSDBi can be done in
O(m × ): for each sequence (sid, s) of SDB, checking if σ occurs in s is O()
and there are m sequences. The total complexity of function GetFreqItems is
O(m×(+d)). Lines (9-11) can be achieved in O(×d). So, the whole complexity
is O(m×+m×(+d)+×d) = O(m×+m×d+×d). The space complexity of
the filtering algorithm lies in the storage of the PSDB internal data structure.
In the worst case, we have to store  pseudo-projected databases. Since each
pseudo-projected database requires O(m), the worst case space complexity is
O(m × ). 2

4.5 Encoding of SPM Constraints

Usual SPM constraints (see Section 2.2) can be reformulated in a straightforward
way. Let P be the unknown pattern. i=
- Minimum size constraint: size(P, min ) ≡ i=1min (Pi = )
- Item constraint:let V be a subset of items, l and u two integers s.t. 0 ≤ l ≤ u ≤
. item(P, V ) ≡ t∈V Among(P, {t}, l, u) enforces that items of V should occur
 PREFIX-PROJECTION Global Constraint for Sequential Pattern Mining 237

Table 2. Dataset Characteristics.

dataset #SDB #I avg (#s) maxs∈SDB (#s) type of data

Leviathen 5834 9025 33.81 100 book
Kosarak 69999 21144 7.97 796 web click stream
FIFA 20450 2990 34.74 100 web click stream
BIBLE 36369 13905 21.64 100 bible
Protein 103120 24 482 600 protein sequences
data-200K 200000 20 50 86 synthetic dataset
PubMed 17527 19931 29 198 bio-medical text

at least l times and at most u times in P . To forbid items of V to occur in P , l

and u must be set to 0.
- Regular expression constraint: let Areg be the deterministic finite automaton
encoding the regular expression exp. reg(P, exp) ≡ Regular(P, Areg ).

5 Experimental Evaluation
This section reports experiments on several real-life datasets from [3,6,18] of
large size having varied characteristics and representing different application
domains (see Table 2). Our objective is (1) to compare our approach to existing
CP methods as well as to state-of-the-art methods for SPM in terms of scalability
which is a major issue of existing CP methods, (2) to show the flexibility of our
approach allowing to handle different constraints simultaneously.
Experimental Protocol. The implementation of our approach was carried
out in the Gecode solver4 . All experiments were conducted on a machine with a
processor Intel X5670 and 24 GB of memory. A time limit of 1 hour has been
used. For each dataset, we varied the minsup threshold until the methods are
not able to complete the extraction of all patterns within the time limit.  was
set to the length of the longest sequence of SDB. The implementation and the
datasets used in our experiments are available online5 . We compare our approach
(indicated by PP) with:
1. two CP encodings [12], the most efficient CP methods for SPM: global-p.f
and decomposed-p.f;
2. state-of-the-art methods for SPM : and cSpade;
3. SMA [18] for SPM under regular expressions.
We used the author’s cSpade implementation 6 for SPM, the publicly avail-
able implementations of by Y. Tabei 7 and the SMA implementation 8 for SPM
under regular expressions. The implementation 9 of the two CP encodings was
4
http://www.gecode.org
5
https://sites.google.com/site/prefixprojection4cp/
6
http://www.cs.rpi.edu/∼zaki/www-new/pmwiki.php/Software/
7
https://code.google.com/p/prefixspan/
8
http://www-kdd.isti.cnr.it/SMA/
9
https://dtai.cs.kuleuven.be/CP4IM/cpsm/
238 A. Kemmar et al.

Fig. 1. Comparing PP with global-p.f for SPM on real-life datasets: CPU times.

carried out in the Gecode solver. All methods have been executed on the same
machine.
(a) Comparing with CP Methods for SPM. First we compare PP with
the two CP encodings global-p.f and decomposed-p.f (see Section 3.2). CPU
times (in logscale for BIBLE, Kosarak and PubMed) of the three methods are
shown on Fig. 1. First, decomposed-p.f is the least performer method. On all
the datasets, it fails to complete the extraction within the time limit for all
values of minsup we considered. Second, PP largely dominates global-p.f on
all the datasets: PP is more than an order of magnitude faster than global-p.f.
The gains in terms of CPU times are greatly amplified for low values of minsup.
On BIBLE (resp. PubMed), the speed-up is 84.4 (resp. 33.5) for minsup equal
to 1%. Another important observation that can be made is that, on most of
the datasets (except BIBLE and Kosarak), global-p.f is not able to mine for
patterns at very low frequency within the time limit. For example on FIFA, PP is
able to complete the extraction for values of minsup up to 6% in 1, 457 seconds,
while global-p.f fails to complete the extraction for minsup less than 10%.
To complement the results given by Fig. 1, Table 3 reports for different
datasets and different values of minsup, the number of calls to the propagate
routine of Gecode (column 5), and the number of nodes of the search tree (column
6). First, PP explores less nodes than global-p.f. But, the difference is not huge
(gains of 45% and 33% on FIFA and BIBLE respectively). Second, our approach
is very effective in terms of number of propagations. For PP, the number of
propagations remains small (in thousands for small values of minsup) compared
to global-p.f (in millions). This is due to the huge number of reified constraints
used in global-p.f to encode the subsequence relation. On the contrary, our
Prefix-Projection global constraint does not require any reified constraints
nor any extra variables to encode the subsequence relation.
 PREFIX-PROJECTION Global Constraint for Sequential Pattern Mining 239

Table 3. PP vs. global-p.f.

CPU times (s) #PROPAGATIONS #NODES
Dataset minsup (%) #PATTERNS
PP global-p.f PP global-p.f PP global-p.f
20 938 8.16 129.54 1884 11649290 1025 1873
18 1743 13.39 222.68 3502 19736442 1922 3486
16 3578 24.39 396.11 7181 35942314 3923 7151
FIFA
14 7313 44.08 704 14691 65522076 8042 14616
12 16323 86.46 1271.84 32820 126187396 18108 32604
10 40642 185.88 2761.47 81767 266635050 45452 81181
10 174 1.98 105.01 363 4189140 235 348
8 274 2.47 153.61 575 5637671 362 548
6 508 3.45 270.49 1065 8592858 669 1016
BIBLE
4 1185 5.7 552.62 2482 15379396 1575 2371
2 5311 15.05 1470.45 11104 39797508 7048 10605
1 23340 41.4 3494.27 49057 98676120 31283 46557
5 2312 8.26 253.16 4736 15521327 2833 4619
4 3625 11.17 340.24 7413 20643992 4428 7242
3 6336 16.51 536.96 12988 29940327 7757 12643
PubMed
2 13998 28.91 955.54 28680 50353208 17145 27910
1 53818 77.01 2581.51 110133 124197857 65587 107051
99.99 127 165.31 219.69 264 26731250 172 221
99.988 216 262.12 411.83 451 44575117 293 390
99.986 384 467.96 909.47 805 80859312 514 679
Protein
99.984 631 753.3 1443.92 1322 132238827 845 1119
99.982 964 1078.73 2615 2014 201616651 1284 1749
99.98 2143 2315.65 − 4485 − 2890 −

(b) Comparing with ad hoc Methods for SPM. Our second experiment
compares PP with state-of-the-art methods for SPM. Fig. 2 shows the CPU times
of the three methods. First, cSpade obtains the best performance on all datasets
(except on Protein). However, PP exhibits a similar behavior as cSpade, but it is
less faster (not counting the highest values of minsup). The behavior of cSpade
on Protein is due to the vertical representation format that is not appropriated
in the case of databases having large sequences and small number of distinct
items, thus degrading the performance of the mining process. Second, PP which
also uses the concept of projected databases, clearly outperforms on all datasets.
This is due to our filtering algorithm combined together with incremental data
structures to manage the projected databases. On FIFA, is not able to complete
the extraction for minsup less than 12%, while our approach remains feasible
until 6% within the time limit. On Protein, fails to complete the extraction for
all values of minsup we considered. These results clearly demonstrate that our
approach competes well with state-of-the-art methods for SPM on large datasets
and achieves scalability while it is a major issue of existing CP approaches.
(c) SPM under size and item constraints. Our third experiment aims at
assessing the interest of pushing simultaneously different types of constraints. We
impose on the PubMed dataset usual constraints such as the minimum frequency
and the minimum size constraints and other useful constraints expressing some
linguistic knowledge such as the item constraint. The goal is to retain sequential
patterns which convey linguistic regularities (e.g., gene - rare disease relation-
ships) [3]. The size constraint allows to remove patterns that are too small w.r.t.
the number of items (number of words) to be relevant patterns. We tested this
constraint with min set to 3. The item constraint imposes that the extracted
patterns must contain the item GENE and the item DISEASE. As no ad hoc
240 A. Kemmar et al.

Fig. 2. Comparing Prefix-Projection with state-of-the-art algorithms for SPM.

Table 4. PP vs. global-p.f under minimum size and item constraints.

CPU times (s) #PROPAGATIONS #NODES
Dataset minsup (%) #PATTERNS
PP global-p.f PP global-p.f PP global-p.f
5 279 6.76 252.36 7878 12234292 2285 4619
4 445 8.81 339.09 12091 16475953 3618 7242
3 799 12.35 535.32 20268 24380096 6271 12643
PubMed
2 1837 20.41 953.32 43088 42055022 13888 27910
1 7187 49.98 2574.42 157899 107978568 52508 107051

method exists for this combination of constraints, we only compare PP with

global-p.f. Fig. 3 shows the CPU times and the number of sequential patterns
extracted with and without constraints. First, pushing simultaneously the two
constraints enables to reduce significantly the number of patterns. Moreover, the
CPU times for PP decrease slightly whereas for global-p.f (with and without
constraints), they are almost the same. This is probably due to the weak com-
munication between the m exists-embedding reified global constraints and the
two constraints. This reduces significantly the quality of the whole filtering. Sec-
ond (see Table 4), when considering the two constraints, PP clearly dominates
global-p.f (speed-up value up to 51.5). Moreover, the number of propagations
performed by PP remains very small as compared to global-p.f. Fig. 3c com-
pares the two methods under the minimum size constraint for different values
of min , with minsup fixed to 1%. Once again, PP is always the most performer
method (speed-up value up to 53.1). These results also confirm what we observed
previously, namely the weak communication between reified global constraints
and constraints imposed on patterns (i.e., size and item constraints).
(d) SPM under regular constraints. Our last experiment compares PP-REG
against two variants of SMA: SMA-1P (SMA one pass) and SMA-FC (SMA Full
Check). Two datasets are considered from [18]: one synthetic dataset (data-
 PREFIX-PROJECTION Global Constraint for Sequential Pattern Mining 241

Fig. 3. Comparing PP with global-p.f under minimum size and item constraints on
PubMed.

Fig. 4. Comparing Prefix-Projection with SMA for SPM under RE constraint.

200k), and one real-life dataset (Protein). For data-200k, we used two RE: RE10 ≡
A∗ B(B|C)D∗ EF ∗ (G|H)I ∗ and RE14 ≡ A∗ (Q|BS ∗ (B|C))D∗ E(I|S)∗ (F |H)G∗ R.
For Protein, we used RE2 ≡ (S|T ) . (R|K) (where . represents any symbol).
Fig. 4 reports CPU-times comparison. On the synthetic dataset, our approach is
very effective. For RE14, our method is more than an order of magnitude faster
than SMA. On Protein, the gap between the 3 methods shrinks, but our method
remains effective. For the particular case of RE2, the Regular constraint can be
substituted by restricting the domain of the first and third variables to {S, T }
and {R, K} respectively (denoted as PP-SRE), thus improving performances.

6 Conclusion

We have proposed the global constraint Prefix-Projection for sequential

pattern mining. Prefix-Projection uses a concise encoding and provides an
efficient filtering based on specific properties of the projected databases, and anti-
monotonicity of the frequency constraint. When this global constraint is integrated
into a CP solver, it enables to handle several constraints simultaneously. Some
of them like size, item membership and regular expression are considered in this
paper. Another point of strength, is that, contrary to existing CP approaches
for SPM, our global constraint does not require any reified constraints nor any
extra variables to encode the subsequence relation. Finally, although Prefix-
Projection is well suited for constraints on sequences, it would require to be
adapted to handle constraints on subsequence relations like gap.
242 A. Kemmar et al.

Experiments performed on several real-life datasets show that our approach

clearly outperforms existing CP approaches and competes well with ad hoc meth-
ods on large datasets and achieves scalability while it is a major issue of CP
approaches. As future work, we intend to handle constraints on set of sequential
patterns such as closedness, relevant subgroup and skypattern constraints.

Acknowledgments. The authors would like to thank the anonymous referees for
their valuable comments. This work is partly supported by the ANR (French Research
National Agency) funded projects Hybride ANR-11-BS002-002.

References
1. Agrawal, R., Srikant, R.: Mining sequential patterns. In: Yu, P.S., Chen, A.L.P.
(eds.) ICDE, pp. 3–14. IEEE Computer Society (1995)
2. Ayres, J., Flannick, J., Gehrke, J., Yiu, T.: Sequential pattern mining using a
bitmap representation. In: KDD 2002, pp. 429–435. ACM (2002)
3. Béchet, N., Cellier, P., Charnois, T., Crémilleux, B.: Sequential pattern mining to
discover relations between genes and rare diseases. In: CBMS (2012)
4. Beldiceanu, N., Contejean, E.: Introducing global constraints in CHIP. Journal of
Mathematical and Computer Modelling 20(12), 97–123 (1994)
5. Coquery, E., Jabbour, S., Saı̈s, L., Salhi, Y.: A SAT-based approach for discovering
frequent, closed and maximal patterns in a sequence. In: ECAI, pp. 258–263 (2012)
6. Fournier-Viger, P., Gomariz, A., Gueniche, T., Soltani, A., Wu, C., Tseng, V.:
SPMF: A Java Open-Source Pattern Mining Library. J. of Machine Learning Resea.
15, 3389–3393 (2014)
7. Garofalakis, M.N., Rastogi, R., Shim, K.: Mining sequential patterns with regular
expression constraints. IEEE Trans. Knowl. Data Eng. 14(3), 530–552 (2002)
8. Guns, T., Nijssen, S., Raedt, L.D.: Itemset mining: A constraint programming
perspective. Artif. Intell. 175(12–13), 1951–1983 (2011)
9. Kemmar, A., Ugarte, W., Loudni, S., Charnois, T., Lebbah, Y., Boizumault, P.,
Crémilleux, B.: Mining relevant sequence patterns with cp-based framework. In:
ICTAI, pp. 552–559 (2014)
10. Li, C., Yang, Q., Wang, J., Li, M.: Efficient mining of gap-constrained subsequences
and its various applications. ACM Trans. Knowl. Discov. Data 6(1), 2:1–2:39 (2012)
11. Métivier, J.P., Loudni, S., Charnois, T.: A constraint programming approach for
mining sequential patterns in a sequence database. In: ECML/PKDD Workshop
on Languages for Data Mining and Machine Learning (2013)
12. Negrevergne, B., Guns, T.: Constraint-based sequence mining using constraint
programming. In: Michel, L. (ed.) CPAIOR 2015. LNCS, vol. 9075, pp. 288–305.
Springer, Heidelberg (2015)
13. Novak, P.K., Lavrac, N., Webb, G.I.: Supervised descriptive rule discovery: A uni-
fying survey of contrast set, emerging pattern and subgroup mining. Journal of
Machine Learning Research 10 (2009)
14. Pei, J., Han, J., Mortazavi-Asl, B., Pinto, H., Chen, Q., Dayal, U., Hsu, M.:
PrefixSpan: Mining sequential patterns by prefix-projected growth. In: ICDE,
pp. 215–224. IEEE Computer Society (2001)
15. Pei, J., Han, J., Wang, W.: Mining sequential patterns with constraints in large
databases. In: CIKM 202, pp. 18–25. ACM (2002)
 PREFIX-PROJECTION Global Constraint for Sequential Pattern Mining 243

16. Pesant, G.: A regular language membership constraint for finite sequences of vari-
ables. In: Wallace, M. (ed.) CP 2004. LNCS, vol. 3258, pp. 482–495. Springer,
Heidelberg (2004)
17. Srikant, R., Agrawal, R.: Mining sequential patterns: Generalizations and perfor-
mance improvements. In: EDBT, pp. 3–17 (1996)
18. Trasarti, R., Bonchi, F., Goethals, B.: Sequence mining automata: A new tech-
nique for mining frequent sequences under regular expressions. In: ICDM 2008,
pp. 1061–1066 (2008)
19. Yan, X., Han, J., Afshar, R.: CloSpan: mining closed sequential patterns in large
databases. In: Barbará, D., Kamath, C. (eds.) SDM. SIAM (2003)
20. Yang, G.: Computational aspects of mining maximal frequent patterns. Theor.
Comput. Sci. 362(1–3), 63–85 (2006)
21. Zaki, M.J.: Sequence mining in categorical domains: Incorporating constraints. In:
Proceedings of the 2000 ACM CIKM International Conference on Information and
Knowledge Management, McLean, VA, USA, November 6–11, pp. 422–429 (2000)
22. Zaki, M.J.: SPADE: An efficient algorithm for mining frequent sequences. Machine
Learning 42(1/2), 31–60 (2001)
 On Tree-Preserving Constraints

Shufeng Kong1 , Sanjiang Li1(B) , Yongming Li2 , and Zhiguo Long1

1
QCIS, FEIT, University of Technology Sydney, Sydney, Australia
{Shufeng.Kong,Zhiguo.Long}@student.uts.edu.au, Sanjiang.Li@uts.edu.au
2
College of Computer Science, Shaanxi Normal University, Xi’an, China
liyongm@snnu.edu.cn

Abstract. Tree convex constraints are extensions of the well-known row

convex constraints. Just like the latter, every path-consistent tree convex
constraint network is globally consistent. This paper studies and com-
pares three subclasses of tree convex constraints which are called chain-,
path- and tree-preserving constraints respectively. While the tractabil-
ity of the subclass of chain-preserving constraints has been established
before, this paper shows that every chain- or path-preserving constraint
network is in essence the disjoint union of several independent connected
row convex constraint networks, and hence (re-)establish the tractability
of these two subclasses of tree convex constraints. We further prove that,
when enforcing arc- and path-consistency on a tree-preserving constraint
network, in each step, the network remains tree-preserving. This ensures
the global consistency of the tree-preserving network if no inconsistency
is detected. Moreover, it also guarantees the applicability of the par-
tial path-consistency algorithm to tree-preserving constraint networks,
which is usually more efficient than the path-consistency algorithm for
large sparse networks. As an application, we show that the class of tree-
preserving constraints is useful in solving the scene labelling problem.

1 Introduction
Constraint satisfaction problems (CSPs) have been widely used in many areas,
such as scene labeling [10], natural language parsing [15], picture processing [16],
and spatial and temporal reasoning [5,14]. Since deciding consistency of CSP
instances is NP-hard in general, lots of efforts have been devoted to identify
tractable subclasses. These subclasses are usually obtained by either restricting
the topology of the underlying graph of the constraint network (being a tree or
having treewidth bounded by a constant) or restricting the type of the allowed
constraints between variables (cf. [17]).
In this paper, we are mainly interested in the second type of restriction.
Montanari [16] shows that path-consistency is sufficient to guarantee that a
network is globally consistent if the relations are all monotone. Van Beek and
Dechter [17] generalise monotone constraints to a larger class of row convex
constraints, which are further generalised to tree convex constraints by Zhang
and Yap [20]. These constraints also have the nice property that every path-
consistent constraint network is globally consistent.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 244–261, 2015.
DOI: 10.1007/978-3-319-23219-5 18
 On Tree-Preserving Constraints 245

However, neither row convex constraints nor tree convex constraints are
closed under composition and intersection, the main operations of path-consistent
algorithms. This means enforcing path-consistency may destroy row and tree-
convexity. Deville et al. [6] propose a tractable subclass of row convex constraints,
called connected row convex (CRC) constraints, which are closed under composi-
tion and intersection. Zhang and Freuder [18] also identify a tractable subclass for
tree convex constraints, called locally chain convex and strictly union closed con-
straints. They also propose the important notion of consecutive constraints. Kumar
[13] shows that the subclass of arc-consistent consecutive tree convex (ACCTC)
constraints is tractable by providing a polynomial time randomised algorithm. But,
for the ACCTC problems, “it is not known whether there are efficient deterministic
algorithms, neither is it known whether arc- and path-consistency ensures global
consistency on those problems.” [18]
In this paper, we study and compare three subclasses of tree convex con-
straints which are called, respectively, chain-, path- and tree-preserving con-
straints. Chain-preserving constraints are exactly “locally chain convex and
strictly union closed” constraints and ACCTC constraints are strictly contained
in the subclass of tree-preserving constraints. We first show that every chain-
or path-preserving constraint network is in essence the disjoint union of several
independent CRC constraint networks and then prove that enforcing arc- and
path-consistency on a tree-preserving constraint network ensures global consis-
tency. This provides an affirmative answer to the above open problem raised
in [18]. Note also that our result is more general than that of Kumar [13] as
we do not require the constraint network to be arc-consistent. Moreover, when
enforcing arc- and path-consistent on a tree-preserving constraint network, in
each step, the network remains tree-preserving. This guarantees the applicabil-
ity of the partial path-consistency algorithm [2] to tree-preserving constraint
networks, which is usually more efficient than the path-consistency algorithm
for large sparse networks. We further show that a large subclass of the trihedral
scene labelling problem [10,12] can be modelled by tree-preserving constraints.
In the next section, we introduce basic notations and concepts that will be
used throughout the paper. Chain-, path-, and tree-preserving constraints are
discussed in Sections 3, 4, and 5, respectively. Application of tree-preserving
constraints in the scene labelling problem is shown in Section 6. Section 7 briefly
discusses the connection with majority operators [11] and concludes the paper.

2 Preliminaries

Let D be a domain of a variable x. A graph structure can often be associated

to D such that there is a bijection between the vertices in the graph and the
values in D. If the graph is connected and acyclic, i.e. a tree, then we say it is
a tree domain of x. Tree domains arise naturally in e.g. scene labeling [18] and
combinatorial auctions [4]. We note that, in this paper, we have a specific tree
domain Dx for each variable x.
246 S. Kong et al.

In this paper, we distinguish between tree and rooted tree. Standard notions
from graph theory are assumed. In particular, the degree of a node a in a graph
G, denoted by deg(a), is the number of neighbors of a in G.

Definition 1. A tree is a connected graph without any cycle. A tree is rooted if

it has a specified node r, called the root of the tree. Given a tree T , a subgraph I
is called a subtree of T if I is connected. An empty set is a subtree of any tree.
Let T be a (rooted) tree and I a subtree of T . I is a path ( chain, resp.) in T
if each node in I has at most two neighbors (at most one child, resp.) in I. Given
two nodes p, q in T , the unique path that connects p to q is denoted by πp,q .
Suppose a is a node of a tree T . A branch of a is a connected component of
T \ {a}.

Throughout this paper, we always associate a subtree with its node set.
Definition 2. A binary constraint has the form (xδy), where x, y are two vari-
ables with domains Dx and Dy and δ is a binary relation from Dx to Dy , or
δ ⊆ Dx × Dy . For simplicity, we often denote by δ this constraint. A value
u ∈ Dx is supported if there exists a value v in Dy s.t. (u, v) ∈ δ. In this case,
we say v is a support of u. We say a subset F of Dx is unsupported if every
value in F is not supported. Given A ⊆ Dx , the image of A under δ is defined
as δ(A) = {b ∈ Dy : (∃a ∈ A)(a, b) ∈ δ}. For A = {a} that contains only one
value, without confusion, we also use δ(a) to represent δ({a}).
A binary constraint network consists of a set of variables V = {x1 , x2 , ..., xn }
with a finite domain Di for each variable xi ∈ V , and a set Δ of binary constraints
over the variables of V . The usual operations on relations, e.g., intersection (∩),
composition (◦), and inverse (−1 ), are applicable to constraints. As usual, we
assume that there is at most one constraint for any ordered pair of variables
(x, y). Write δxy for this constraint if it exists. In this paper, we always assume
δxy is the inverse of δyx , and if there is no constraint for (x, y), we assume δxy is
the universal constraint.

Definition 3. [8, 9] A constraint network Δ over n variables is k-consistent iff

any consistent instantiation of any distinct k − 1 variables can be consistently
extended to any k-th variable. We say Δ is strongly k-consistent iff it is j-
consistent for all j ≤ k; and say Δ is globally consistent if it is strongly n-
consistent. 2- and 3-consistency are usually called arc- and path-consistency
respectively.

Definition 4. Let x, y be two variables with finite tree domains Tx = (Dx , Ex )

and Ty = (Dy , Ey ) and δ a constraint from x to y. We say δ, w.r.t. Tx and Ty , is

- tree convex if the image of every value a in Dx (i.e. δ(a)) is a (possibly empty)
subtree of Ty ;
- consecutive if the image of every edge in Tx is a subtree in Ty ;
- path-preserving if the image of every path in Tx is a path in Ty .
- tree-preserving if the image of every subtree in Tx is a subtree in Ty .
 On Tree-Preserving Constraints 247

In case Tx and Ty are rooted, we say δ, w.r.t. Tx and Ty , is

- chain-preserving if the image of every chain in Tx is a chain in Ty .
Chain-preserving constraints are exactly those “locally chain convex and strictly
union closed” constraints defined in [18].
CRC constraints are special tree convex constraints defined over chain
domains. The following definition of CRC constraints is equivalent to the one
given in [6].
Definition 5. Let x, y be two variables with finite tree domains Tx and Ty ,
where Tx and Ty are chains. A constraint δ from x to y is connected row convex
(CRC), w.r.t. Tx and Ty , if both δ and δ −1 are chain-preserving.
The class of CRC constraints is tractable and closed under intersection, inverse,
and composition [6].
Definition 6. A binary constraint network Δ over variables in V and tree
domains Tx (x ∈ V ) is called tree convex, chain-, path-, or tree-preserving if
every constraint δ ∈ Δ is tree convex, chain-, path-, or tree-preserving, respec-
tively. A CRC constraint network is defined similarly.
Proposition 1. Every chain-, path-, or tree-preserving constraint (network) is
consecutive and every path-preserving constraint (network) is tree-preserving.
Moreover, every arc-consistent consecutive tree convex (ACCTC) constraint
(network) is tree-preserving.
Not every consecutive tree convex constraint (or chain-preserving constraint)
is tree-preserving, but such a constraint becomes tree-preserving if it is arc-
consistent.
Lemma
n 1. [20] Let T be a tree and suppose ti (i = 1, .., n) are subtrees of T .
Then i=1 ti is nonempty iff ti ∩ tj is nonempty for every 1 ≤ i = j ≤ n.
Lemma 2. Let T be a tree and t, t subtrees of T . Suppose {u, v} is an edge in
T . If u ∈ t and v ∈ t , then t ∪ t is a subtree of T ; if, in addition, u ∈ t and
v ∈ t, then t ∩ t = ∅.
Using Lemma 1, Zhang and Yap [20] proved
Theorem 1. A tree-convex constraint network is globally consistent if it is path-
consistent.

3 Chain-Preserving Constraints
Zhang and Freuder [18] have proved that any consistent chain-preserving network
can be transformed to an equivalent globally consistent network by enforcing arc-
and path-consistency. This implies that the class of chain-preserving constraints
is tractable. We next show that every chain-preserving constraint network Δ can
be uniquely divided into a small set of k CRC constraint networks Δ1 , ..., Δk s.t.
Δ is consistent iff at least one of Δi is consistent.
We first recall the following result used in the proof of [18, Theorem 1].
248 S. Kong et al.

Proposition 2. Let Δ be a chain-preserving constraint network over tree

domains Tx (x ∈ V ). If no inconsistency is detected, then Δ remains chain-
preserving after enforcing arc-consistency.
We note that these tree domains over variables in V may need adjustment
in the process of enforcing arc-consistency. Here by adjustment we mean adding
edges to the tree structure so that it remains connected when unsupported values
are deleted.
Definition 7. Let T be a tree with root. A chain [a, a∗ ] in T is called an irre-
ducible perfect chain (ip-chain) if (i) a is the root or has one or more siblings;
(ii) a∗ is a leaf node or has two or more children; and (iii) every node in [a, a∗ )
has only one child.
Note that it is possible that a = a∗ . In fact, this happens when a is the root or
has one or more siblings and has two or more children. An ip-chain as defined
above is a minimum chain which satisfies (1) in the following lemma.
Lemma 3. Suppose δxy and δyx are arc-consistent and chain-preserving w.r.t.
rooted trees Tx and Ty . Assume [a, a∗ ] ⊆ Tx is an ip-chain. Then

δyx (δxy ([a, a∗ ])) = [a, a∗ ] (1)

and δxy ([a, a∗ ]) is also an ip-chain in Ty .

Proof. W.l.o.g., we suppose â is the parent of a, a is a sibling of a, and
a1 , a2 , ..., ak (k ≥ 2) are the children of a∗ .
Because δxy and δyx are arc-consistent and chain-preserving, δxy ([a, a∗ ]) is a
non-empty chain in Ty , written [b, b∗ ], and so is δyx ([b, b∗ ]). Suppose δyx ([b, b∗ ])
is not [a, a∗ ]. This implies that either â or one of a1 , a2 , ..., ak is in δyx ([b, b∗ ]).
Suppose â ∈ δyx ([b, b∗ ]). Then there exists b̂ ∈ [b, b∗ ] such that (â, b̂) ∈ δxy .
By b̂ ∈ [b, b∗ ] = δxy ([a, a∗ ]), we have a+ ∈ [a, a∗ ] s.t. (a+ , b̂) ∈ δxy . Therefore,
[â, a+ ] is contained in δyx (δxy ({â})). Recall that a is a sibling of a. Because
δyx (δxy ([â, a ])) contains â, a , a+ , it cannot be a chain in Tx . A contradiction.
Therefore, â ∈ δyx ([b, b∗ ]).
Suppose, for example, a1 ∈ δyx ([b, b∗ ]). Then there exist b ∈ [b, b∗ ] s.t.
(a1 , b ) ∈ δxy and ā ∈ [a, a∗ ] s.t. (ā, b ) ∈ δxy . We have δyx (δxy ({ā}) ⊇ [ā, a1 ] and
δyx (δxy ([ā, a2 ]) contains {ā, a1 , a2 }, which is not a subset of a chain. Therefore,
ai ∈ δyx ([b, b∗ ]).
So far, we have proved δyx (δxy ([a, a∗ ])) = [a, a∗ ]. We next show [b, b∗ ] is also
an ip-chain. First, we show every node in [b, b∗ ) has only one child. Suppose not
and b ∈ [b, b∗ ) has children b1 , b2 with b1 ∈ (b , b∗ ]. Since δxy ([â, a∗ ]) is a chain
that contains [b, b∗ ], we know (â, b2 ) is not in δxy . Furthermore, as δyx ({b , b2 }) is
a chain in Tx and the image of b2 is disjoint from [a, a∗ ], we must have (ai , b2 ) ∈
δxy for some child ai of a∗ . Note that then δxy ([a, ai ]) contains [b, b∗ ] and b2 and
thus is not a chain. This contradicts the chain-preserving property of δxy . Hence,
every node in [b, b∗ ) has only one child. In other words, [b, b∗ ] is contained in an
ip-chain [u, v].
 On Tree-Preserving Constraints 249

By the result we have proved so far, we know δxy (δyx ([u, v])) = [u, v] and
δyx ([u, v]) is contained in an ip-chain in Tx . Because [a, a∗ ] = δyx ([b, b∗ ]) ⊆
δyx ([u, v]) is an ip-chain, we know δyx ([u, v]) is exactly [a, a∗ ]. Therefore, we
have [u, v] = δxy (δyx ([u, v])) = δxy ([a, a∗ ]) = [b, b∗ ]. This proves that [b, b∗ ] is an
ip-chain in Ty .

Using the above result, we can break Tx into a set of ip-chains by deleting
the edges from each node a to its children if a has two or more children. Write
Ix for the set of ip-chains of Tx . Similar operation and notation apply to Ty . It
is clear that two different ip-chains in Ix are disjoint and δxy naturally gives rise
to a bijection from Ix to Iy .
Lemma 4. Suppose Δ is an arc-consistent and chain-preserving constraint net-
work over tree domains Tx (x ∈ V ). Fix a variable x ∈ V and let Ix = {Ix1 , ..., Ixl }
be the set of ip-chains of Tx . Then, for every y = x in V , the set of ip-chains
in Ty is {δxy (Ix1 ), δxy (Ix2 ), ..., δxy (Ixl )}. Write Δi for the restriction of Δ to Ixi .
Then each Δi is a CRC constraint network and Δ is consistent iff at least one
Δi is consistent.
The following result asserts that the class of chain-preserving constraints is
tractable.
Theorem 2. Let Δ be a chain-preserving constraint network. If no inconsis-
tency is detected, then enforcing arc- and path-consistency determines the con-
sistency of Δ and transforms Δ into a globally consistent network.
Proof. First, by Proposition 2, we transform Δ into an arc-consistent and
chain-preserving constraint network if no inconsistency is detected. Second, by
Lemma 4, we reduce the consistency of Δ to the consistency of the CRC con-
straint networks Δ1 , ..., Δl . By [6], we know enforcing path-consistency trans-
forms a CRC constraint network into a globally consistent one if no inconsistency
is detected. If enforcing arc- and path-consistency does not detect any inconsis-
tency, then the result is a set of at most l globally consistent CRC networks Δi ,
the union of which is globally consistent and equivalent to Δ.

Lemma 4 also suggests that we can use the variable elimination algorithm for
CRC constraints [19] to more efficiently solve chain-preserving constraints.

4 Path-Preserving Constraints
At first glance, path-preserving constraints seem to be more general than chain-
preserving constraints, but Fig. 1(a,b) show that they are in fact incomparable.
We show the class of path-preserving constraints is also tractable by estab-
lishing its connection with CRC constraints.
We have the following simple results.
Lemma 5. Suppose δxy and δyx are path-preserving (tree-preserving) w.r.t. tree
 
domains Tx and Ty . Let t be a subtree of Tx and δxy and δyx the restrictions of
 
δxy and δyx to t. Then both δxy and δyx are path-preserving (tree-preserving).
250 S. Kong et al.

Fig. 1. (a) A chain- but not path-preserving constraint; (b) A path- but not chain-
preserving constraint; (c) A tree-preserving but neither path- nor chain-preserving
constraint.

Lemma 6. Suppose δxy is nonempty and path-preserving (tree-preserving)

w.r.t. tree domains Tx and Ty . If v ∈ Ty has no support in Tx under δyx , then
all supported nodes of Ty are in the same branch of v. That is, every node in
any other branch of v is not supported under δyx .

Proof. Suppose a, b are two supported nodes in Ty . There exist u1 , u2 in Tx s.t.

−1
u1 ∈ δyx (a) and u2 ∈ δyx (b). By δyx = δxy , we have a ∈ δxy (u1 ) and b ∈
δxy (u2 ). Hence a, b ∈ δxy (πu1 ,u2 ). Since δxy is path-preserving (tree-preserving),
δxy (πu1 ,u2 ) is a path (tree) in Ty . If a, b are in two different branches of v, then
πa,b must pass v and hence we must have v ∈ δxy (πu1 ,u2 ). This is impossible as
v has no support.


It is worth noting that this lemma does not require δyx to be path- or tree-
preserving.
The following result then follows directly.
Proposition 3. Let Δ be a path-preserving (tree-preserving) constraint network
over tree domains Tx (x ∈ V ). If no inconsistency is detected, then Δ remains
path-preserving (tree-preserving) after enforcing arc-consistency.

Proof. Enforcing arc-consistency on Δ only removes values which have no sup-

port under some constraints. For any y ∈ V , if v is an unsupported value in Ty ,
then, by Lemma 6, every supported value of Ty is located in the same branch of v.
Deleting all these unsupported values from Ty , we get a subtree t of Ty . Applying
Lemma 5, the restricted constraint network to t remains path-preserving (tree-
preserving).


Definition 8. Let T be a tree. A path π in T is maximal if there exists no path

π  in T that strictly contains π.

We need three additional lemmas to prove the main result.

Lemma 7. Suppose δxy and δyx are arc-consistent and path-preserving w.r.t.
tree domains Tx and Ty . If π is a maximal path in Tx , then δxy (π) is a maximal
path in Ty .
 On Tree-Preserving Constraints 251

Lemma 8. Suppose δxy and δyx are arc-consistent and path-preserving w.r.t. Tx
and Ty . Assume a is a node in Tx with deg(a) > 2. Then there exists a unique
node b ∈ Ty s.t. (a, b) ∈ δxy . Moreover, deg(a) = deg(b).
Proof. Suppose π = a0 a1 ...ak is a maximal path in Tx and π ∗ = b0 b1 ...bl is its
image under δxy in Ty . W.l.o.g. we assume k, l ≥ 1. Suppose ai is a node in π
s.t. deg(ai ) > 2 and a ∈ π is another node in Tx s.t. {ai , a } is an edge in Tx .
Suppose δxy (ai ) = [bj , bj  ] and j  > j.
Because π is a maximal path and π ∗ is its image, we know δxy (a ) ∩ π ∗ = ∅.
Consider the edge {a , ai }. Since δxy ({a , ai }) is a path in Ty , there exists a
node b ∈ δxy (a ) s.t. either {b , bj } or {b , bj  } is an edge in Ty . Suppose w.l.o.g.
{b , bj  } is in Ty . Note that π ∗ = [b0 , bl ] is contained in the union of δxy ([a , a0 ])
and δxy ([a , ak ]). In particular, bl is in either δxy ([a , a0 ]) or δxy ([a , ak ]). Let
us assume bl ∈ δxy ([a , a0 ]). Then bl , bj , bj  , b (which are not on any path) are
contained in the path δxy ([a , a0 ]), a contradiction. Therefore, our assumption
that δxy (ai ) = [bj , bj  ] and j  > j is incorrect. That is, the image of ai under δxy
is a singleton, say, {bj }. We next show deg(bj ) = deg(ai ).
Because δxy (ai ) = {bj }, the image of each neighbor of ai in Tx must contain
a neighbor of bj , as δxy is path-preserving. Moreover, two different neighbors
ai , ai of ai cannot map to the same neighbor bj of bj . This is because the image
of bj under δyx , which is a path in Tx , contains ai and ai , and hence also
contains ai . This contradicts the assumption δxy (ai ) = {bj }. This shows that
deg(ai ) = deg(bj ).

Definition 9. Let T be a tree. A path π from a to a∗ in T is called an irreducible
perfect path (ip-path) if (i) every node on path π has degree 1 or 2; and (ii) any
neighbour of a (or a∗ ) that is not on π has degree 3 or more.
Let Fx = {a ∈ Tx : deg(a) > 2} and Fy = {b ∈ Ty : deg(b) > 2}. Then δxy ,
when restricted to Fx , is a bijection from Fx to Fy . Removing all edges incident
to a node in Fx , we obtain a set of pairwise disjoint paths in Tx . These paths are
precisely the ip-paths of Tx . Write Px for this set. Then δxy induces a bijection
from Px to Py .
Lemma 9. Suppose Δ is an arc-consistent and path-preserving constraint net-
work over tree domains Tx (x ∈ V ). Fix a variable x ∈ V and let Px =
{πx1 , ..., πxl } be the set of ip-paths in Tx . Then, for every y = x, the set of ip-
paths in Ty is {δxy (πx1 ), ..., δxy (πxl )}. Write Δi for the restriction of Δ to πxi .
Then each Δi is a CRC constraint network and Δ is consistent iff at least one
Δi is consistent.
Thus the class of path-preserving constraints is tractable.
Theorem 3. Let Δ be a path-preserving constraint network. If no inconsistency
is detected, then enforcing arc- and path-consistency determines the consistency
of Δ and transforms Δ into a globally consistent network.
The proof is analogous to that of Theorem 2. Lemma 9 suggests that we can use
the variable elimination algorithm for CRC constraints [19] to more efficiently
solve path-preserving constraints.
252 S. Kong et al.

5 Tree-Preserving Constraints
It is easy to see that every arc-consistent chain- or path-preserving constraint is
tree-preserving, but Fig. 1(c) shows that the other direction is not always true.
In this section, we show that the class of tree-preserving constraints is
tractable. Given a tree-preserving constraint network Δ, we show that, when
enforcing arc- and path-consistency on Δ, in each step, the network remains
tree-preserving. Hence, enforcing arc- and path-consistency on Δ will transform
it to an equivalent globally consistent network if no inconsistency is detected.
Moreover, we show that the partial path-consistency algorithm (PPC) of [2] is
applicable to tree-preserving constraint networks. PPC is more efficient than
path-consistency algorithm for large sparse constraints.

5.1 Enforcing Arc- and Path-Consistency Preserves Tree-Preserving

Unlike CRC and chain-preserving constraints, removing a value from a domain

may change the tree-preserving property of a network. Instead, we need to
remove a ‘trunk’ from the tree domain or just keep one branch.

Definition 10. Suppose a = b are two nodes of a tree T that are not neighbors.
The trunk between a, b, written Mab , is defined as the connected component of
T \{a, b} which contains all internal nodes of πa,b (see Fig.2). The M-contraction
of T by Ma,b , denoted by T  Ma,b , is the tree obtained by removing nodes in
Ma,b and adding an edge {a, b} to T .

Fig. 2. Mab is a trunk of tree T .

Lemma 10. Suppose δxy and δyx are arc-consistent and tree-preserving w.r.t.
tree domains Tx and Ty . Suppose a, b are two nodes in Tx s.t. δxy (a) ∪ δxy (b)
is not connected in Ty . Then there exist r, s ∈ Ty s.t. r ∈ δxy (a), s ∈ δxy (b),
and δyx (Mr,s ) ⊆ Ma,b . Let Ty∗ be the domain obtained by deleting from Ty all
nodes v s.t. δyx (v) ⊆ Ma,b . Then Ty∗ becomes a tree if we add the edge {r, s}.
Moreover, δxy and δyx remain arc-consistent and tree-preserving when restricted
to Tx  Ma,b and Ty∗ .
 On Tree-Preserving Constraints 253

Proof. Choose r ∈ δxy (a) and s ∈ δxy (b) such that the path πr,s from r to s in
Ty is a shortest one among {πr ,s : r ∈ δxy (a), s ∈ δxy (b)}. In particular, we
have πr,s ∩ (δxy (a) ∪ δxy (b)) = {r, s}. We assert that the image of every node v
in Mr,s under δyx is contained in Ma,b . Suppose otherwise and there exists u in
Tx \ Ma,b s.t. (u, v) ∈ δxy . Assume that u is in the same connected component as
a. Since the subtree δyx (πv,s ) contains u and b, it also contains a. This implies
that there is a node v  on πv,s which is in δxy (a). This is impossible as v ∈ Mr,s
and δxy (a) ∩ πr,s = {r}. Therefore δyx (v) ⊆ Ma,b for any v ∈ Mr,s . Hence
δyx (Mr,s ) ⊆ Ma,b holds.
It is clear that, when restricted to Tx  Ma,b and Ty  Mr,s , δxy ({a, b})
is connected and so is δyx ({r, s}). For any other edge {a , b } in Tx  Ma,b , by
δyx (Mr,s ) ⊆ Ma,b , δxy ({a , b }) ∩ Mr,s = ∅ and the image of {a , b } is unchanged
(hence connected) after the M-contraction of Ty . This shows that δxy is consec-
utive when restricted to Tx  Ma,b . Furthermore, since every node in Tx  Ma,b
is supported in Ty  Mr,s , we know δxy is also tree-preserving when restricted
to Tx  Ma,b .
It is possible that there is a node v ∈ Ty  Mr,s s.t. δyx (v) ⊆ Ma,b . We assert
that any v like this has at most one branch in Ty \Mr,s s.t. there is a node v  in the
branch which is supported under δyx by a node in Tx \ Ma,b . Because δxy is tree-
preserving when restricted to Tx Ma,b , this follows immediately from Lemma 6.
This implies that, if we remove all these nodes v s.t. δyx (v) ⊆ Ma,b from Ty Mr,s ,
the domain is still connected. As a consequence, the two constraints remain arc-
consistent and tree preserving when restricted to Tx  Ma,b and Ty∗ .


In general, we have

Lemma 11. Let Δ be an arc-consistent and tree-preserving constraint network

over tree domains Tx (x ∈ V ). Suppose x ∈ V and Ma,b is a trunk in Tx . When
restricted to Tx  Ma,b and enforcing arc-consistency, Δ remains tree-preserving
if we modify each Ty (y ∈ V ) by deleting all unsupported nodes and adding some
edges.

Proof (Sketch). The result follows from Lemmas 10 and 5. One issue we need
to take care of is how two trunks interact. Suppose x, y, z are three different
variables and M , M  are trunks to be contracted from Tx and Ty respectively.
Applying Lemma 10 to the constraints between x and z and, separately, to the
constraints between y and z, we get two different trunks, say Ma,b and Mc,d ,
to be contracted from the same tree domain Tz . Can we do this (i.e. applying
Lemma 10) one by one? Does the order of the contractions matter? To answer
these questions, we need to know what is exactly the union of two trunks. There
are in essence ten configurations as shown in Fig. 3. The union of two trunks can
be the whole tree, a branch, a trunk, or two disjoint trunks. If the union is the
whole tree, then the network is inconsistent; if it is a branch, then we can remove
it directly; if it is a trunk or two disjoint trunks, then we can use Lemma 10 to
contract them one by one in either order.

254 S. Kong et al.

Fig. 3. Possible configurations of trunks Ma,b and Mc,d .


Lemma 12. Assume δxy and δxy are two arc-consistent and tree-preserving
∗ 
constraints w.r.t. trees Tx and Ty . Let δxy = δxy ∩ δxy . Suppose u ∈ Tx and
 ∗
δxy (u) ∩ δxy (u) = ∅. Then the supported values of δxy in Tx are in at most two
branches of u.
∗
Proof. Suppose u1 , u2 , u3 are three supported values of δxy in Tx that are in

three different branches of u. Take wi ∈ δxy (ui ) ∩ δxy (ui ). For each i, we have

either wi ∈ δxy (u) or wi ∈ δxy (u). Recall that πwi ,wj denotes the unique path
πwi ,wj that connects wi to wj (1 ≤ i = j ≤ 3). There are two subcases. (1) One
node is on the path that connects the other two. Suppose w.l.o.g. w3 is between
w1 and w2 . If w3 ∈ δxy (u), then there exist v1 ∈ πw1 ,w3 and v2 ∈ πw3 ,w2 s.t.
v1 , v2 ∈ δxy (u). This is because the image of πw1 ,w3 and the image of πw3 ,w2
(under δyx ) both contain u. That is, v1 , v2 ∈ δxy (u) but w3 ∈ δxy (u). Since
δxy (u) is a subtree and w3 is on the path πv1 ,v2 , this is impossible. The case when

w3 ∈ δxy (u) is analogous. (2) The three nodes are in three different branches
of a node w. In this case, we note there exists a node between any path πwi ,wj
which is a support of u under δxy . It is easy to see that w itself is a support of u

under δxy . Similarly, we can show w is also a support of u under δxy . Since both
∗
subcases lead to a contradiction, we know the supported values of δxy are in at
most two branches of u.


Actually, this lemma shows


Corollary 1. Assume δxy and δxy are two arc-consistent and tree-preserving

constraints w.r.t. trees Tx and Ty . Then those unsupported values of δxy ∩ δxy in
Tx are in a unique set of pairwise disjoint branches and trunks.
Similar to Lemma 10, we have
 On Tree-Preserving Constraints 255


Lemma 13. Suppose δxy and δxy are arc-consistent and tree-preserving con-
 ∗ 
straints w.r.t. trees Tx and Ty and so are δyx and δyx . Let δxy = δxy ∩ δxy .
∗ ∗
Assume {u, v} is an edge in Tx s.t. δxy (u) ∪ δxy (v) is disconnected in Ty . Then
∗ ∗
there exist r ∈ δxy (u) and s ∈ δxy (v) s.t. every node in Mr,s is unsupported
∗
under δyx .
∗ ∗
Proof. Write Tr = δxy (u) and Ts = δxy (v). Clearly, Tr and Ts are nonempty
subtrees of Ty . Since they are disconnected, there exist r ∈ Tr , s ∈ Ts s.t.
πr,s ∩ (Tr ∪ Ts ) = {r, s} (see Fig. 4 for an illustration). Write A = δxy (u),
 
B = δxy (v), C = δxy (u) and D = δxy (v). We show every node in Mr,s is not
∗
supported under δyx .
Suppose w is an arbitrary internal node on πr,s . We first show w is not
∗
supported under δyx . Note w ∈ A ∪ B, w ∈ C ∪ D, w ∈ A ∩ C, and w ∈ B ∩ D.
There are two cases according to whether w ∈ A. If w ∈ A, then we have w ∈ C,
w ∈ D, and w ∈ B. If w ∈ A, then we have w ∈ B, w ∈ D, and w ∈ C. Suppose
w.l.o.g. w ∈ A. By w ∈ A = δxy (u), we have u ∈ δyx (w); by w ∈ B = δxy (v), we
 
have v ∈ δyx (w). Similarly, we have u ∈ δyx (w) and v ∈ δyx (w). Thus subtree
 ∗
δyx (w) is disjoint from subtree δyx (w). This shows δyx (w) = ∅ and hence w is
∗
not supported under δyx .
Second, suppose w1 is an arbitrary node in Mr,s s.t. w1 is in a different branch
of w to r and s, i.e. πw,w1 ∩ (Tr ∪ Ts ) = ∅. We show w1 is not supported under
∗
δyx either.
Again, we assume w ∈ A. In this case, we have u ∈ δyx (w) ⊆ δyx (πw,w1 )
 
and v ∈ δyx (w) ⊆ δyx (πw,w1 ). As πw,w1 ∩ (Tr ∪ Ts ) = ∅, we have πw,w1 ∩ Tr =
πw,w1 ∩ A ∩ C = ∅. As πw,w1 ∩ A = ∅ and A ∩ C = ∅, by Lemma 1, we must
 
have πw,w1 ∩ δxy (u) = ∅. This shows u ∈ δyx (πw,w1 ). Similarly, we can show

v ∈ δyx (πw,w1 ). Thus subtree δyx (πw,w1 ) is disjoint from subtree δyx (πw,w1 ) and,
∗ ∗
hence, δyx (πw,w1 ) = ∅. This proves that w1 is not supported under δyx either.
In summary, every node in Mr,s is unsupported.


Fig. 4. Illustration of proof of Lemma 13.

Proposition 4. [18] Assume δxz and δzy are two tree-preserving constraints
w.r.t. trees Tx , Ty , and Tz . Then their composition δxz ◦ δzy is tree-preserving.
256 S. Kong et al.

At last, we give the main result of this section.

Theorem 4. Let Δ be a tree-preserving constraint network. If no inconsistency

is detected, then enforcing arc- and path-consistency determines the consistency
of Δ and transforms Δ into a globally consistent network.

Proof. If we can show that Δ is still tree-preserving after enforcing arc and
path-consistency, then by Theorem 1 the new network is globally consistent if
no inconsistency is detected.
By Proposition 3, Δ remains tree-preserving after enforcing arc-consistency.
To enforce path-consistency on Δ, we need to call the following updating rule

δxy ← δxy ∩ (δxz ◦ δzy ) (2)

for x, y, z ∈ V until the network is stable.

Suppose Δ is arc-consistent and tree-preserving w.r.t. trees Tx for x ∈ V
before applying (2). Note that if δ ∗ = δxy ∩ (δxz ◦ δzy ) (as well as its converse)
is arc-consistent, then δ ∗ (u) is nonempty for any node u in Tx . By Corollary 1,
no branches or trunks need to be pruned in either Tx or Ty . Furthermore, by
Lemma 13, δ ∗ (u) ∪ δ ∗ (v) is connected for every edge {u, v} in Tx as there are no
unsupported nodes in Ty under the converse of δ ∗ . Therefore δ ∗ is arc-consistent
and consecutive, hence, tree-preserving.
If δ ∗ is not arc consistent, then we delete all unsupported values from Tx
and Ty and enforce arc-consistency on Δ. If no inconsistency is detected then
we have an updated arc-consistent and tree-preserving network by Lemma 11.
Still write Δ for this network and recompute δ ∗ = δxy ∩ (δxz ◦ δzy ) and repeat
the above procedure until either inconsistency is detected or δ ∗ is arc-consistent.
Note that, after enforcing arc-consistency, the composition δxz ◦ δzy may have
changed.
Once arc-consistency of δ ∗ is achieved, we update δxy with δ ∗ and continue
the process of enforcing path-consistency until Δ is path-consistent or an incon-
sistency is detected.


5.2 Partial Path-Consistency

The partial path-consistency (PPC) algorithm was first proposed by Bliek and
Sam-Haroud [2]. The idea is to enforce path consistency on sparse graphs by
triangulating instead of completing them. Bliek and Sam-Haroud demonstrated
that, as far as CRC constraints are concerned, the pruning capacity of path
consistency on triangulated graphs and their completion are identical on the
common edges.
An undirected graph G = (V, E) is triangulated or chordal if every cycle of
length greater than 3 has a chord, i.e. an edge connecting two non-consecutive
vertices of the cycle. For a constraint network Δ = {vi δij vj : 1 ≤ i, j ≤ n}
over V = {v1 , ..., vn }, the constraint graph of Δ is the undirected graph G(Δ) =
(V, E(Δ)), for which we have (vi , vj ) ∈ E(Δ) iff δij is not a universal constraint.
 On Tree-Preserving Constraints 257

Given a constraint network Δ and a graph G = (V, E), we say Δ is partial path-
consistent w.r.t. G iff for any 1 ≤ i, j, k ≤ n with (vi , vj ), (vj , vk ), (vi , vk ) ∈ E
we have δik ⊆ δij ◦ δjk [2].
Theorem 5. Let Δ be a tree-preserving constraint network. Suppose G = (V, E)
is a chordal graph such that E(Δ) ⊆ E. Then enforcing partial path-consistency
on G is equivalent to enforcing path-consistency on the completion of G, in the
sense that the relations computed for the constraints in G are identical.
Proof. The proof is similar to the one given for CRC constraints [2, Theorem 3].
This is because, (i) when enforcing arc- and path-consistency on a tree-preserving
constraint network, in each step, we obtain a new tree-preserving constraint
network; and (ii) path-consistent tree convex constraint networks are globally
consistent.

Remark 1. Note that our definition and results of tree-preserving constraints
can be straightforwardly extended to domains with acyclic graph structures
(which are connected or not). We call such a structure a forest domain. Given a
tree-preserving constraint network Δ over forest domains F1 , ..., Fn of variables
v1 , ..., vn . Suppose Fi consists of trees ti,1 , ..., ti,ki . Note that the image of each
tree, say ti,1 , of Fi under constraint Rij is a subtree t of Fj . Assume t is con-
tained in the tree tj,s of forest Fj . Then the image of tj,s under constraint Rji
is a subtree of ti,1 . This establishes, for any 1 ≤ i = j ≤ n, a 1-1 correspondence
between trees in Fi and trees in Fj if the image of each tree is nonempty. In
this way, the consistency of Δ is reduced to the consistency of several parallel
tree-preserving networks over tree domains.

6 Tree-Preserving Constraints and the Scene Labelling

Problem
The scene labelling problem [10] is a classification problem where all edges in
a line-drawing picture have to be assigned a label describing them. The scene
labelling problem is NP-complete in general. This is true even in the case of the
trihedral scenes, i.e. scenes where no four planes share a point [12].
Labels used in the scene labelling problem are listed as follows:
‘+’ The edge is convex which has both of its corresponding planes visible;
‘−’ The edge is concave which has both of its corresponding planes visible;
‘→’ Only one plane associated with the edge is visible, and when one moves in
the direction indicated by the arrow, the pair of associated planes is to the
right.
In the case of trihedral scenes, there are only four basic ways in which three
plane surfaces can come together at a vertex [10]. A vertex projects in the picture
into a ‘V ’, ‘W ’, ‘Y ’ or ‘T ’-junction (each of these junction-types may appear
with an arbitrary rotation in a given picture). A complete list of the labelled
line configurations that are possible in the vicinity of a node in a picture is given
in Fig. 5.
258 S. Kong et al.

Fig. 5. Possible labelled line configurations of a junction in a picture and their corre-
sponding forest structures.

In this section, we show that (i) every instance of the trihedral scene labelling
problem can be modelled by a tree convex constraint network; (ii) a large sub-
class of the trihedral scene labelling problem can be modelled by tree-preserving
constraints; (iii) there exists a scene labelling instance which can be modelled
by tree-preserving constraints but not by chain- or CRC constraints.
A CSP for the scene labelling problem can be formulated as follows. Each
junction in the line-drawing picture is a variable. The domains of the vertices are
the possible configurations as shown in Fig. 5. The constraints between variables
are simply that, if two variables share an edge, then the edge must be labeled
the same at both ends.
Proposition 5. Every instance of the trihedral scene labelling problem can be
modelled by a tree convex constraint network. Furthermore, there are only 39
possible configurations of two neighbouring nodes in 2D projected pictures of 3D
trihedral scenes, and 29 out of these can be modelled by tree-preserving con-
straints.

Proof. The complete list of these configurations and their corresponding tree
convex or tree-preserving constraints is attached as an appendix. Note that we
do not consider T-junctions in line drawing pictures since they decompose into
unary constraints.


As a consequence, we know that these 29 configurations of the scene labelling

problem with 2D line-drawing pictures can be solved by the path-consistency
algorithm in polynomial time. Moreover, since it is NP-hard to decide if a trihe-
dral scene labelling instance is consistent, we have the following corollary.
Corollary 2. The consistency problem of tree convex constraint networks is NP-
complete.
 On Tree-Preserving Constraints 259

We next give a scene labelling instance which can be modelled by tree-

preserving constraints but not by chain-preserving or CRC constraints. Consider
the line drawing in the left of the following figure and the constraints for the
drawing listed in the right. One can easily verify that all constraints are tree-
preserving w.r.t. the forest structures listed in Fig. 5, but, for example, δ21 is
not chain-preserving for the forest structures illustrated in Fig. 5 and δ25 is not
CRC.

7 Further Discussion and Conclusion

In this paper, when formulating a CSP, we allow different variables have different
tree domains. Feder and Vardi [7] and many other authors (see e.g. [1,11]) also
considered CSPs which have a common domain D for all variables. These CSPs
are called one-sorted in [3]. For one-sorted tree-preserving CSPs, we could also
define a majority operator [11] under which the set of tree-preserving constraints
is closed. This implies that the class of one-sorted tree-preserving CSPs has
bounded strict width [7] and hence tractable. Indeed, such a majority operator ρ
is defined as follows: for any three nodes a, b, c in a tree domain T , define ρ(a, b, c)
as the node d which is the intersection of paths πa,b , πa,c , and πb,c . Following [3],
it is straightforward to extend this result to multi-sorted tree-preserving CSPs.
In this paper, we identified two new tractable subclasses of tree convex con-
straint which are called path- and tree-preserving constraints, and proved that
a chain- or path-preserving constraint network is in essence the disjoint union
of several independent CRC constraint networks, and hence (re-)established the
tractability of these constraints. More importantly, we proved that when enforc-
ing arc- and path-consistency on a tree-preserving constraint network, in each
step, the network remains tree-preserving. This implies that enforcing arc- and
260 S. Kong et al.

path-consistency will change a tree-preserving constraint network into a globally

consistent constraint network. This also implies that the efficient partial path-
consistent algorithm for large sparse networks is applicable for tree-preserving
constraint network. As an application, we showed that a large class of the tri-
hedral scene labelling problem can be modelled by tree-preserving constraints.
This shows that tree-preserving constraints are useful in real world applications.

Acknowledgments. We sincerely thank the anonymous reviewers of CP-15 and

IJCAI-15 for their very helpful comments. The majority operator was first pointed
out to us by two reviewers of IJCAI-15. This work was partially supported by ARC
(FT0990811, DP120103758, DP120104159) and NSFC (61228305).

References
1. Barto, L., Kozik, M.: Constraint satisfaction problems solvable by local consistency
methods. Journal of ACM 61(1), 3:1–3:19 (2014)
2. Bliek, C., Sam-Haroud, D.: Path consistency on triangulated constraint graphs.
In: IJCAI 1999, pp. 456–461 (1999)
3. Bulatov, Andrei A., Jeavons, Peter G.: An algebraic approach to multi-sorted
constraints. In: Rossi, Francesca (ed.) CP 2003. LNCS, vol. 2833, pp. 183–198.
Springer, Heidelberg (2003)
4. Conitzer, V., Derryberry, J., Sandholm, T.: Combinatorial auctions with structured
item graphs. In: AAAI 2004, pp. 212–218 (2004)
5. Dechter, R., Meiri, I., Pearl, J.: Temporal constraint networks. Artificial Intelli-
gence 49(1–3), 61–95 (1991)
6. Deville, Y., Barette, O., Hentenryck, P.V.: Constraint satisfaction over connected
row convex constraints. Artificial Intelligence 109(1–2), 243–271 (1999)
7. Feder, T., Vardi, M.Y.: The computational structure of monotone monadic snp and
constraint satisfaction: A study through datalog and group theory. SIAM Journal
on Computing 28(1), 57–104 (1998)
8. Freuder, E.C.: Synthesizing constraint expressions. Communications of the ACM
21(11), 958–966 (1978)
9. Freuder, E.C.: A sufficient condition for backtrack-free search. Journal of the ACM
29(1), 24–32 (1982)
10. Huffman, D.A.: Impossible objects as nonsense sentences. Machine Intelligence
6(1), 295–323 (1971)
11. Jeavons, P., Cohen, D.A., Cooper, M.C.: Constraints, consistency and closure.
Artificial Intelligence 101(1–2), 251–265 (1998)
12. Kirousis, L.M., Papadimitriou, C.H.: The complexity of recognizing polyhedral
scenes. In: FOCS 1985, pp. 175–185 (1985)
13. Kumar, T.K.S.: Simple randomized algorithms for tractable row and tree convex
constraints. In: AAAI 2006, pp. 74–79 (2006)
14. Li, S., Liu, W., Wang, S.: Qualitative constraint satisfaction problems: An extended
framework with landmarks. Artificial Intelligence 201, 32–58 (2013)
15. Maruyama, H.: Structural disambiguation with constraint propagation. In: ACL
1990, pp. 31–38 (1990)
16. Montanari, U.: Networks of constraints: Fundamental properties and applications
to picture processing. Information Sciences 7, 95–132 (1974)
 On Tree-Preserving Constraints 261

17. Van Beek, P., Dechter, R.: On the minimality and global consistency of row-convex
constraint networks. Journal of the ACM 42(3), 543–561 (1995)
18. Zhang, Y., Freuder, E.C.: Properties of tree convex constraints. Artificial Intelli-
gence 172(12–13), 1605–1612 (2008)
19. Zhang, Y., Marisetti, S.: Solving connected row convex constraints by variable
elimination. Artificial Intelligence 173(12), 1204–1219 (2009)
20. Zhang, Y., Yap, R.H.C.: Consistency and set intersection. In: IJCAI 2003,
pp. 263–270 (2003)
 Modeling and Solving Project Scheduling
with Calendars

Stefan Kreter1 , Andreas Schutt2,3(B) , and Peter J. Stuckey2,3

1
Operations Research Group, Institute of Management and Economics,
Clausthal University of Technology, 38678 Clausthal-Zellerfeld, Germany
stefan.kreter@tu-clausthal.de
2
Optimisation Research Group, National ICT Australia, Melbourne, Australia
{andreas.schutt,peter.stuckey}@nicta.com.au
3
Department of Computing and Information Systems, The University of Melbourne,
Melbourne, VIC 3010, Australia

Abstract. Resource-constrained project scheduling with the objective

of minimizing project duration (RCPSP) is one of the most studied
scheduling problems. In this paper we consider the RCPSP with gen-
eral temporal constraints and calendar constraints. Calendar constraints
make some resources unavailable on certain days in the scheduling period
and force activity execution to be delayed while resources are unavail-
able. They arise in practice from, e.g., unavailabilities of staff during
public holidays and weekends. The resulting problems are challenging
optimization problems. We develop not only four different constraint
programming (CP) models to tackle the problem, but also a specialized
propagator for the cumulative resource constraints taking the calendar
constraints into account. This propagator includes the ability to explain
its inferences so it can be used in a lazy clause generation solver. We
compare these models, and different search strategies on a challenging
set of benchmarks using a lazy clause generation solver. We close 83 of
the open problems of the benchmark set, and show that CP solutions
are highly competitive with existing Mip models of the problem.

1 Introduction
The resource-constrained project scheduling problem with general temporal and
calendar constraints (RCPSP/max-cal) is an extension of the well-known RCPSP
and RCPSP/max (see, e.g., [14, Chap. 2]) through calendars. The RCPSP/max-
cal can be given as follows. For a set of activities, which require time and renew-
able resources for their execution, execution time intervals must be determined
in a way that minimum and maximum time lags between activities are satisfied,
the prescribed resource capacities are not exceeded, and the project duration is
minimized. The difference with RCPSP/max is that a calendar is given for each
renewable resource type that describes for each time period whether the resource
type is available or unavailable. Time periods of unavailability can occur, e.g.,
due to weekends or public holidays. The activities and time lags are dependent

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 262–278, 2015.
DOI: 10.1007/978-3-319-23219-5 19
 Modeling and Solving Project Scheduling with Calendars 263

resource 3
3 1 6 @
- @
1  2 3 @ 3
@

{1, 2, 3} {3}
@ 2 @
@
0 @ 0 @
R
@ 1
@ -t
0 5 1 2 3 4 5 6 7 8 9 10

∅ @  ∅ resource 2
@ 3 2
R 3  - 4
6 @ @ 4
@ 3
@ @
@ @ 3
@ @
{1, 2, 3} {1, 2} @ @
@ @
(a) Logic diagram of the example project 1 @ @
@ @ -t
1 2 3 4 5 6 7 8 9 10

resource 1
6 J

6 4
Cal3
JJ
J

3
Cal2
JJ
JJ
Cal1 -t
1
-t
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

(b) Time periods of unavailability (c) Possible resource allocation

Fig. 1. Illustrative Example for RCPSP/max-cal

on the resource calendars, too, and some activities can be interrupted for the
duration of a break while others cannot be interrupted due to technical reasons.
For the interruptible activities a start-up phase is given during which the activ-
ity is not allowed to be paused. Concerning the renewable resource types one
distinguishes resource types that stay engaged or are blocked, respectively, dur-
ing interruptions of activities that require it and resource types that are released
and can be used to carry out other activities during interruptions.
Our motivation for developing CP models for the RCPSP/max-cal and using
lazy clause generation to solve it lies in the very good results obtained by
[18,19,20] solving RCPSP and RCPSP/max by lazy clause generation.
Example 1. Figure 1 shows an illustrative example with six activities and three
renewable resource types. The project start (activity 0) and the project end
(activity 5) are fictitious activities, i.e., they do not require time or resources. A
logic diagram of the project is given in Fig. 1(a) where each activity is represented
by a node with the duration given above and the set of resource types used by
the activity below the node. The arcs between the nodes represent time lags.
The calendars of the three renewable resource types are depicted in Fig. 1(b).
If there is a box with an X for a resource type k and time t, then resource type k
is not available at time t. Resource type 1 is always available and can be thought
of as a machine. Resource types 2 and 3 can be thought of as different kinds of
staff where resource type 2 (3) has a five-day (six-day) working week. In addition,
assume that resource type 1 stays engaged or is blocked, respectively, during a
264 S. Kreter et al.

break of an activity that requires resource type 1 for its execution while resource
types 2 and 3 are released during interruptions of activities.
A possible resource allocation of the three renewable resource types is shown
in Fig. 1(c). Activity 3 requires all renewable resource types for its execution.
Since resource type 2 is not available in periods 6 and 7, activity 3 is interrupted
during these periods. While resource type 1 stays engaged during the interrup-
tion, resource type 3 can be used to carry out activity 2 in period 6. 

Few authors have dealt with calendars in project scheduling so far. A time
planning method for project scheduling with the same calendar for each resource
type is introduced in [23]. In [6] the RCPSP/max with different calendars for
each renewable resource type is investigated for the first time but the start-
up phase of the interruptible activities are not taken into account. [6] proposes
methods to determine the earliest and latest start and completion times for the
project activities and priority rule methods. Procedures to determine the earliest
and latest start times if a start-up phase is taken into account are presented in
[7] and [14, Sect. 2.11]. In addition, they sketch how priority-rule methods for
the RCPSP/max can be adapted for calendars. In the approach in [7] and [14,
Sect. 2.11] all resources stay engaged during interruptions of activities. Within
the priority-rule methods in [6,7], and [14, Sect. 2.11] the procedures to deter-
mine the earliest and latest start times must be carried out in each iteration.
Recently, a new time planning method, three binary linear model formulations,
and a scatter search procedure for the RCPSP/max-cal were developed in [9].
Moreover, Kreter et al. [9] introduce a benchmark test set which is based on
the UBO test set for RCPSP/max [8]. The time planning method determines all
time and calendar feasible start times for the activities and absolute time lags
depending on the start times of the activities once in advance and then uses this
throughout the scatter search.
In CP, the works [3,4] respectively propose calendar constraints/rules for
ILOG Schedule and Cosytech CHIP. The former [3] was generalized to intensity
functions of activities in IBM ILOG CP Optimizer, while breaks of activities
extend the length between their start and end times, only resource types that
stay engaged can be modeled directly. The latter [4] introduces constraint rules
in the global constraint diffn for parallel machine scheduling.
A practical application where calendars must be considered as well as other
additional constraints can be found in batch scheduling [21]. Problems that are
related to the RCPSP/max-cal are treated in [22,5]. An alternative approach to
include calendars into project scheduling that makes use of calendar independent
start-start, start-end, end-start, and end-end time lags is proposed in [22] and [5]
studies the RCPSP with non-preemptive activity splitting, where an activity in
process is allowed to pause only when resource levels are temporarily insufficient.

2 Problem Description
In this section we describe the RCPSP/max-cal formally and give an example
instance. We use identifiers and definitions from [9]. In what follows, we assume
 Modeling and Solving Project Scheduling with Calendars 265

that a project consists of a set V := {0, 1, . . . , n, n + 1}, n ≥ 1, of activities,

where 0 and n + 1 represent the begin and the end of the project, respectively.
Each activity i has a processing time pi ∈ N0 . Activities i with pi > 0 are called
real activities and the set of real activities is denoted by V r ⊂ V . Activities 0
and n + 1 as well as milestones, which specify significant events of the project
and have a duration of pi = 0, form the set V f = V \ V r of fictitious activities.
A project completion deadline d ∈ N has to be determined in order to define
the time horizon of the calendars and the time axis is divided into intervals
[0, 1), [1, 2), . . . , [d−1, d) where a unit length time interval [t−1, t) is also referred
to as time period t. The set of renewable resource types is denoted by R and for
each renewable resource type k ∈ R a resource capacity Rk ∈ N is given that
must not be exceeded at any point in time. The amount of resource type k that
is used constantly during the execution of activity i ∈ V is given by rik ∈ N0 .
For fictitious activities i ∈ V f rik := 0 holds for all k ∈ R. For each resource
type a resource calendar is given.
Definition 1. A calendar for resource k ∈ R is a step function Calk (·) :
[0, d) → {0, 1} continuous from the right at the jump points, where the condition

1, if period [t , t + 1 ) is a working period for k
Calk (t) :=
0, if period [t , t + 1 ) is a break period for k

is satisfied.
With Ri := {k ∈ R | rik > 0} indicating the set of resource types that is used
to carry out activity i ∈ V , an activity calendar Ci (·) : [0, d) → {0, 1} can be
determined from the resource calendars as follows:

mink∈Ri Calk (t), if Ri
= ∅
Ci (t) :=
1, otherwise.

Then, for every activity i and a point in time t ∈ T := {0, 1, . . . , d} functions

next breaki (t) and next starti (t) give the start time and the end time of the
next break after time t in calendar Ci , respectively.

next breaki (t) := min{τ ∈ T | τ > t ∧ Ci (τ ) = 0}

next starti (t) := min{τ ∈ T | τ > t ∧ Ci (τ ) = 1 ∧ Ci (τ − 1) = 0}

When calendars are present, we have to distinguish activities that can be inter-
rupted for the duration of a break in the underlying activity calendar and activ-
ities that are not allowed to be interrupted. The set of (break-)interruptible
activities is denoted by V bi ⊂ V and the set of non-interruptible activities is
given by V ni = V \ V bi , where V f ⊆ V ni holds. The execution of an activity
i ∈ V bi must be interrupted at times t with Ci (t) = 0, and the execution must
be continued at the next point in time τ > t with Ci (τ ) = 1. Si ∈ T indicates the
start time and Ei ∈ T represents the end of activity i ∈ V . Since the jump points
in the calendars Calk , k ∈ R, are all integer valued, the points in time where an
activity is interrupted or continued are integer valued, too. The completion time
266 S. Kreter et al.

t−1
of activity i ∈ V can be determined by Ei (Si ) := min{t | τ =Si Ci (τ ) = pi }.
For each activity i ∈ V a start-up phase εi ∈ N0 is given during which activity
i is not allowed to be interrupted. For all activities i ∈ V ni εi := pi holds. We
assume that the underlying project begins at time 0, i.e., S0 := 0. Then, the
project duration equals Sn+1 . In addition, we assume that no activity i ∈ V can
be in execution before the project start, i.e., Si ≥ 0, or after the project end,
i.e., Ei ≤ Sn+1 .
Between the activities a set A of minimum and maximum time lags is given.
W.l.o.g. these time lags are defined between the start times of the activities (see
[6,9]). For each time lag i, j ∈ A, a resource set Rij ⊆ R and a length δij ∈ Z
are given, from which we can compute a calendar Cij (·) : [0, d) → {0, 1} for each
time lag by 
mink∈Rij Calk (t), if Rij
= ∅
Cij (t) :=
1, otherwise
i.e., at least tu time units mustelapse after the start of activity i before activity j
t−1
can start where tu = min{t | τ =Si Cij (τ ) = δij }.
With parameter ρk we indicate whether renewable resource types k ∈ R
stay engaged or are blocked, respectively, during interruptions of activities that
require it (ρk = 1) or are released and can be used to carry out other activities
during interruptions (ρk = 0). A vector S = (S0 , S1 , . . . , Sn+1 ) of all activity
start times is called a schedule. Given a schedule S and point in time t the set
of all real activities i ∈ V r that are started before but not completed at time
t is called the active set and can be determined by A(S, t) := {i ∈ V r | Si ≤
t < Ei (Si )}. Then, the resource utilization rkcal (S, t) of resource k ∈ R at time t
according to schedule S can be computed by
 
rkcal (S, t) := rik + rik ρk .
i∈A(S,t)|Ci (t)=1 i∈A(S,t)|Ci (t)=0

With the introduced notation the following mathematical formulation for the
RCPSP/max-cal can be given (cf. [6]):

Minimize Sn+1 (1)

Si +εi −1
subject to Ci (t) = εi i∈V (2)
t=Si
Sj −1 Si −1
Cij (t) − Cij (t) ≥ δij i, j ∈ A (3)
t=Si t=Sj

rkcal (S, t) ≤ Rk k ∈ R, t ∈ T \ {d} (4)

Si ∈ T i∈V (5)

The aim of the RCPSP/max-cal is to find a schedule that minimizes the project
makespan (1) and satisfies the calendar constraints (2), time lags (3), and
resource capacities (4).
Each project can be represented by an activity-on-node network where each
activity i ∈ V is represented by a node and each time lag i, j ∈ A is given by
an arc from node i to node j with weights δij and Rij . The activity duration as
 Modeling and Solving Project Scheduling with Calendars 267

3,13,R1 - 1,1 Cal1 (t)

1  2
 -7,R 2 @ 1 r
6
0,∅ 2,1,2 0,0,3
@1,R2 -t
0,0 @ 0,0 2 4 6 8 10
@
R Cal2 (t)
0  5
-10,∅ r
6 rq r
0,0,0@ 0,∅ 0,0,0 1
2,R4 r rq -t
@ 2 4 6 8 10
@ 3,2
3,R3 - 2,2 Cal3 (t)
@R 3  4 1 r
6 qr r
1,2,1 -3,R3 2,1,0 r rq -t
2 4 6 8 10
pi , εi pj , εj
δij , Rij
Legend: i - j
ri1 , ri2 , ri3 rj1 , rj2 , rj3

Fig. 2. Activity-on-node network and resource calendars

well as the start-up phase is given above node i in an activity-on-node network

and the resource requirements of activity i ∈ V are given below node i. For
the case where time lags depend on calendars, the label-correcting and triple
algorithm (see, e.g., [2, Sects. 5.4 and 5.6]) can be adapted and integrated in
a time planning procedure that determines a set Wi for each activity i ∈ V
containing all start times that are feasible due to the time lags and calendar
constraints, i.e., this procedure determines the solution space of the resource
relaxation of the RCPSP/max-cal (problem (1)–(3), (5)) [9]. In addition to the
sets Wi , the time planning procedure in [9] determines the “absolute” durations
of each activity and time lag with respect to the activities start times. The
absolute duration of an activity i ∈ V is denoted by pi (Si ) := Ei (Si ) − Si and
the absolute time lag for i, j ∈ A by dij (t) for each t ∈ Wi .
Example 2. Figure 2 shows the problem of Ex. 1 again, but now filled with
information for the activites start-up phases and resource requirements as well
as information for the time lags.
Activities 0, 2, 4, and 5 are non-interruptible while activities 1 and 3 form
the set V bi and therefore can be interrupted for the duration of a break in the
underlying activity calendar. By applying the determination rules from above
Cal1 = C0 = C5 = C01 = C03 = C50 , Cal2 = C1 = C3 = C4 = C12 = C34 =
C43 = C45 , and Cal3 = C2 = C21 = C25 hold for the activity and time lag
calendars. Since both time lags between activities 3 and 4 depend on the same
calendar and p3 = δ34 = −δ43 , activity 4 must be started when activity 3 ends
or more precisely at the next point in time after the end of activity 3 where the
calendar equals 1. The arc from the project end (node 5) to the project start
(node 0) represents an upper bound on the planning horizon of d = 10.
For the given example the time planning procedure from [9] determines
the sets W0 = {0}, W1 = {0, 1, 2, 3, 4}, W2 = {3, 4, 5, 7, 8, 9}, W3 = {0, 2, 3},
268 S. Kreter et al.

r1cal (S, t) r2cal (S, t)

6 s rs 6 s rs
3 3
2 s sr3 s sr 2 3 s rs s rs
1 1 s sr 4 1 s sr 3 3 s sr
s rs 3 s- t s rs 1 s sr 4 s- t
5 10 5 10
r3cal (S, t)

6 s rs s rs
3
2 s sr3
1 1 s sr2 s sr
s rs 3 s rs3 s -t
5 10

Fig. 3. Resource profiles of schedule S = (0, 1, 5, 3, 8, 10)

W4 = {3, 7, 8}, and W5 = {5, 6, 7, 8, 9, 10}. For example, activity 4 cannot start
at times 5 or 6 since there is a break in calendar C4 from 5 to 7. Moreover, activ-
ity 4 cannot start at time 4 because it has to be executed without interruptions.
Due to the time lag between activities 3 and 4, activity 3 cannot start at time
1, because if activity 3 started at time 1 activity 4 must start at time 4.
For the time- and calendar-feasible schedule S = (0, 1, 5, 3, 8, 10) the resource
profiles are given in Fig. 3. As already mentioned in the introduction resource
type 1 stays engaged during interruptions (ρ1 = 1) while resource types 2 and
3 are released during interruptions (ρ2 = ρ3 = 0). If the inequality Rk ≥ 3 is
fullfilled for each k ∈ R, schedule S is resource feasible and therefore a feasible
solution for the given example. 2

3 Models for RCPSP/max-cal

In this section, we present four different ways of modeling the RCPSP/max-cal.
The first three approaches use only well-known constraints from finite domain
propagation, while a new constraint to model the resource restrictions of the
RCPSP/max-cal and a corresponding propagator are used in the fourth model.

3.1 Model timeidx (Time Indexed Formulation)

In preprocessing, the time planning procedure of [9] is used to determine the
sets Wi of all time- and calendar-feasible start times for each activity i ∈ V and
S i ∈ Wi i∈V (6)
must be satisfied. Since the absolute time lags between the activities are depen-
dent on the start time of activity i for each i, j ∈ A, element constraints are
used to ensure that the correct values are taken into account.
element(Si , dij , dij ) i, j ∈ A (7)
 Modeling and Solving Project Scheduling with Calendars 269

Thereby, dij is an array that contains for all Si ∈ Wi the corresponding dij (Si )
value. Then, the constraints modelling time lags are

Sj − Si ≥ dij i, j ∈ A (8)

Absolute durations of the activities i ∈ V are used and the correct assignment
is ensured again by element constraints, where pi is an array containing for all
Si ∈ Wi the coresponding pi (Si ) value.

element(Si , pi , pi ) i∈V (9)

We implement the resource constraints using a time-indexed decomposition with

binary variables bit for each real activity i ∈ V r and point in time t ∈ T where
bit is true when i runs at t.

bit ↔ Si ≤ t ∧ t < Si + pi i ∈ V r, t ∈ T (10)


bit rik (Ci (t) + (1 − Ci (t)) ρk ) ≤ Rk k ∈ R, t ∈ T (11)
i∈V r

Model timeidx can now be given by: Minimize Sn+1 subject to (6) − (11).

3.2 Model 2cap (Doubling Resource Capacity)

Usually global propagators should be used to implement the resource constraints,

since more information is taken into account during propagation. This model
and the next make use of the global cumulative propagator [1] that explains
its propagation [16]. If the resource k ∈ R under investigation stays engaged
during interruptions of activities that require k for their execution, i.e., ρk = 1,
the global cumulative propagator can be used directly with the absolute activity
durations. If we regard the absolute duration of each activity i ∈ V and assume
that activity i requires rik units of resource k ∈ R with ρk = 0 at each point
in time {Si , . . . , Ei (Si ) − 1}, there can be resource overloads at break times of
an activity even if the corresponding schedule is feasible. One way to handle
resources k ∈ R with ρk = 0 is to determine points in time Rtimes k where
there exist an activity that can be in execution and another activity that can
be interrupted, double the resource capacity Rk , introduce a set Vkd of dummy
activities that require exactly Rk units of resource k at each point in time t ∈
T \ Rtimes
k , and use the global cumulative propagator:

cumulative(S, p , rk , Rk ) k ∈ R : (ρk = 1 ∨ Rtimes

k = ∅) (12)
cumulative(S ∪ S d , p ∪ pd , rk ∪ rkd , 2 Rk ) k ∈ R : (ρk = 0 ∧ Rtimes
k
= ∅) (13)

Note that rk is a vector containing the resource requirements on resource k of all

activities i ∈ V and that the vectors S d , pd , and rkd contain start times, absolute
durations, and resource requirements on resource k, respectively, for all j ∈ Vkd .
270 S. Kreter et al.

In addition, some decomposed constraints from (10) and (11) are required to
enforce non-overload of resource k at times Rtimes
k .

bit ↔ Si ≤ t ∧ t < Si + pi i ∈ V r, t ∈ Rtimes
k (14)
k∈R:ρk =0

bit rik Ci (t) ≤ Rk k ∈ R : ρk = 0, t ∈ Rtimes
k (15)
i∈V r

For all k ∈ R with ρk = 0 the set Rtimes

k is defined as follows.

Rtimes
k := {t ∈ T | ∃ i, j ∈ V : rik > 0 ∧ rjk > 0 ∧ min Wi ≤ t < Ei (max Wi )∧
min Wj ≤ t < Ej (max Wj ) ∧ Ci (t)
= Cj (t)}

Model 2cap can be achieved by deleting constraints (10) and (11) from model
timeidx and adding constraints (12)–(15) instead.
Example 3. Regarding the example project from Fig. 2 on page 267, resource 3
is the only resource where Rtimes
k
= ∅. We can see in Fig. 3 on page 268 that
in time period 6 activity 2 is in execution and activity 3 is interrupted. Hence
Rtimes
3 = {5}. The solution presented in Fig. 3 is resource feasible for R3 = 3 but
cumulative does not know that activity 3 is interrupted and detects a resource
overload if resource limit R3 = 3 is used. By doubling the resource capacity and
introducing a set V3d of dummy activities requiring 3 resources in all periods
but 6, the cumulative of (13) does not detect a resource overload. The reason
for the decomposed constraint (15) for time point 5 is clear when we imagine
another activity 2 that requires resource type 3 for its execution and could be
in execution in time period 6 just like activity 2, then for any solution where
both activities 2 and 2 are in execution in time period 6 there is a resource
overload, which the cumulative does not detect when the resource capacity is
doubled. 


3.3 Model addtasks (Adding Split Tasks)

Another way to handle resources k ∈ R with ρk = 0 is to introduce for
each interruptible activity i ∈ V bi a set Addi := {ai1 , ai2 , . . . , ai|Addi | } of addi-
tional (non-interruptible) activities that cover only those points in time t ∈
{Si , . . . , Ei (Si )−1} with Ci (t) = 1, i.e., resource k is released during an interrup-
tion of activity i. For the start times and processing times of activites aij ∈ Addi
the following equalities must be guaranteed.

Sai1 = Si i ∈ V bi (16)
Saij = next starti (Saij−1 ) i ∈ V bi , j ∈ {2, . . . , |Addi |} (17)
j−1

paij = min(next breaki (Saij ), pi − paih ) i ∈ V bi , j ∈ {1, . . . , |Addi |} (18)
h=1
raij ,k = rik i ∈ V bi , j ∈ {1, . . . , |Addi |} (19)
 Modeling and Solving Project Scheduling with Calendars 271

Thereby, next breaki (t) gives the start time of the next break after time t in
calendar Ci and next starti (t) gives the end time of the next break as defined
in Sect. 2. Finally, the resource requirement of each additional activity aij ∈ Addi
is set equal to rik and the global cumulative propagator can be used:
cumulative(S, p , rk , Rk ) k ∈ R : ρk = 1 (20)
cumulative(S a , pa , rka , Rk ) k ∈ R : ρk = 0 (21)
In constraints (21), the vectors S a , pa , and rka contain not only the start times,
durations, and resource requirements of the additional activities aij , i ∈ V bi , j ∈
{1, . . . , |Addi |}, but also the start times, durations, and resource requirements
of the non-interruptible activities i ∈ V ni .
Model addtasks can be achieved by deleting constraints (10) and (11) from
model timeidx as well as adding constraints (16)–(21) instead.

3.4 Model cumucal (Global Calendar Propagator)

For our fourth model for RCPSP/max-cal, we created a global cumulative prop-
agator that takes calendars into account and named it cumulative calendar.
The fourth model (cumucal) can be achieved by deleting constraints (9), (10),
and (11) from model timeidx as well as adding constraints (22)
cumulative calendar(S, p, rk , Rk , C, ρk ) k∈R (22)
with p being the vector of all constant processing times pi and C being the vector
of all activity calendars Ci , i ∈ V .
The cumulative calendar propagator is made up of two parts, a time-table
consistency check and filtering. The basic ideas of these two parts are the same as
in the cumulative propagator of [18], but non-trivial adaptions were necessary to
consider calendars. These adaptions are described in the following. The compul-
sory part [10] of an activity i ∈ V is the time interval [ub(Si ), lb(Si ) + pi (lb(Si ))),
where lb(Si ) (ub(Si )) represents the current minimum (maximum) value in the
domain of Si . If ρk = 1 for the resource k ∈ R then activity i requires rik units
of resource k at each point in time of its compulsory part. Otherwise (ρk = 0),
activity i requires rik units of resource k only at points in time of its compulsory
part where Ci (t) = 1. The intervals where an activity requires resource k within
its compulsory part are named the calendar compulsory parts. At the begin
of the cumulative calendar propagator the calendar compulsory parts of all
activities are determined and a resource profile including all these parts is built.
Within the consistency check, resource overloads in this profile are detected. If
an overload of the resource k occurs in the time interval [s, e) involving the set
of activities Ω, the following conditions hold:
ub(Si ) ≤ s ∧ lb(Si ) + pi (lb(Si )) ≥ e i∈Ω
(1 − ρk ) · Ci (t) + ρk = 1 i ∈ Ω, t ∈ [s, e)

rik > Rk
i∈Ω
272 S. Kreter et al.

In a lazy clause generation solver integer domains are represented using Boolean
variables. Each variable x with initial domain D0 (x) = {l, . . . , u} is represented
by two sets of Boolean variables x = d, l ≤ d ≤ u and x ≤ d, l ≤ d < u which
define which values are in D(x). A lazy clause generation solver keeps the two
representations of the domain in sync. In order to explain the resource overload,
we use a pointwise explanation [18] at T imeD, which is the nearest integer to
the mid-point of [s, e).
∀i ∈ Ω : back(i, T imeD + 1) ≤ Si  ∧ Si ≤ T imeD → f alse
 t−1
max{τ ∈ T | z=τ Ci (z) = pi } if Ci (t − 1) = 1
back(i, t) := t−1
max{τ ∈ T | z=τ Ci (z) = pi − 1} if Ci (t − 1) = 0.
The definition by cases for back(i, t) is necessary to guarantee the execution of
activity i at time t−1, if Si = t−back(i, t) holds. If for a time t with Ci (t−1) = 0
back(i, t) would be calculated with the first case, then Ei (t − back(i, t)) < t and
the explanation would be incorrect. 
If there exists a proper subset of activities Ω  ⊂ Ω with i∈Ω  rik > Rk , the
explanation of the resource overload is done on set Ω  . Sometimes more than
one such subset exists. In this situation the lexicographic least set of activities
is chosen as was done in [18].
Time-table filtering is also based on the resource profile of calendar com-
pulsory parts of all activities. In a filtering without explanations the height of
the calendar compulsory parts concerning one time period or a time interval
is given. For an activity the profile is scanned through to detect time intervals
where it cannot be executed. The lower (upper) bound of an activity’s start
time is updated to the first (last) possible time period with respect to those
time intervals and the activity calendar. If we want to explain the new lower
(upper) bound we need to know additionally which activities have the calendar
compulsory parts of those time intervals.
A profile is a triple (A, B, C) where A = [s, e) is a time interval, B the set of
all activities that have a calendar compulsory part in the time interval A, and
C the sum of the resource requirements rik of all activities in B. Here, we only
consider profiles with a maximal time interval A with respect to B and C, i.e.,
no other profile ([s , e ), B, C) exists where s = e or e = s.
Let us consider the case when the lower bound of the start time variable
for activity i can be maximally increased from its current value lb(Si ) to a new
value LB(i) using time-table filtering (the case of decreasing upper bounds is
analogous and omitted). Then there exists a sequence of profiles [D1 , . . . , Dp ]
where Dh = ([sh , eh ), Bh , Ch ) with e0 = lb(Si ) and ep = LB(i) such that
∀h : 1 ≤ h ≤ p; Ch + rik > Rk ∧ sh < eh−1 + pi (eh−1 )
In Sect. 2, we introduced pi (t) only for t ∈ Wi . Note that pi (t) can be calculated
in the same way for t ∈ / Wi , where pi (t) takes the value d − t if less than pi
working periods are following after t in calendar Ci . In addition, if ρk = 0 is
satisfied then
∀h : 1 ≤ h ≤ p; ∃ t ∈ [sh , eh ) : Ci (t) = 1
 Modeling and Solving Project Scheduling with Calendars 273

Hence each profile Dh pushes the start time of activity i to eh .

Again we use pointwise explanations based on single time points. Unlike
the consistency case, we may need to pick a set of time points no more than
the absolute duration of activity i apart to explain the increasing of the lower
bound of Si over the time interval. For a profile with length greater than the
absolute processing time of activity i we may need to pick more than one time
point in a profile. Let {t1 , . . . , tm } be a set of time points such that t0 = lb(Si ),
tm + 1 = LB(i), ∀1 ≤ l ≤ m : tl−1 + pi (tl−1 ) ≥ tl and there exists a mapping
P (tl ) of time points to profiles such that ∀1 ≤ l ≤ m : sP (tl ) ≤ tl < eP (tl ) . Then
we build a pointwise explanation for each time point tl , 1 ≤ l ≤ m

back(i, tl + 1) ≤ Si  ∧ (back(j, tl + 1) ≤ Sj  ∧ Sj ≤ tl ) → tl + 1 ≤ Si 
j∈Bh

Example 4. We illustrate cumulative calendar for the example network from

Fig. 2. To explain both the time-table consistency check and the time-table
filtering we are using two different cases. For the first case (consistency check),
we assume that in the current search node lb(S1 ) = 3, ub(S1 ) = 4, lb(S2 ) =
8, ub(S2 ) = 9, lb(S3 ) = ub(S3 ) = 3, and lb(S4 ) = ub(S4 ) = 8 holds, i.e., activities
3 and 4 are already fixed. We examine the cumulative calendar for resource
type 1 with a resource capacity of R1 = 2. The calendar compulsory parts
are [4, 8) for activity 1, [3, 8) for activity 3, and [8, 10) for activity 4. Note
that activity 2 is not taken into account since r21 = 0 and that the calendar
compulsory parts equal the compulsory parts for this example because ρ1 = 1.
The compulsory parts of activities 1 and 3 cover the interval [4, 8) and a resource
overload of resource 1 occurs, since r11 + r31 = 2 + 1 = 3 > 2 = R1 . A pointwise
explanation of the resource overload is done at T imeD = 6:

3 ≤ S1  ∧ S1 ≤ 6 ∧ 3 ≤ S3  ∧ S3 ≤ 6 → f alse

For activities i = 1 and i = 3, respectively, Ci (T imeD − 1) = 0 is satisfied

and back(i, T imeD) is calculated through the second case. Without case differ-
entiation for back(i, t) only the first case would be considered, resulting that
back(1, T imeD) would equal 1 and the explanation would be wrong.
For the second case (time-table filtering), we assume that in the current
search node lb(S1 ) = ub(S1 ) = 0, lb(S2 ) = 3, ub(S2 ) = 8, lb(S3 ) = ub(S3 ) = 3,
and lb(S4 ) = ub(S4 ) = 8 holds. We examine the cumulative calendar for
resource type 3 with a resource capacity of R3 = 3. Activity 2 is the only task
where the start time is not fixed and the consistency check detects no resource
overload. The calendar compulsory parts are [0, 3) for activity 1, [3, 5), [7, 8) for
activity 3, and [8, 10) for activity 4. For the profile (A, B, C) with A = [3, 5),
B = {3}, and C = 1 the condition C + r23 = 1 + 3 > 3 = R3 is satisfied and
therefore the lower bound for variable S2 can be increased to LB(2) = 5. Since
the activity duration p2 equals 1 a pointwise explanation is done for t0 = 3 and
t1 = 4. The explanation for t0 = 3 is 3 ≤ S2  ∧ 1 ≤ S3  ∧ S3 ≤ 3 → 4 ≤ S2 
and for t1 = 4 it is 4 ≤ S2  ∧ 2 ≤ S3  ∧ S3 ≤ 4 → 5 ≤ S2 . 

274 S. Kreter et al.

3.5 Time Granularity Considerations

All models depend on the granularity chosen for the time. If the granularity
increases then the size of T increases respectively. Thus, the number of linear
constraints and auxiliary Boolean variables increases for the models timeidx and
2cap, especially for the former. Moreover, filtering algorithms for the element
constraints (used in all models) might be negatively affected due to a larger size
of the input arrays. The implemented time-table consistency check, filtering, and
explanation generation for resource overloads and start time bounds updates in
cumulative calendar depend on the granularity, too. Their respective runtime
complexity are O(x × y log(x × y) + x × z), O(x2 × z × y), and O(x × z) where
x is the number of tasks, y − 1 is maximal possible number of interruptions of
any task and z the maximal possible absolute duration of any task.

4 Experiments and Conclusion

We conducted extensive experiments on Dell PowerEdge R415 machines running
CentOS 6.5 with 2x AMD 6-Core Opteron 4184, 2.8GHz, 3M L2/6M L3 Cache
and 64 GB RAM. We used MiniZinc 2.0.1 [13] and the lazy clause generation [15]
solver chuffed rev 707.
A runtime limit of 10 minutes was imposed excluding runtimes needed for
pre-processing, initial solution generation, and compiling the MiniZinc models
to solver-dependent FlatZinc models. We used the same benchmarks and ini-
tial solutions as in [9], which are available at www.wiwi.tu-clausthal.de/en/chairs/
unternehmensforschung/research/benchmark-instances/.
Since instances with 10 or 20 activities could easily be solved within a
few seconds by any combination of solver, model, and search, we concentrate
on instances with 50 and 100 activities. The average runtime needed for pre-
processing and initial solution generation are less than a few seconds for instances
with 50 activities and less than 30 seconds for instances with 100 activities,
respectively.

4.1 Comparing Search Strategies

For finding the shortest project duration, we employ a branch-and-bound strat-
egy for which we investigate following four different search combinations. Those
seem likely to be most suitable based on our previous experience on solving
scheduling problems using lazy clause generation (see, e.g., [18,19,17]).
ff: Selects the variable with the smallest domain size and assigns the minimal
value in the domain to it.
vsids: Selects the literal with the highest activity counter and sets it to true,
where the literal is a part of the Boolean representation of the integer vari-
ables, i.e., x = v, x ≤ v, where x is an integer variable and v ∈ D(x).
Informally, the activity counter records the recent involvement of the literal
in conflicts and all activity counters are simultaneously decayed periodically.
 Modeling and Solving Project Scheduling with Calendars 275

Table 1. Comparison of search strategies on instances with 50 activities.

cmp(179) all(180)
model search #opt #feas #inf #un avg. rt avg. #cp avg. rt avg. #cp
cumucal alt 161 0 19 0 1.13 3847 1.71 5236
cumucal ff 160 1 19 0 7.54 16401 10.83 16310
cumucal hs 161 0 19 0 1.33 5358 1.80 6349
cumucal vsids 161 0 19 0 4.27 18495 4.79 19445

Table 2. Comparison of search strategies on instances with 100 activities.

cmp(140) all(180)
model search #opt #feas #inf #un avg. rt avg. #cp avg. rt avg. #cp
cumucal alt 158 11 11 0 7.58 11498 56.18 25170
cumucal ff 150 16 10 4 13.73 14305 82.51 16420
cumucal hs 152 17 11 0 20.33 34900 85.20 45109
cumucal vsids 133 36 11 0 76.87 146172 185.61 122457

The activity counter of a literal is increased during conflict analysis when

the literal is related to the conflict. It is an adaption of the variable state
independent decaying sum heuristic [12]. The search vsids is combined with
Luby restarts [11] and a restart base of 100 conflicts.
hs: The search starts off with ff and then switches to vsids after 1000 conflicts.
alt: The search alternates between ff and vsids starting with ff. It switches
from one to the other after each restart where we use the same restart policy
and base as for vsids.

Tables 1 and 2 show the results of chuffed on the cumucal model using dif-
ferent search strategies on instances with 50 and 100 activities, respectively. The
search strategies behave similar with the other models. We show the number of
instances proven optimal (#opt), not proven optimal but where feasible solutions
were found (#feas), proven infeasible (#inf), and where nothing was determined
(#un). We compare the average runtime in seconds (avg. rt) and average number
of choice points to solve (avg. #cp), on two subsets of each benchmark. The cmp
subset are all the benchmarks where all solvers proved optimality or infeasibility,
and all is the total set of benchmarks.
The alt search is clearly the fastest, also leading to the lowest average number
of nodes explored in comparison to the rest. Interestingly, the performance of
vsids significantly decays from instances with 50 activities to those ones with
100 activities in proportion to alt and ff. This decay also affects hs, but not
so dramatically. The strength of the alt method is the combination of integer
based search in ff which concentrates on activities that have little choice left,
with the robustness of vsids which is excellent for proving optimality once a
good solution is known.
276 S. Kreter et al.

Table 3. Comparison of models on instances with 50 activities.

cmp(177) all(180)
model search #opt #feas #inf #un avg. rt avg. #cp avg. ft avg. rt avg. #cp
addtasks alt 160 1 19 0 9.10 14232 0.84 15.39 17924
cumucal alt 161 0 19 0 0.92 3203 0.48 1.71 5236
timeidx alt 158 3 19 0 22.20 3484 18.65 31.82 3426
2cap alt 161 0 19 0 1.55 5341 0.72 3.12 9619

Table 4. Comparison of models on instances with 100 activities.

cmp(138) all(180)
model search #opt #feas #inf #un avg. rt avg. #cp avg. ft avg. rt avg. #cp
addtasks alt 139 28 10 3 25.88 26344 4.05 139.47 35525
cumucal alt 158 11 11 0 3.24 5037 2.45 56.18 25170
timeidx alt 131 38 10 1 83.37 4947 78.24 196.63 4031
2cap alt 153 16 11 0 6.13 8798 4.05 78.96 30728

Table 5. Comparison of solvers on instances with 50 activities.

cmp(170) all(180)
model search #opt #feas #inf #un avg. rt avg. #cp avg. ft avg. rt avg. #cp
cumucal chuffed+alt 161 0 19 0 0.59 2198 0.48 1.71 5236
timeidx chuffed+alt 158 3 19 0 10.45 1662 18.65 31.82 3426
timeidx ocpx+free 159 2 19 0 69.95 13383 19.97 83.92 14155
mip 153 7 18 2 222.42 — — 750.25 —

4.2 Comparing Models

Tables 3 and 4 compare the effect of the different models using chuffed and the
best search method alt. As expected, the time-indexed model, timeidx, is the
worst in terms of times due to the large model size, but it propagates effectively
as illustrated by the low number of explored nodes (only ever bettered by cumu-
cal). The model addtasks performs worst with respect to the average number
of nodes, which can be explained by the shorter activities causing weaker time-
table propagation in the cumulative propagator. The best model is cumucal
that takes the advantage of using fixed durations, since the variability is han-
dled directly by the propagator, and because it generates the smallest model.
We also show the average flattening time (avg. ft.) for all benchmarks, where
clearly cumucal is advantageous.

4.3 Comparing Solvers

Table 5 compares the results obtained by chuffed to those obtained by Opturion
CPX 1.0.2 (ocpx), which is available at www.opturion.com/cpx, and the best
solution obtained by any mixed-integer linear programming formulation from [9]
 Modeling and Solving Project Scheduling with Calendars 277

(mip), which is solved using CPLEX 12.6 on an Intel Core i7 CPU 990X with
3.47 GHz and 24GB RAM under Windows 7. For mip the runtime limit was
set to 3 hours and 8 threads were used. To get an idea of the impact of the
machine used, we also ran ocpx with the deterministic search ff on the same
Windows machine. ocpx was more than 3 times faster on that machine. It can
be seen that chuffed and ocpx clearly outperform the state-of-the-art solution
approach, which is mip, and that the machine we used is even slower than the
machine used in [9].
Overall the cumucal model closes all open benchmarks of size 50 and 75 of
size 100, and clearly, we significantly advance the state of the art.

Acknowledgments. NICTA is funded by the Australian Government as represented

by the Department of Broadband, Communications and the Digital Economy and the
Australian Research Council through the ICT Centre of Excellence program. This
work was partially supported by Asian Office of Aerospace Research and Development
(AOARD) grant FA2386-12-1-4056.

References
1. Aggoun, A., Beldiceanu, N.: Extending CHIP in order to solve complex scheduling
and placement problems. Mathematical and Computer Modelling 17(7), 57–73
(1993)
2. Ahuja, R., Magnanti, T., Orlin, J.: Network Flows. Prentice Hall, Englewood Cliffs
(1993)
3. Baptiste, P.: Constraint-Based Scheduling: Two Extensions. Master’s thesis, Uni-
versity of Strathclyde, Glasgow, Scotland, United Kingdom (1994)
4. Beldiceanu, N.: Parallel machine scheduling with calendar rules. In: International
Workshop on Project Management and Scheduling (1998)
5. Cheng, J., Fowler, J., Kempf, K., Mason, S.: Multi-mode resource-constrained
project scheduling problems with non-preemptive activity splitting. Computers &
Operations Research 53, 275–287 (2015)
6. Franck, B.: Prioritätsregelverfahren für die ressourcenbeschränkte Projektplanung
mit und ohne Kalender. Shaker, Aachen (1999)
7. Franck, B., Neumann, K., Schwindt, C.: Project scheduling with calendars. OR
Spektrum 23, 325–334 (2001)
8. Franck, B., Neumann, K., Schwindt, C.: Truncated branch-and-bound, schedule-
construction, and schedule-improvement procedures for resource-constrained
project scheduling. OR Spektrum 23, 297–324 (2001)
9. Kreter, S., Rieck, J., Zimmermann, J.: Models and solution procedures for the
resource-constrained project scheduling problem with general temporal constraints
and calendars. Submitted to European Journal of Operational Research (2014)
10. Lahrichi, A.: Scheduling: The notions of hump, compulsory parts and their use in
cumulative problems. Comptes Rendus de l’Académie des Sciences. Paris, Série 1,
Matématique 294(2), 209–211 (1982)
11. Luby, M., Sinclair, A., Zuckerman, D.: Optimal speedup of Las Vegas algorithms.
Information Processing Letters 47, 173–180 (1993)
12. Moskewicz, M.W., Madigan, C.F., Zhao, Y., Zhang, L., Malik, S.: Chaff: engineer-
ing an efficient SAT solver. In: Proceedings of Design Automation Conference -
DAC 2001, pp. 530–535. ACM, New York (2001)
278 S. Kreter et al.

13. Nethercote, N., Stuckey, P.J., Becket, R., Brand, S., Duck, G.J., Tack, G.: MiniZinc:
towards a standard CP modelling language. In: Bessière, C. (ed.) CP 2007. LNCS,
vol. 4741, pp. 529–543. Springer, Heidelberg (2007)
14. Neumann, K., Schwindt, C., Zimmermann, J.: Project Scheduling with Time Win-
dows and Scarce Resources, 2nd edn. Springer, Berlin (2003)
15. Ohrimenko, O., Stuckey, P.J., Codish, M.: Propagation via lazy clause generation.
Constraints 14(3), 357–391 (2009)
16. Schutt, A.: Improving Scheduling by Learning. Ph.D. thesis, The University of
Melbourne (2011). http://repository.unimelb.edu.au/10187/11060
17. Schutt, A., Feydy, T., Stuckey, P.J.: Explaining time-table-edge-finding propaga-
tion for the cumulative resource constraint. In: Gomes, C., Sellmann, M. (eds.)
CPAIOR 2013. LNCS, vol. 7874, pp. 234–250. Springer, Heidelberg (2013)
18. Schutt, A., Feydy, T., Stuckey, P.J., Wallace, M.G.: Explaining the cumulative
propagator. Constraints 16(3), 250–282 (2011)
19. Schutt, A., Feydy, T., Stuckey, P.J., Wallace, M.G.: Solving RCPSP/max by lazy
clause generation. Journal of Scheduling 16(3), 273–289 (2013)
20. Schutt, A., Feydy, T., Stuckey, P.J., Wallace, M.G.: A satisfiability solving app-
roach. In: Schwindt, C., Zimmermann, J. (eds.) Handbook on Project Management
and Scheduling, vol. 1, pp. 135–160. Springer International Publishing (2015)
21. Schwindt, C., Trautmann, N.: Batch scheduling in process industries: An applica-
tion of resource-constrained project scheduling. OR Spektrum 22, 501–524 (2000)
22. Trautmann, N.: Calendars in project scheduling. In: Fleischmann, B., Lasch, R.,
Derigs, U., Domschke, W., Rieder, U. (eds.) Operations Research Proceedings 2000,
pp. 388–392. Springer, Berlin (2001)
23. Zhan, J.: Calendarization of timeplanning in MPM networks. ZOR - Methods and
Models of Operations Research 36, 423–438 (1992)
 Deterministic Estimation of the Expected
Makespan of a POS Under Duration Uncertainty

Michele Lombardi, Alessio Bonfietti(B) , and Michela Milano

DISI, University of Bologna, Bologna, Italy

{michele.lombardi2,alessio.bonfietti,michela.milano}@unibo.it

Abstract. This paper is about characterizing the expected makespan of

a Partial Order Schedule (POS) under duration uncertainty. Our analy-
sis is based on very general assumptions about the uncertainty: in par-
ticular, we assume that only the min, max, and average durations are
known. This information is compatible with a whole range of values for
the expected makespan. We prove that the largest of these values and the
corresponding “worst-case” distribution can be obtained in polynomial
time and we present an O(n3 ) computation algorithm. Then, using the-
oretical and empirical arguments, we show that such expected makespan
is strongly correlated with certain global properties of the POS, and we
exploit this correlation to obtain a linear-time estimator. The estimator
provides accurate results under a very large variety of POS structures,
scheduling problem types, and uncertainty models. The algorithm and
the estimator may be used during search by an optimization approach, in
particular one based on Constraint Programming: this allows to tackle a
stochastic problem by solving a dramatically simpler (and yet accurate)
deterministic approximation.

1 Introduction

A Partial Order Schedule (POS) is an acyclic graph G = A, E, where A is a set
of activities ai , and E is a set of directed edges (ai , aj ) representing end-to-start
precedence relations. A POS is a flexible solution to a scheduling problem: some
of the edges derive from the original Project Graph, while the remaining ones are
added by an optimization approach so as to prevent potential resource conflicts.
A POS is very well suited as a solution format in the presence of dura-
tion uncertainty. Before the execution starts, each activity has a candidate start
time. During execution, whenever a duration becomes known, the start times
are updated (in O(n2 ) time) so that no precedence relation is violated: this
guarantees that the schedule remains resource feasible.
In the presence of duration uncertainty, the quality of a POS should be
evaluated via a stochastic metric, which in this paper is the expected value of
the makespan: this is far from being a perfect choice, but still a fair one in many
settings (in particular when the POS should be repeatedly executed, such as in
stream processing applications).


c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 279–294, 2015.
DOI: 10.1007/978-3-319-23219-5 20
280 M. Lombardi et al.

A POS can be obtained essentially in two ways. First, in a constructive

fashion, by adding edges to the graph until all potential resource conflicts
are resolved: this is done by Precedence Constraint Posting (PCP) approaches
[7,10,12,13]. Alternatively, one can obtain a POS from a traditional schedule
with fixed start times via a post-processing step [7,8,12,13]. Despite the fact
that a POS can adapt to uncertain durations, in practice the uncertainty is often
largely disregarded when the graph is built. For example, all the post-processing
methods operate on a schedule obtained for fixed durations. The PCP methods
are capable of dealing with uncertainty, for example via scenario-based optimiza-
tion or via the Sample Average Approximation [5,14]: both approaches require
to sample many duration assignments and to optimize a meta-model, resulting
from the combination of a set of models (one per scenario) connected by chaining
constraints. Despite this, most PCP approaches in the literature either assume to
have fixed durations (e.g. [7,12,13]) or rely on minimal uncertainty information
to perform robust (rather than stochastic) optimization (e.g. Simple Temporal
Networks with Uncertainty [11,15] or the method from [10]).
One of the main reasons for ignoring the uncertainty when searching for an
optimal POS is the additional complexity. In fact, the complexity of a scenario-
based approach depends on the number of samples required to have a satisfactory
accuracy, which can be large for problems with many activities. The second
main reason is the lack of reliable information, either because data has not been
collected or because the POS must be executed in a very dynamic environment.
In this paper, we tackle the problem of estimating the expected makespan
of a POS under duration uncertainty. The POS describes a solution computed
with a classical CP approach with a makespan minimization objective. We use
“deterministic” estimators rather than sampling-based statistics. The goal is pro-
viding an efficient alternative to scenario-based optimization in case the number
of required samples is too high. Moreover, our estimators rely on very general
assumptions on the uncertainty (in particular, we do not require independent
durations) and are therefore well suited for cases in which extensive informa-
tion is not available. Technically, we focus on computing or approximating the
largest value of the expected makespan that is compatible with the available
uncertainty information. Our main contributions are: first, a proof that such a
expected makespan value can be obtained in polynomial time and a O(n3 ) com-
putation algorithm; second, a mixed theoretical/empirical analysis that high-
lights interesting properties of POSs and enables the definition of a linear-time
approximate estimator.
We have evaluated our estimator under an impressive variety of settings,
obtaining reasonably accurate and robust results in most cases. Our algorithm
could be used within a heuristic scheduling method for estimating the value of
the expected makespan. Alternatively, with further research the algorithm could
be turned into a propagator for a global constraint, despite this is a non-trivial
task. Our estimator consists of a non-linear formula that can be embedded in
a CP approach using standard building blocks (sum and min operators): the
resulting CP expression is propagated in O(n). Indeed, this research line stems
 Deterministic Estimation of the Expected Makespan of a POS 281

in part from the unexpectedly good performance of a PCP approach based on

Constraint Programming for robust scheduling, that we presented in [10]: our
estimator in particular is designed to be used in a similar framework.
This paper is a follow-up of a previous work of ours [2], where we reached
similar (although less mature) conclusions under more restrictive assumptions.
The algorithm for computing the expected makespan is presented in Section 2,
the analysis and the estimator are in Section 3, and Section 4 provides some
concluding remarks.

2 Expected Makespan with Worst-Case Distribution

Our analysis is based on very general assumptions about the uncertainty. In
particular, we assume that only the minimal, maximal, and average durations
are known. Formally, the duration of each activity ai can be modeled as a random
variable Di ranging over a known interval [di , di ] and having known expected
value d˜i . Note that the Di variables are not assumed to be independent, and
therefore they are best described in terms of their joint probability distribution.
Assuming that lower completion times are desirable, the makespan for a given
instantiation of the Di variables can be obtained by computing the longest path
in G (often referred to as critical path). Formally, the makespan is a function
T (D) of the duration variables, where we refer as D to the vector of all Di . The
makespan is therefore itself stochastic, and E [T (D)] denotes its expected value.
Because we rely on so little information about the uncertainty, the expected
makespan is not unambiguously defined: the value of E [T (D)] depends on the
joint distribution of the Di , and there exists an infinite number of distributions
that is compatible with known values of d, d, and d. ˜
The expected makespan cannot be higher than T (d), i.e. than the makespan
with maximal durations. Moreover, since the expected value is a linear operator,
E [T (D)] cannot be lower T (d),˜ i.e. the makespan with expected durations.
Reaching stronger conclusions requires more powerful analysis tools, which
will be presented in this paper. In particular, we are interested in the largest
possible value of E [T (D)] that is compatible with the known duration data: in
absence of more information, this allows to obtain a safe estimate. The compu-
tation requires to identify a “worst-case” distribution, among the infinite set of
distributions that are compatible with the given d, d, and d. ˜

Properties of the Worst-case Distribution A generic (joint) probability distribu-

tion can be defined as a function:
P : Ω → [0, 1] (1)
where Ω is a set of scenarios ωk , each representing in our case an assignment
for all Di variables. The set Ω is called the support of the distribution and can
have infinite size. The value P (ωk ) is the probability of scenario ωk . The integral
(or the sum, for discrete distributions) of P over Ω should be equal to 1. Our
method for defining the worst-case distribution relies on a first, very important,
result that is stated in the following theorem:
282 M. Lombardi et al.

Theorem 1. There exists a worst-case distribution such that, in each scenario

with non-zero probability, every Di takes either the value di or the value di .

The proof is in Appendix A. From Theorem 1 we deduce that it is always possible

to define a worst-case
 distribution
  with a support Ω that is a subset of the
Cartesian product ai ∈A di , di , and the therefore with size bounded by 2|A| .

Finding the worst-case distribution This makes it possible to model the con-
struction of a worst-case distribution as an optimization problem. Formally, let
us introduce for each scenario a decision variable pωk ∈ [0, 1] representing the
value of P (ωk ). The assignment of the pωk variables must be such that the known
values of the expected durations are respected:
⎛ ⎞
   d˜i − di
di p ωk + d i ⎝ 1 − pωk ⎠ = d˜i and hence: p ωk = (2)
di − di
ωk :Di =di ωk :Di =di ωk :Di =di

Then a worst-case distribution can be found by solving:


P0 : max z = pωk T (D(ωk )) (3)
ωk ∈Ω
 d˜i − di
subject to: p ωk ≤ ∀ai ∈ A (4)
di − di
ωk :Di =di

p ωk ≤ 1 (5)
ωk ∈Ω

p ωk ≥ 0 ∀ωk ∈ Ω (6)
  
where Ω = ai ∈A di , di . Equation (3) is the makespan definition (D(ωk ) are
the durations in ωk ), Equation (4) corresponds to Equation (2), and Equation (5)
ensures that the total probability mass does not exceed one. It is safe to use a
≤ sign in Equation (4) and (5), because increasing a pωk can only improve the
problem objective and therefore all constraints are tight in any optimal solution.
P0 is linear, but has unfortunately an exponential number of variables.

Tractability of P0: Luckily, P0 has a number of nice properties that make it

much easier to solve. At a careful examination, one can see that P0 is the lin-
ear relaxation of a multi-knapsack problem. Results from LP duality theory [4]
imply that the optimal solution cannot contain more than min(n, m) fractional
variables, where n is the number of variables (|Ω| in our case) and m is the
number of knapsack constraints (i.e. |A| + 1).
Due to Constraint (5), in an optimal solution of P0: either 1) all variables are
integer and there is a single pωk = 1; or 2) all non-zero variables are fractional.
Therefore, the number of non-zero variables in an optimal solution is bounded
by |A| + 1. This means that a worst-case distribution with a support of size at
most |A| + 1 is guaranteed to exist.
 Deterministic Estimation of the Expected Makespan of a POS 283

Algorithm 1. Compute Ewc [T (D)]

Require: A POS G = A, E, plus di , di , and d˜i for each activity
1: Let T = 0 and stot = 1 and let si = (d˜i − di )/(di − di ) for each ai ∈ A
2: while stot > 0 do
3: for all activities ai ∈ A do
4: Let di = di if si > 0, otherwise let di = di
5:  path π over G, usingthe di values as durations.
Find the critical
6: Let pπ = min si : ai ∈ π and di = di
7: If pπ = 0, then pπ = stot
8: for all activities ai on the critical path such that di = di do
9: Set si = si − pπ
10: Set stot = stot − pπ
11: Set T = T + pπ ai ∈π di
12: return T

Furthermore, all the variables of P0 appear with identical weights in all

knapsack constraints (i.e. Equation (4) and (5)). Therefore, P0 can be solved
to optimality by generalizing the solution approach for the LP relaxation of the
classical knapsack problem: namely, one has to repeatedly: 1) pick the variable
pωk with the highest reward T (D(ωk )) in the objective; and 2) increase the value
of pωk until one of the knapsack constraints becomes tight.

A Polynomial Time Algorithm: We present (in Algorithm 1) a practical method

for the computation of E [T (D)] with worst-case distribution, i.e. Ewc [T (D)].
The algorithm improves the basic process that we have just described by relying
on partial scenarios. A partial scenario represents a group of scenarios having
the same makespan, and it is identified by specifying a duration (either di or
di ) for a subset of the activities. Algorithm 1 keeps (line 2) a global probability
budget stot and a vector of probability budgets si to keep track of the slack in
Constraint (5) and Constraints (4), respectively. Then the algorithm repeatedly
identifies a partial scenario that is compatible with the remaining slack and has
maximal makespan. This is done by: 1) Assigning to each activity a duration di .
This is equal to di if the corresponding slack variable si is non-zero, otherwise,
the duration is di (lines 4-5); and 2) finding the critical path π on the POS,
using the durations di (line 6).
At each iteration, the partial scenario is specified by: 1) the set of activities
on the critical path that have maximal duration di ; and 2) the set of all activities
for which di = di . The idea is that if both sets of activities take their prescribed
duration, then the critical path is π. This process always identifies the partial
scenario that is compatible with the current slack and has the largest makespan.
The partial scenario is then inserted in the joint distribution with its largest
possible probability: this is determined (with an exception discussed later) by
smallest slack of the activities for which di = di in the partial scenario. The
probability pπ of the partial scenario is used to update the slack variables at
lines 9-11. Finally, the expected makespan variable T is incremented by the
284 M. Lombardi et al.

probability of the partial scenario, multiplied by its associated makespan (i.e.

the sum of durations of the activities on the critical path).
The mentioned exception is that the probability pπ computed at line 7 may
be 0 in case all activities on π have di = di . If this happens, it means that the
probability of the current partial scenario is not limited by Constraints (4), but
only by Constraint (5). Hence, the scenario probability can be set to the value
of the remaining global slack (line 8). When the global slack becomes zero, the
algorithm returns the final value of Ewc [T (D)].
By construction, at each iteration a slack variable si becomes zero. If all
the si become zero, then the next partial scenario will have a critical path with
only “short” activities and the probability update at line 8 will trigger. Overall,
the algorithm can perform at most |A| + 1 iterations, in agreement with the
theoretical prediction made for problem P0. The complexity of each iteration
is dominated by the O(|A|2 ) longest path computation. The overall complexity
is therefore O(|A|3 ), which is a remarkable result given that in principle the
worst-case distribution had to be found among a set with infinite size.

3 Estimating Ewc [T(D)]

Algorithm 1 allows an exact computation of the expected makespan with worst
case distribution, but its scalability is limited by the O(n3 ) complexity. Moreover,
the added difficulty of embedding a procedural component within an optimiza-
tion approach may make the method less applicable in practice.
With the aim to address such limitations, we have devised a low-complexity,
approximate, estimator based on global graph properties that are reasonably
easy to measure. In detail, the estimator is given by the following formula:
 

 1  d˜i − di ai ∈A di
τ (T , T ) = T + T − T min 1, (7)
|A| di − di T
ai ∈A

Most of the terms in Equation (7) are constants that can be computed via a
pre-processing step. The only input terms that change at search time are T
and T , which are simplified notations for T (di ) and T (di ): both values can be
efficiently made available as decision variables in a PCP approach via (e.g.)
the methods from the Constraint Programming approach in [10]. The estimator
formula is non-linear, but it is not difficult to embed in a CP model by using
widely available building blocks: the resulting encoding is propagated in O(|A|).
The estimator is obtained via a non-trivial process that is discussed in detail
in the remainder of this section. Its accuracy and robustness are discussed in
Section 3.3.

3.1 A Simplified POS Model

The estimator from Equation (7) is obtained by reasoning on a simplified POS
model. Specifically, such model is a layered graph with constant width, and
identical minimal duration di = δ and maximal duration di = δ for each activity.
 Deterministic Estimation of the Expected Makespan of a POS 285

A layered graph is a directed acyclic graph whose activities are organized in

layers: there are edges between each activity in k-th layer and each activity in
the (k+1)-th layer, and no edge between activities in the same layer. In a layered
graph with constant width, all layers contain the same number of activities.
The rationale behind our simplified model is that, when a POS is constructed,
new edges are added to the original graph so as to prevent possible resource con-
flicts. This process leads to an “iterative flattening” of the POS (see [3]), making
it closer in structure to a layered graph with constant width. The assumption
on the identical minimal and maximal durations is instead introduced only to
increase the tractability.

Evaluating the Model: The effectiveness of using a layered graph as a model for
more complex POSs is difficult to assess in a theoretical fashion, but can be
checked empirically. Basically, if the model works, it should lead to reasonably
accurate predictions of (e.g.) the value of E [T (D)] with worst-case distribution.
This evaluation can be done by: 1) generating graphs with different number
of activities, different width values, and different durations parameters; then 2)
using Algorithm 1 to obtain the expected makespan with worst-case distribution.
Then the results should be compared with the results obtained for a set of POSs
representing solutions to real scheduling problems: if the results are similar, it
will be an indication of the model quality. In the remainder of this section, we
will use this approach to investigate the impact of changing the width on the
value of Ewc [T (D)].

Measuring the Width: The width w of a layered graph is (by definition) the
number of activities in each layer. The width of a real POS is instead a much
fuzzier concept. In our evaluation, an approximate width value is computed as:

|A| T 1 
w= with: nL = and: avg(di ) = di (8)
nL avg(di ) |A|
ai ∈A

Intuitively, we estimate the width as the ratio between the number of activities
|A| and the (estimated) number of layers nL . The number of layers is obtained by
dividing the makespan with maximal durations T by the average di . By applying
algebraic simplifications, we obtain the compact formula:

di
w = ai ∈A (9)
T
We use the same symbol (i.e. w) for both layered graphs and real POSs, because
Equation (9) returns the “true” width when applied to our simplified model.

Evaluation Setup We generated layered graphs with different numbers of activ-

ities, and for each one we varied the width from |A| (single layer) to 1. In order
to obtain data points for every (integer) value of w, we allow for graphs with
quasi-constant width: namely, if |A| is not a multiple of w, the right-most layers
are permitted to contain w − 1 activities.
286 M. Lombardi et al.

As a “real” counterpart for our empirical evaluation we employ a large bench-

mark of POSs representing solutions of Resource Constrained Project Scheduling
Problems (RCPSP) and Job Shop Scheduling Problems (JSSP). The POSs for
the RCPSP have been obtained by solving the j30, j60, and j90 instances from
the PSPlib [6], respectively having |A| = 30, 60, 90. The POSs for the JSSP have
been obtained by solving the instances in the Taillard benchmark, from size
15 × 15 to 30 × 20: in this case the number of activities ranges from 225 to 600.
Both the RCPSP and the JSSP instances have been solved with a classical
CP approach [1], the SetTimes search strategy [9], and a makespan minimization
objective. The POS have been obtained via the post-processing algorithm from
[13] from all the schedules found during the optimization process, so that our
collection contains 5,753 POSs coming from optimal, slightly suboptimal, and
quite far from optimal schedules.
The instances in the PSPLIB and the Taillard benchmarks are originally
deterministic, so we had to introduce some uncertainty artificially. In particular,
the problem files specify a fixed duration for each activity, that we treat as the
maximal duration di . The values of di and d˜i are instead generated at random.
In this particular experimentation, the minimal duration of all activities was
fixed to 0 for both the real POSs and the layered graphs. The maximum dura-
tion is fixed to 1 for the layered graphs (more on this later). All the average
durations d˜i were randomly generated using a large variety of different schemes
(see Section 3.3 for more details).

Makespan Normalization: Intuitively, the value of Ewc [T (D)] should depend on

the graph structure, on the characteristics of the uncertainty, and on the scale of
the instance. Factoring out the dependence on the scale is important in order to
understand the dependence on the structure and on the uncertainty. After some
attempts, we have found that is possible to normalize the scale of the expected
makespan without introducing distortions by using this simple formula:
Ewc [T (D)]
normalized (Ewc [T (D)]) = (10)
ai ∈A di

i.e. by dividing the original value by the sum of the maximal durations. This
normalization allows to compare on a uniform scale the expected makespan of
graphs with widely different durations and number of activities. For our simpli-
fied model, the calculation makes the value of d completely irrelevant, which is
the reason for fixing the maximal duration to 1 in the experimentation.

Some Results: We have obtained plots reporting the value of Ewc [T (D)] over
the graph width. Figure 1 shows the plots for the simplified model (with 64
activities) on the left, and for all the real POSs on the right. The values of d˜i in
both cases have been generated uniformly at random between di (which is always
0) and 0.5 di . The colors represent the data point density, which is benchmark
dependent and unfortunately not very informative.
The results are very interesting. First, the general shape of the plots does
not appear to depend on the number of activities: this can be inferred by the
 Deterministic Estimation of the Expected Makespan of a POS 287

Fig. 1. Value of Ewc [T (D)] over the graph width for the simplified model (left) and
the real POSs (right)

plot for the real POSs, for which the number of activities ranges from 30 to 600.
For the simplified model the plots with 16 and 32 activities (not reported) are
almost indistinguishable from the plot with 64.
The most important result, however, is that the behavior of the simplified
model and of the real graphs is remarkably similar, in terms of both shape and
scale. This is particularly striking given the wide range of graph structures con-
tained in our benchmark, and suggests that our simplified model is indeed well
suited for studying the behavior of real POSs.
A necessary condition for the similarity to show up is that the values of
the expected durations d˜i must be obtained using the same random generation
scheme. This is not surprising, as it simply means that the characteristics of
the uncertainty change the degree by which the value of w affects the expected
makespan: this aspect will be tackled in the next section.

3.2 Deriving the Estimator

In this section, we use a mix of theoretical and empirical arguments applied to
the simplified model from Section 3.1 to derive the estimator from Equation (7).
In general terms, we are looking for some kind of formula that, based on global
properties of the instance and the schedule can approximately predict the value
of Ewc [T (D)]. Thanks to the simplicity of the layered graph model, this may be
doable via Probability Theory results.

Random POS Construction: Let us assume to have a set of A of activities,

with the corresponding minimal, maximal, and expected durations di , di , and
d˜i . In our simplified model all minimal durations are equal to δ and all maximal
durations are equal to δ: however, in this discussion we will use the di , di notation
whenever possible to maintain a clearer analogy with general POSs.
Since we are looking for a formula that depends on global properties, we can
assume to have access to general information (e.g. the width or the number of
288 M. Lombardi et al.

layers), but not to the location of each activity within the graph1 . Therefore,
we will assume that the layer where each activity appears is determined at
random. This is a stochastic process and we are interested in the expected value
of Ewc [T (D)] w.r.t. the random mapping: i.e., we are dealing with a double
expectation in the form Emap [Ewc [T (D)]].

Single-layer Ewc [T (D)]: As a first step, we will focus on the inner expectation,
i.e. on the computation of E [T (D)] with worst case duration. At this stage, we
can assume the activity positions to be known so that the expected makespan
can be obtained by running Algorithm 1 (or solving P0). Since adjacent layers
are fully connected, processing each layer separately is equivalent to processing
the POS as whole. Now, consider the problem objective in P0:

max z = pωk T (ωk ) (11)
ωk ∈Ω

For a single layer of the simplified model it can be rewritten as:


max z = δ + (δ − δ) p ωk (12)
ωk :∃Di =δ

where we have exploited the fact that ∀ai ∈ A : di = δ, di = δ, and therefore the
single-layer makespan is equal to δ unless at least one activity takes maximal
duration. Therefore, the whole problem can be rewritten for a single-layer as:

P1 : max z = δ + (δ − δ) p ωk (13)
ωk :∃Di =δ
 d˜i − di
subject to: p ωk ≤ ∀ai ∈ A (14)
Di − Di
ωk :Di =di

p ωk ≤ 1 (15)
ωk ∈Ω 

p ωk ≥ 0 ∀ωk ∈ Ω  (16)

where A is the set of activities in the layer and Ω  = ai ∈A {di , di }. As already
mentioned, we have not replaced di , di with δ, δ unless it was strictly necessary
to obtain an important simplification.
Problem P1 is simple enough to admit a closed form solution. In particular, all
scenarios where at least one activity takes the maximal duration are symmetrical.
Moreover, it is always worthwhile to increase the probability of such scenarios
as much as possible. Therefore the solution of P1 is always given by:
 
 d˜i − d
∗ i
z = δ + (δ − δ) min 1, (17)
a ∈A
Di − Di
i


where (d˜i − di )/ Di − Di is the probability that ai takes maximal duration.
1
If we had access to this, we could use Algorithm 1.
 Deterministic Estimation of the Expected Makespan of a POS 289

Expectation w.r.t. the Random Mapping In our stochastic POS construction

process, the choice

of the activities in each layer is random. Therefore the value
of (d˜i − di )/ Di − Di can be seen as a random variable in Equation (17):
 w−1 

∗
z = δ + (δ − δ) min 1, Qk (18)
k=0

and selecting an activity for the current layer is equivalent to instantiating a Qk .

For the first instantiation, the probability to pick a certain value of vk for
Qk is given by the number of activities in A such that (d˜i − di )/(Di − Di ) =
vk . For subsequent instantiations the probabilities will be different, because an
activity cannot be inserted in two positions in the POS. In probabilistic terms,
the Qk variables are non-independent, which complicates enormously the task
of defining a probabilistic model.
We address this issue by simply disregarding the constraint and allowing
activity duplication to occur. We have empirically checked the accuracy of this
approximation, which appear to be very good in terms of expected value. This
last simplification is enough to ensure that: 1) the Qk variables are independent;
and 2) the Qk variables have the same discrete distribution, which is defined by
the frequency of occurrence of each (d˜i − di )/(Di − Di ) among the ai ∈ A.
Having independent and identically distributed Qk implies that the value
of Equation (18) will be identical for all the layers. Therefore a approximate
formula for Emap [Ewc [T (D)]] of the random POS construction process is:
  w−1 

nL E δ + (δ − δ) min 1, Qk (19)
k=0

where we recall that nL = |A|/w is the number of layers.

The Estimator Formula: Computing the expected value in Equation (19) is

non-trivial, due to the presence of the “min” operator. It is however possible to
address this issue by introducing some additional approximations. After explor-
ing several alternatives, we have settled for replacing the sum of Qk variables
with a product. By exploiting the linearity of the expectation operator, we get:
nL δ + nL (δ − δ)E [min (1, w Qk )] (20)
where it should be noted that nL δ and nL δ correspond for the simplified model
to T and T . Since Qk has a discrete distribution, the expected value can be
computed by: 1) multiplying by w every value in the distribution of Qk ; then 2)
capping each result at 1; and finally 3) computing the average over all activities:
 
1  d˜i − di
E [min (1, wQk )] = min 1, w (21)
|A| di − di
a ∈A
i

by combining Equation (20) and (21), replacing nL δ with T , replacing nL δ with

T , and replacing w with the formula from Equation (9), we obtain the estimator
original estimator formula.
290 M. Lombardi et al.

It is useful to have a visual interpretation of

the α and β parameters of the beta distri-
bution. The plot on the left depicts five dif-
ferent shapes of the probability density func-
tion, which correspond to the configurations
used in the experiments. The α parameter is
intuitively related to the probability of the
lower values, while β is related to the higher
values. If α = β = 1, all values are equally
likely and we get a uniform distribution. If
α < 1 or β < 1 the probability mass tends to
cluster respectively on the lowest and highest
value. Conversely, if α or β are > 1, then the
probability mass tends to move to the cen-
ter, and the distribution assumes a bell-like
shape.

Fig. 2. Visual interpretation of the α and β parameters in the beta distribution.

3.3 Empirical Evaluation of the Estimator

We have evaluated the estimator on the same benchmark used for assessing
the effectiveness of the simplified model (see Section 3.1). We have compared
the accuracy of our estimator τ (T , T ) with that of two baseline predictors: 1)
the makespan with worst case durations T (d), which is an upper bound for
Ewc [T (D)]; and 2) and the makespan with expected durations T (d), ˜ which is a
˜
lower bound. In particular T (d) proved to be a surprisingly effective estimator
for the expected makespan in a previous evaluation of ours [2].
We recall that the maximal durations for each POS are specified by the
instance file. The minimal and the average durations are randomly generated,
in this case using a beta distribution Beta(α, β). This was chosen because it is
bounded and very flexible. The parameter configurations that we have employed
are presented in Figure 2. In detail, the values of di and d˜i are obtained as:
di
di = Beta(αd , βd ) d˜i = di + (di − di ) Beta(αd˜, βd˜) (22)
2
where di ranges in [0, di /2], and d˜i ranges in [di , di ]. We consider four parameter
configurations for di and five configurations for d˜i . Our goal was to generate a
large variety of duration scenarios that are somehow representative of real world
conditions. For instance the configuration αd˜ = 3, βd˜ = 3 could model a situation
where the activities are subject to many small disruptions. The configuration
αd˜ = 0.3, βd˜ = 3 leads to average durations very close to the minimal ones,
which may be a model for unlikely, but very disruptive, machine faults.
Table 1 shows the average and standard deviation of the relative prediction
error for the compared estimators. We use percentage errors computed as:
Ewc [T (D)] − x
Err(x) = 100 (23)
Ewc [T (D)]
where x is the prediction given by an estimator. Negative errors represent over-
estimations, while positive errors are under-estimations. As expected, the T (d)
˜ under-estimates.
predictor always over-estimates, while T (d)
 Deterministic Estimation of the Expected Makespan of a POS 291

As the number of activities grows, Ewc [T (D)] gets closer to T (d): intuitively,
since we are reasoning with a worst-case distribution, more activities mean more
things that can go wrong. This trend reflects on the performance of T (d) (which
gets better), and appears to have a adverse effect on the accuracy of τ and T (d). ˜
When the minimal durations tend to be higher (i.e. when αd ≥ 1 and βd < 0),
the range of the random durations is reduced and all predictions are more precise.
The T (d)˜ estimator benefits the most from this situation.
In general, the τ estimator tends to outperform the other estimators: the
accuracy is much better than T (d), and significantly better than T (d) ˜ in most
cases. It must be said that using T (d)˜ as an estimator for Ewc [T (D)] lead to sur-
prisingly good results, especially in terms of standard deviation: this is consistent
with our findings form [2], and deserves further investigation. The τ estimator
really shines when the d˜i values are small, i.e. for αd˜ < 0: this is the situation
that models the presence of machine faults, which is also the best suited for
reasoning in terms of worst-case distribution (since a fault tends to affect all the
activities in the system).

4 Concluding Remarks

We have tackled the problem of estimating the expected makespan of a POS

under duration uncertainty, assuming access to very limited information about
the probability distribution. We have provided an O(n3 ) algorithm for comput-
ing the largest expected makespan that is compatible with the available informa-
tion, and we have proposed a linear-time estimator to approximately predict the
same quantity. Both the algorithm and the estimator could be employed within
an optimization approach to obtain an estimate of the expected makespan, in
cases where scenario-based optimization has prohibitive complexity, or extensive
duration data is not available.
Our results are very encouraging, and we believe that further improvements
are possible. First, it should be possible to exploit better the striking similarity
of behavior between real POSs and layered graphs, which we observed in [2] and
confirmed in this paper. Second, it may be possible to obtain a more accurate
and still efficient approach by exploiting additional information (e.g. a covariance
matrix). Third, we have largely disregarded the strong correlation between T (d) ˜
and the expected makespan, which we had observed in [2] and was confirmed in
our experimental evaluation: it should be possible to exploit this correlation in
order to improve the accuracy of the estimate.
Additional possible directions for future research involve the design of an
estimator for the makespan variance, or for specific quantiles in its distribution.
Such an estimator would allow to state “approximate chance constraints”, and
allow an efficient solution of many more practical problems. Finally, it may be
possible to replace our scalar estimate with ranges, similarly to what is done in
statistics with confidence intervals. A the present time, we have not started to
investigate any of the last two mentioned research directions.
292 M. Lombardi et al.

5 Appendix A
This appendix contains the proof to Theorem 1.
Theorem 1. There exists a worst-case distribution such that, in each scenario
with non-zero probability, every Di takes either the value di or the value di .
Proof (By contradiction and induction). Let us assume that the worst case distri-
bution contains a scenario ωk such that one Di is equal to some value vk ∈ (di , di )
and P (ωk ) > 0. Let ω k and ω k be two scenarios that are identical to ωk in all
assignments, except that Di takes respectively its minimal and maximal value.
Let us consider shifting the probability P (ωk ) from ωk to ω k and ω k , and let
ΔP (ω k ) ≥ 0 and ΔP (ω k ) ≥ 0 be the two probability variations. Then it must
hold:
ΔP (ωk ) + ΔP (ωk ) = P (ωk ) (24)
in order to preserve the total probability mass, and:
di ΔP (ωk ) + di ΔP (ωk ) = vk P (ωk ) (25)

to keep the expected value d˜i unchanged. We recall that in each scenario the
makespan is given by the length of the critical path, and that the expected
makespan is given by the integral (or sum) of T (D(ωk ))P (ωk ). Now, let us
distinguish some cases:
1) ai is on the critical path in ωk . This implies that ai is on the critical
path also in ω k , and therefore shifting the probability causes the expected
makespan to increase by (di − vk )ΔP (ω k ) units. Then:
– If ai is on the critical path in ωk , then the expected makespan is also
decreased by (vk − di )ΔP (ω k ) units. However, by combining Equa-
tions (24) and (25) we obtain that:
(di − vk )ΔP (ω k ) − (vk − di )ΔP (ω k ) = 0 (26)
i.e. the net change of the expected makespan is zero.
– If ai is not on the critical path in ω k , then the expected makespan is
decreased by less than (vk − di )ΔP (ω k ) units, and the net change is
positive (i.e. the expected makespan is increased).
2) ai is not on the critical path in ωk . This implies that ai is not on the
critical path also in ω k , and therefore that shifting probability to ω k leaves
the expected makespan unchanged. Then:
– If ai is on the critical path in ω k , the expected makespan is increased by
a quantity in the range [0, (di − vk )ΔP (ω k )).
– If ai is not on the critical path in ω k , the expected makespan is again
unchanged.
Therefore, by reducing to zero the probability P (ωk ), the expected makespan
either is increased or stays unchanged. This procedure can be repeated until
there is no scenario with non-zero probability where Di takes a value different
from di or di .

 Deterministic Estimation of the Expected Makespan of a POS 293

Table 1. Partial horizontal line

J30
(ad̃i , bd̃i ) −→ (0.3,0.3) (0.3,3) (1,1) (3,3)
(adi , bdi ) τ T(d̃) T(d) τ T(d̃) T(d) τ T(d̃) T(d) τ T(d̃) T(d)
Err -1.08 5.06 -37.19 -3.91 9.37 -148.18 -7.75 11.32 -31.82 -12.23 15.87 -27.92
(0.3,0.3)
Std 7.77 2.72 13.38 5.36 4.57 32.46 6.94 3.79 10.29 6.51 3.99 8.27
Err 1.55 7.54 -52.16 2.44 24.83 -352.00 -6.83 17.14 -41.85 -13.18 24.20 -35.18
(0.3,3.0)
Std 10.07 3.87 19.37 6.77 8.80 113.50 8.32 4.87 13.75 8.06 4.61 10.57
Err -1.62 2.66 -26.68 -3.22 3.54 -85.42 -6.38 6.22 -23.62 -9.90 9.33 -21.15
(3.0,0.3)
Std 5.55 1.67 8.54 3.38 2.19 10.22 4.85 2.59 7.03 4.82 2.97 6.01
Err -0.70 5.14 -37.30 -3.02 10.13 -153.98 -7.04 11.33 -32.04 -11.91 16.20 -27.98
(1.0,1.0)
Std 7.71 2.66 13.60 4.79 4.52 30.28 6.42 3.64 10.17 6.28 3.88 8.10
Err -0.03 5.29 -38.33 -2.17 10.25 -160.43 -6.95 11.54 -32.73 -11.72 16.46 -28.09
(3.0,3.0)
Std 7.42 2.75 13.65 4.53 4.39 29.59 6.59 3.62 10.15 6.44 3.75 8.09
J60
Err -2.92 6.16 -33.85 -7.35 12.62 -133.52 -11.11 13.10 -27.88 -16.20 18.25 -23.59
(0.3,0.3)
Std 6.66 2.45 10.35 5.12 4.32 23.12 5.81 3.25 7.72 5.33 3.38 6.15
Err 0.07 8.82 -44.82 -0.34 33.56 -268.74 -11.47 19.79 -35.42 -18.73 27.21 -29.12
(0.3,3.0)
Std 8.10 3.39 14.08 6.30 7.83 66.40 7.54 4.06 9.63 6.62 3.68 7.42
Err -3.19 3.32 -25.27 -6.40 4.93 -81.94 -9.31 7.54 -21.56 -13.37 10.95 -18.81
(3.0,0.3)
Std 4.90 1.61 7.01 3.52 2.32 8.50 4.46 2.32 5.54 4.27 2.57 4.81
Err -2.34 6.10 -33.65 -6.61 13.79 -139.56 -10.83 13.31 -28.08 -16.30 18.50 -23.87
(1.0,1.0)
Std 6.57 2.47 10.04 4.82 4.53 22.94 5.77 3.21 7.57 5.50 3.32 6.21
Err -1.84 6.22 -34.20 -5.66 14.53 -145.00 -10.48 13.52 -28.10 -16.10 18.69 -23.90
(3.0,3.0)
Std 6.63 2.46 10.24 4.85 4.61 22.46 5.72 3.16 7.41 5.37 3.17 6.14
J90
Err -3.94 6.59 -32.30 -9.82 14.67 -125.92 -12.44 13.87 -26.23 -17.41 19.20 -22.03
(0.3,0.3)
Std 6.28 2.35 9.03 5.20 4.16 19.42 5.43 2.92 6.54 4.93 3.17 5.54
Err -1.16 9.62 -42.30 -1.25 38.04 -233.89 -13.54 20.81 -33.02 -20.17 28.48 -26.67
(0.3,3.0)
Std 7.39 3.19 11.32 6.35 6.83 52.10 6.77 3.74 8.34 6.04 3.35 6.54
Err -3.71 3.62 -24.17 -8.44 5.78 -79.91 -10.55 8.16 -20.61 -14.43 11.71 -17.88
(3.0,0.3)
Std 4.59 1.55 6.08 3.70 2.42 8.03 4.26 2.25 5.03 3.88 2.51 4.24
Err -3.46 6.67 -32.51 -8.80 15.97 -131.57 -12.31 14.23 -26.29 -17.24 19.42 -21.97
(1.0,1.0)
Std 5.85 2.36 8.44 5.08 4.33 19.47 5.53 2.95 6.59 4.86 3.02 5.40
Err -2.73 6.83 -32.88 -7.77 17.12 -136.15 -12.12 14.35 -26.59 -17.36 19.80 -22.28
(3.0,3.0)
Std 5.84 2.36 8.55 5.07 4.54 19.52 5.47 2.91 6.63 5.00 2.92 5.47
Taillard
Err -7.34 11.16 -27.74 -14.69 27.53 -94.63 -12.85 20.99 -17.84 -12.29 26.59 -12.63
(0.3,0.3)
Std 4.16 2.36 5.69 3.69 3.77 10.76 3.51 2.64 4.09 3.42 2.26 3.50
Err -5.74 15.93 -34.45 -6.04 55.75 -134.53 -14.01 29.17 -20.96 -13.88 36.39 -14.33
(0.3,3.0)
Std 4.66 3.04 6.73 4.14 3.54 17.82 4.07 2.97 4.90 3.69 2.43 3.79
Err -6.97 6.92 -22.08 -13.61 13.95 -66.55 -11.20 13.85 -14.98 -10.62 17.87 -10.87
(3.0,0.3)
Std 3.35 1.76 4.15 3.18 2.68 6.65 2.94 2.14 3.29 2.92 2.10 2.98
Err -7.17 11.29 -28.19 -13.26 29.75 -96.61 -12.84 21.31 -17.97 -12.31 26.88 -12.65
(1.0,1.0)
Std 3.95 2.44 5.27 3.73 3.46 11.00 3.44 2.54 4.00 3.36 2.34 3.46
Err -6.49 11.45 -28.09 -11.98 31.37 -98.08 -12.72 21.43 -17.98 -12.28 27.13 -12.62
(3.0,3.0)
Std 3.92 2.45 5.34 3.81 3.39 11.24 3.51 2.56 4.05 3.35 2.24 3.44
294 M. Lombardi et al.

References
1. Baptiste, P., Le Pape, C., Nuijten, W.: Constraint-based scheduling. Kluwer Aca-
demic Publishers (2001)
2. Bonfietti, A., Lombardi, M., Milano, M.: Disregarding duration uncertainty in
partial order schedules? yes, we can!. In: Simonis, H. (ed.) CPAIOR 2014. LNCS,
vol. 8451, pp. 210–225. Springer, Heidelberg (2014)
3. Cesta, A., Oddi, A., Smith, S.F.: Iterative flattening: a scalable method for solving
multi-capacity scheduling problems. In: AAAI/IAAI, pp. 742–747 (2000)
4. Kellerer, H., Pferschy, U., Pisinger, D.: Knapsack problems. Springer Science &
Business Media (2004)
5. Kleywegt, A.J., Shapiro, A., Homem-de Mello, T.: The sample average approxima-
tion method for stochastic discrete optimization. SIAM Journal on Optimization
12(2), 479–502 (2002)
6. Kolisch, R., Sprecher, A.: Psplib-a project scheduling problem library: Or software-
orsep operations research software exchange program. European Journal of Oper-
ational Research 96(1), 205–216 (1997)
7. Laborie, P.: Complete MCS-based search: application to resource constrained
project scheduling. In: Proc. of IJCAI, pp. 181–186. Professional Book Center
(2005)
8. Laborie, P., Godard, D.: Self-adapting large neighborhood search: application to
single-mode scheduling problems. In: Proc. of MISTA (2007)
9. Le Pape, C., Couronné, P., Vergamini, D., Gosselin, V.: Time-versus-capacity com-
promises in project scheduling. AISB Quarterly, 19 (1995)
10. Lombardi, M., Milano, M., Benini, L.: Robust scheduling of task graphs under
execution time uncertainty. IEEE Trans. Computers 62(1), 98–111 (2013)
11. Morris, P., Muscettola, N., Vidal, T.: Dynamic control of plans with temporal
uncertainty. In: Proc. of IJCAI, pp. 494–499. Morgan Kaufmann Publishers Inc.
(2001)
12. Policella, N., Cesta, A., Oddi, A., Smith, S.F.: From precedence constraint posting
to partial order schedules: A CSP approach to Robust Scheduling. AI Communi-
cations 20(3), 163–180 (2007)
13. Policella, N., Smith, S.F., Cesta, A., Oddi, A.: Generating robust schedules through
temporal flexibility. In: Proc. of ICAPS, pp. 209–218 (2004)
14. Tarim, S.A., Manandhar, S., Walsh, T.: Stochastic constraint programming: A
scenario-based approach. Constraints 11(1), 53–80 (2006)
15. Vidal, T.: Handling contingency in temporal constraint networks: from consistency
to controllabilities. Journal of Experimental & Theoretical Artificial Intelligence
11(1), 23–45 (1999)
A Parallel, Backjumping Subgraph Isomorphism
Algorithm Using Supplemental Graphs

Ciaran McCreesh(B) and Patrick Prosser

University of Glasgow, Glasgow, Scotland

c.mccreesh.1@research.gla.ac.uk, patrick.prosser@glasgow.ac.uk

Abstract. The subgraph isomorphism problem involves finding a

pattern graph inside a target graph. We present a new bit- and thread-
parallel constraint-based search algorithm for the problem, and experi-
ment on a wide range of standard benchmark instances to demonstrate its
effectiveness. We introduce supplemental graphs, to create implied con-
straints. We use a new low-overhead, lazy variation of conflict directed
backjumping which interacts safely with parallel search, and a counting-
based all-different propagator which is better suited for large domains.

1 Introduction
The subgraph isomorphism family of problems involve “finding a copy of” a
pattern graph inside a larger target graph; applications include bioinformatics
[3], chemistry [31], computer vision [12,37], law enforcement [7], model check-
ing [33], and pattern recognition [9]. These problems have natural constraint
programming models: we have a variable for each vertex in the pattern graph,
with the vertices of the target graph being the domains. The exact constraints
vary depending upon which variation of the problem we are studying (which we
discuss in the following section), but generally there are rules about preserving
adjacency, and an all-different constraint across all the variables.
This constraint-based search approach dates back to works by Ullmann [39]
and McGregor [25], and was improved upon in the LV [20] and VF2 [11] algo-
rithms. More recently, ILF [41], LAD [36] and SND [1] are algorithms which take
a “deep thinking” approach, using strong inference at each stage of the search.
This is powerful, but we observe LAD or SND sometimes make less than one
recursive call per second with larger target graphs, and cannot always explore
enough of the search space to find a solution in time. This motivates an alter-
native approach: on the same hardware, we will be making 104 to 106 recursive
calls per core per second. The main features of our algorithm are:

1. We introduce supplemental graphs, which generalise some of the ideas in

SND. The key idea is that a subgraph isomorphism i : P  T induces a
subgraph isomorphism F (i) : F (P )  F (T ), for certain functors F . This is
C. McCreesh—This work was supported by the Engineering and Physical Sciences
Research Council [grant number EP/K503058/1].

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 295–312, 2015.
DOI: 10.1007/978-3-319-23219-5 21
296 C. McCreesh and P. Prosser

used to generate implied constraints: we may now look for a mapping which
is simultaneously a subgraph isomorphism between several carefully selected
pairs of graphs.
2. We use weaker inference than LAD and SND: we do not achieve or maintain
arc consistency. We introduce a cheaper, counting-based all-different prop-
agator which has better scalability for large target graphs, and which has
very good constant factors on modern hardware thanks to bitset encodings.
3. We describe a clone-comparing variation of conflict-directed backjumping,
which does not require conflict sets. We show that an all-different propagator
can produce reduced conflict explanations, which can improve backjumping.
4. We use thread-parallel preprocessing and search, to make better use of mod-
ern multi-core hardware. We explain how parallel search may interact safely
with backjumping. We use explicit, non-randomised work stealing to offset
poor early heuristic choices during search.

Although weaker propagation and backjumping have fallen out of fashion in gen-
eral for constraint programming, here this approach usually pays off. In section 4
we show that over a large collection of instances commonly used to compare sub-
graph isomorphism algorithms, our solver is the single best.

2 Definitions, Notation, and a Proposition

Throughout, our graphs are finite, undirected, and do not have multiple edges
between pairs of vertices, but may have loops (an edge from a vertex to itself).
We write V(G) for the vertex set of a graph G, and N(G, v) for the neighbours
of a vertex v in G (that is, the vertices adjacent to v). The degree of v is the
cardinality of its set of neighbours. The neighbourhood degree sequence of v,
denoted S(G, v), is the sequence consisting of the degrees of every neighbour
of v, from largest to smallest. A vertex is isolated if it has no neighbours. By
v ∼G w we mean vertex v is adjacent to vertex w in graph G. We write G[V ]
for the subgraph of G induced by a set of vertices V .
A non-induced subgraph isomorphism is an injective mapping i : P  T
from a graph P to a graph T which preserves adjacency—that is, if v ∼P w
then we require i(v) ∼T i(w) (and thus if v has a loop, then i(v) must have a
loop). The non-induced subgraph isomorphism problem is to find such a mapping
from a given pattern graph P to a given target graph T . (The induced subgraph
isomorphism problem additionally requires that if v ∼P w then i(v) ∼T i(w),
and variants also exist for directed and labelled graphs; we discuss only the
non-induced version in this paper. All these variants are NP-complete.)
If R and S are sequences of integers, we write R  S if there exists a subse-
quence of S with length equal to that of R, such that each element in R is less
than or equal to the corresponding element in S. For a set U and element v, we
write U − v to mean U \ {v}, and U + v to mean U ∪ {v}.
A path in a graph is a sequence of distinct vertices, such that each successive
pair of vertices are adjacent; we also allow a path from a vertex to itself, in which
 A Parallel, Backjumping Subgraph Isomorphism Algorithm 297

case the first and last vertices in the sequence are the same (and there is a cycle).
The distance between two vertices is the length of a shortest path between them.
We write Gd for the graph with vertex set V(G), and edges between v and w
if the distance between v and w in G is at most d. We introduce the notation
G[c,l] for the graph with vertex set V(G), and edges between vertices v and w
(not necessarily distinct) precisely if there are at least c paths of length exactly
l between v and w in G. The following proposition may easily be verified by
observing that injectivity means paths are preserved:

Proposition 1. Let i : P  T be a subgraph isomorphism. Then i is also

1. a subgraph isomorphism id : P d  T d for any d ≥ 1, and
2. a subgraph isomorphism i[c,l] : P [c,l]  T [c,l] for any c, l ≥ 1.

The (contrapositive of the) first of these facts is used by SND, which dynamically
performs distance-based filtering during search. We will instead use the second
fact, at the top of search, to generate implied constraints.

3 A New Algorithm
Algorithm 1 describes our approach. We begin (line 3) with a simple check that
there are enough vertices in the pattern graph for an injective mapping to exist.
We then (line 4) discard isolated vertices in the pattern graph—such vertices may
be greedily assigned to any remaining target vertices after a solution is found.
This reduces the number of variables which must be copied when branching. Next
we construct the supplemental graphs (line 5) and initialise domains (line 6).
We then (line 7) use a counting-based all-different propagator to reduce these
domains further. Finally, we perform a backtracking search (line 8). Each of
these steps is elaborated upon below.

3.1 Preprocessing and Initialisation

Following Proposition 1, in Algorithm 2 we construct a sequence of supplemental
graph pairs from our given pattern and target graph. We will then search for
a mapping which is simultaneously a mapping from each pattern graph in the

Algorithm 1. A non-induced subgraph isomorphism algorithm

1 nonInducedSubgraphIsomorphism (Graph P, Graph T ) → Bool
2 begin
3 if | V(P)| > | V(T )| then return false
4 Discard isolated vertices in P
5 L ← createSupplementalGraphList(P, T )
6 D ← init(V(P), V(T ), L)
7 if countingAllDifferent(D) = Success then return false
8 return search(L, D) = Success
298 C. McCreesh and P. Prosser

Algorithm 2. Supplemental graphs for Algorithm 1

1 createSupplementalGraphList (Graph P, Graph T ) → GraphPairs
2 begin 
return (P, T ), (P [1,2] , T [1,2] ), (P [2,2] , T [2,2] ), (P [3,2] , T [3,2] ),
3 
(P [1,3] , T [1,3] ), (P [2,3] , T [2,3] ), (P [3,3] , T [3,3] )

sequence to its paired target graph—this gives us implied constraints, leading to

additional filtering during search.
Our choice of supplemental graphs is somewhat arbitrary. We observe that
distances of greater than 3 rarely give additional filtering power, and constructing
G[c,4] is computationally very expensive (for unbounded l, the construction is
NP-hard). Checking c > 3 is also rarely beneficial. Our choices work reasonably
well in general on the wide range of benchmark instances we consider, but can
be expensive for trivial instances—thus there is potential room to improve the
algorithm by better selection on an instance by instance basis [23].
Algorithm 3 is responsible for initialising domains. We have a variable for
each vertex in the (original) pattern graph, with each domain being the vertices
in the (original) target graph. It is easy to see that a vertex of degree d in the
pattern graph P may only be mapped to a vertex in the target graph T of degree
d or higher: this allows us to perform some initial filtering. By extension, we may
use compatibility of neighbourhood degree sequences: v may only be mapped
to w if S(P, v)  S(T, w) [41]. Because any subgraph isomorphism P  T is
also a subgraph isomorphism F (P )  F (T ) for any of our supplemental graph
constructions F , we may further restrict initial domains by considering only the
intersection of filtered domains using each supplemental graph pair individually
(line 5). At this stage, we also enforce the “loops must be mapped to loops”
constraint.
Following this filtering, some target vertices may no longer appear in any
domain, in which case R will be reduced on line 6. If this happens, we iteratively
repeat the domain construction, but do not consider any target vertex no longer

Algorithm 3. Variable initialisation for Algorithm 1

1 init (Vertices V , Vertices R, GraphPairs L) → Domains
2 begin
3 repeat
4 foreach v ∈ V do  

5 Dv ← (P, T )∈L w ∈ R : v ∼ v ⇒ w ∼ w ∧ S(P, v) S(T [R], w)
P T

6 R ← v∈V Dv
7 until R is unchanged
8 return D
 A Parallel, Backjumping Subgraph Isomorphism Algorithm 299

Algorithm 4. Recursive search for Algorithm 1

1 search (GraphPairs L, Domains D) → Fail F or Success
2 begin
3 if D = ∅ then return Success
4 Dv ← a domain in D with minimum size, tiebreaking on static degree in P
5 F ← {v}
6 foreach v  ∈ Dv ordered by static degree in T do
7 D ← clone(D)
8 case assign(L, D , v, v  ) of
9 Fail F  then F ← F ∪ F 
10 Success then
11 case search(L, D − Dv ) of
12 Success then return Success
13 Fail F  then
14 if  w ∈ F  such that Dw = Dw

then return Fail F 

15 F ←F ∪F

16 return Fail F

in R when calculating degree sequences. (Note that for performance reasons, we

do not recompute supplemental graphs when this occurs.)

3.2 Search and Inference

Algorithm 4 describes our recursive search procedure. If every variable has

already been assigned, we succeed (line 3). Otherwise, we pick a variable (line 4)
to branch on by selecting the variable with smallest domain, tiebreaking on
descending static degree only in the original pattern graph (we tried other vari-
ations, including using supplemental graphs for calculating degree, and domain
over degree, but none gave an overall improvement). For each value in its domain
in turn, ordered by descending static degree in the target graph [13], we try
assigning that value to the variable (line 8). If we do not detect a failure, we
recurse (line 11).
The assignment and recursive search both either indicate success, or return
a nogood set of variables F which cannot all be instantiated whilst respecting
assignments which have already been made. This information is used to prune
the search space: if a subproblem search has failed (line 13), but the current
assignment did not remove any value from any of the domains involved in the
discovered nogood (line 14), then we may ignore the current assignment and
backtrack immediately. In fact this is simply conflict-directed backjumping [27]
in disguise: rather than explicitly maintaining conflict sets to determine culprits
(which can be costly when backjumping does nothing [2,14]), we lazily create
the conflict sets for the variables in F  as necessary by comparing D before the
current assignment with the D created after it. Finally, as in backjumping, if
300 C. McCreesh and P. Prosser

Algorithm 5. Variable assignment for Algorithm 5

1 assign (GraphPairs L, Domains D, Vertex v, Vertex v  ) → Fail F or Success
2 begin
3 Dv ← {v  }
4 foreach Dw ∈ D − Dv do
5 Dw ← Dw − v 
6 foreach (P, T ) ∈ L do
7 if v ∼P w then Dw ← Dw ∩ N(T, v  )
8 if Dw = ∅ then return Fail {w}
9 return countingAllDifferent(D)

none of the assignments are possible, we return with a nogood of the current
variable (line 5) combined with the union of the nogoods of each failed assignment
(line 9) or subsearch (line 15).
For assignment and inference, Algorithm 5 gives the value v  to the domain
Dv (line 3), and then infers which values may be eliminated from the remaining
domains. Firstly, no other domain may now be given the value v  (line 5). Sec-
ondly, for each supplemental graph pair, any domain for a vertex adjacent to v
may only be mapped to a vertex adjacent to v  (line 7). If any domain gives a
wipeout, then we fail with that variable as the nogood (line 8).
To enforce the all-different constraint, it suffices to remove the assigned value
from every other domain, as we did in line 5. However, it is often possible to
do better. We can sometimes detect that an assignment is impossible even if
values remain in each variable’s domain (if we can find a set of n variables whose
domains include strictly less than n values between them, which we call a failed
Hall set), and we can remove certain variable-value assignments that we can
prove will never occur (if we can find a set of n variables whose domains include
only n values between them, which we call a Hall set, then those values may be
removed from the domains of any other variable). The canonical way of doing
this is to use Régin’s matching-based propagator [30].
However, matching-based filtering is expensive and may do relatively little,
particularly when domains are large, and the payoff may not always be worth the
cost. Various approaches to offsetting this cost while maintaining the filtering
power have been considered [15]. Since we are not maintaining arc consistency in
general, we instead use an intermediate level of inference which is not guaranteed
to identify every Hall set: this can be thought of as a heuristic towards the
matching approach. This is described in Algorithm 6.
The algorithm works by performing a linear pass over each domain in turn,
from smallest cardinality to largest (line 4). The H variable contains the union
of every Hall set detected so far; initially it is empty. The A set accumulates the
union of domains seen so far, and n contains the number of domains contributing
to A. For each new domain we encounter, we eliminate any values present in
previous Hall sets (line 6). We then add that domain’s values to A and increment
 A Parallel, Backjumping Subgraph Isomorphism Algorithm 301

Algorithm 6. Counting-based all-different propagation

1 countingAllDifferent (Domains D) → Fail F or Success
2 begin
3 (F, H, A, n) ← (∅, ∅, ∅, 0)
4 foreach Dv ∈ D from smallest cardinality to largest do
5 F ←F +v
6 Dv ← Dv \ H
7 (A, n) ← (A ∪ Dv , n + 1)
8 if Dv = ∅ or |A| < n then return Fail F
9 if |A| = n then (H, A, n) ← (H ∪ A, ∅, 0)
10 return Success

n (line 7). If we detect a failed Hall set, we fail (line 8). If we detect a Hall set,
we add those values to H, and reset A and n, and keep going (line 9).
It is important to note that this approach may fail to identify some Hall
sets, if the initial ordering of domains is imperfect. However, the algorithm runs
very quickly in practice: the sorting step is O(v log v) (where v is the number of
remaining variables), and the loop has complexity O(vd) (where d is the cost of
a bitset operation over a target domain, which we discuss below). We validate
this trade-off experimentally in the following section.
In case a failure is detected, the F set of nogoods we return need only include
the variables processed so far, not every variable involved in the constraint. This
is because an all-different constraint implies an all-different constraint on any
subset of its variables. A smaller set of nogoods can increase the potential for
backjumping (and experiments verified that this is beneficial in practice).
We have been unable to find this algorithm described elsewhere in the liter-
ature, although a sort- and counting-based approach has been used to achieve
bounds consistency [28] (but our domains are not naturally ordered) and as a
preprocessing step [29]. Bitsets (which we discuss below) have also been used to
implement the matching algorithm [19].

3.3 Bit- and Thread-Parallelism

The use of bitset encodings for graph algorithms to exploit hardware parallelism
dates back to at least Ullmann’s algorithm [39], and remains an active area of
research [32,40]. We use bitsets here: our graphs are stored as arrays of bit vec-
tors, our domains are bit vectors, the neighbourhood intersection in Algorithm 5
is a bitwise-and operation, the unions in Algorithm 4 and Algorithm 6 are bitwise-
or operations, and the cardinality check in Algorithm 6 is a population count
(this is a single instruction in modern CPUs).
In addition to the SIMD-like parallelism from bitset encodings, we observed
two opportunities for multi-core thread parallelism in the algorithm:
302 C. McCreesh and P. Prosser

Graph and domain construction. We may parallelise the outer for loops involved
in calculating neighbourhood degree sequences and in initialising the domains
of variables in Algorithm 3. Similarly, constructing each supplemental graph in
Algorithm 2 involves an outer for loop, iterating over each vertex in the input
graph. These loops may also be parallelised, with one caveat: we must be able
to add edges to (but not remove edges from) the output graph safely, in parallel.
This may be done using an atomic “or” operation.

Search. Viewing the recursive calls made by the search function in Algorithm 4
as forming a tree, we may explore different subtrees in parallel. The key points
are:

1. We do not know in advance whether the foreach loop (Algorithm 4 line 6)

will exit early (either due to a solution being found, or backjumping). Thus
our parallelism is speculative: we make a single thread always preserve the
sequential search order, and use any additional threads to precompute sub-
sequent entries in the loop which might be used. This may mean we get no
speedup at all, if our speculation performs work which will not be used.
2. The search function, parallelised without changes, could attempt to exit
early due to backjumping. We rule out this possibility by refusing to pass
knowledge to the left: that is, we do not allow speculatively-found backjump-
ing conditions to change the return value of search. This is for safety [38]
and reproducibility: value-ordering heuristics can alter the performance of
unsatisfiable instances when backjumping, and allowing parallelism to select
a different backjump set could lead to an absolute slowdown [26]. To avoid
this possibility, when a backjump condition is found, we must cancel any
speculative work being done to the right of its position, and cannot cancel
any ongoing work to the left. This means that unlike in conventional back-
tracking search without learning, we should not expect a linear speedup for
unsatisfiable instances.
(In effect we are treating the foreach loop as a parallel lazy fold, so that a
subtree does not depend upon items to its left. Backjumping conditions are
left-zero elements [21], although we do not have a unique zero.)
3. If any thread finds a solution, we do succeed immediately, even if this involves
passing knowledge to the left. If there are multiple solutions, this can lead to
a parallel search finding a different solution to the one which would be found
sequentially—since the solution we find is arbitrary, this is not genuinely
unsafe. However, this means we could witness a superlinear (greater than n
from n threads) speedup for satisfiable instances [4].
4. For work stealing, we explicitly prioritise subproblems highest up and then
furthest left in the search tree. This is because we expect our value-ordering
heuristics to be weakest early on in search [18], and we use parallelism to
offset poor choices early on in the search [6,24].
 A Parallel, Backjumping Subgraph Isomorphism Algorithm 303

4 Experimental Evaluation
Our algorithm was implemented1 in C++ using C++11 native threads, and was
compiled using GCC 4.9.0. We performed our experiments on a machine with
dual Intel Xeon E5-2640 v2 processors (for a total of 16 cores, and 32 hardware
threads via hyper-threading), running Scientific Linux 6.6. For the comparison
with SND in the following section, we used Java HotSpot 1.8.0 11. Runtimes
include preprocessing and thread launch costs, but not the time taken to read
in the graph files from disk (except in the case of SND, which we were unable
to instrument).
For evaluation, we used the same families of benchmark instances that were
used to evaluate LAD [36] and SND [1]. The “LV” family [20] contains graphs
with various interesting properties from the Stanford Graph Database, and
the 793 pattern/target pairs give a mix of satisfiable and unsatisfiable queries.
The “SF” family contains 100 scale-free graph pairs, again mixing satisfiable
and unsatisfiable queries. The remainder of these graphs come from the Vflib
database [11]: the “BVG” and “BVGm” families are bounded degree graphs (540
pairs all are satisfiable), “M4D” and “M4Dr” are four-dimensional meshes (360
pairs, all satisfiable), and the “r” family is randomly generated (270 pairs, all
satisfiable). We expanded this suite with 24 pairs of graphs representing image
pattern queries [12] (which we label “football”), and 200 randomly selected pairs
from each of a series of 2D image (“images”) and 3D mesh (“meshes”) graph
queries [37]. The largest number of vertices is 900 for a pattern and 5,944 for a
target, and the largest number of edges is 12,410 for a pattern and 34,210 for a
target; some of these graphs do contain loops. All 2,487 instances are publicly
available in a simple text format2 .

4.1 Comparison with Other Solvers

We compare our implementation against the Abscon 609 implementation of SND
(which is written in Java) [1], Solnon’s C implementation of LAD [36], and the
VFLib C implementation of VF2 [11]. (The versions of each of these solvers we
used could support loops in graphs correctly.)
Note that SND is not inherently multi-threaded, but the Java 8 virtual
machine we used for testing makes use of multiple threads for garbage collec-
tion even for sequential code. On the one hand, this could be seen as giving SND
an unfair advantage. However, nearly all modern CPUs are multi-core anyway,
so one could say that it is everyone else’s fault for not taking advantage of these
extra resources. We therefore present both sequential (from a dedicated imple-
mentation, not a threaded implementation running with only a single thread)
and threaded results for our algorithm.
In Fig. 1 we show the cumulative performance of each algorithm. The value
of the line at a given time for an algorithm shows the total number of instances
1
source code, data, experimental scripts and raw results available at
https://github.com/ciaranm/cp2015-subgraph-isomorphism
2
http://liris.cnrs.fr/csolnon/SIP.html
304 C. McCreesh and P. Prosser

Fig. 1. Cumulative number of benchmark instances solved within a given time, for
different algorithms: at time t, the value is the size of the set of instances whose
runtime is at most t for that algorithm. Parallel results are using 32 threads on a 16
core hyper-threaded system.

which, individually, were solved in at most that amount of time. Our sequential
implementation beats VF2 for times over 0.2s, LAD for times over 0.6s, and
always beats SND. Our parallel implementation beats VF2 for times over 0.06s,
LAD for times over 0.02s, and always beats SND; parallelism gives us an overall
benefit from 12ms onwards. Finally, removing the supplemental graphs from
our sequential algorithm gives an improvement below 10s (due to the cost of
preprocessing), but is beneficial for longer runtimes.
Fig. 2 presents an alternative perspective of these results. Each point repre-
sents an instance, and the shape of a point shows its family. For the y position
for an instance, we use our sequential (top graph) or parallel (bottom graph) run-
time. For the x position, we use the runtime from the virtual best other solver;
the colour of a point indicates which solver this is. For any point below the x = y
diagonal line, we are the best solver. A limit of 108 ms was used—points along
the outer axes represent timeouts.
Although overall ours is the single best solver, VF2 is stronger on triv-
ial instances. This is not surprising: we must spend time constructing supple-
mental graphs. Thus it may be worth using either VF2 or our own algorithm
 A Parallel, Backjumping Subgraph Isomorphism Algorithm 305

108

LAD
107
VF2
SND
106
Our sequential runtime (ms)

LV (sat)
5 LV (unsat)
10
BVG / BVGm
M4D / M4Dr
104
SF (sat)
SF (unsat)
103 r
football
102 images (sat)
images (unsat)
meshes (sat)
101
meshes (unsat)

100
100 101 102 103 104 105 106 107 108
Virtual best other solver runtime (ms)
8
10

107

106
Our parallel runtime (ms)

105

104

103

102

101

100
100 101 102 103 104 105 106 107 108
Virtual best other solver runtime (ms)

Fig. 2. Above, our sequential runtime compared to the virtual best other sequential
solver, for each benchmark instance; below, the same, with our parallel runtimes and
including parallel solvers. For points below the diagonal line, ours is the best solver for
this instance; for points above the diagonal, the point colour indicates the best solver.
306 C. McCreesh and P. Prosser

without supplemental graphs as a presolver, if short runtimes for trivial instances

is desirable—this may be the case in graph database systems where many trivial
queries must be run [16] (although these systems could cache the supplemen-
tal graphs for targets). These results also suggest potential scope for algorithm
portfolios, or instance-specific configuration: for example, we could omit or use
different supplemental graphs in some cases.

4.2 Parallelism
Fig. 3 shows, for each instance, the speedup obtained from parallelism. Except
at very low sequential runtimes, we see a reasonable general improvement. For
some satisfiable instances, we see strongly superlinear speedups. These instances
are exceptionally hard problems [35]: we would have found a solution quickly,
except for a small number of wrong turns at the top of search. Our work stealing
strategy was able to avoid strong commitment to early value-ordering heuristic
choices, providing an alternative to using more complicated sequential search
strategies to offset this issue. (Some of these results were also visible in Fig. 2,
where we timed out on satisfiable instances which another solver found trivial.)
Some non-trivial satisfiable instances exhibited a visible slowdown. This is
because we were using 32 software threads, to match the advertised number of
hardware hyper-threads, but our CPUs only have 16 “real” cores between them.
For these instances parallelism did not reduce the amount of work required to
find a solution, but did result in a lower rate of recursive calls per second on
the sequential search path—this is similar to the risk of introducing a slower
processing element to a cluster [38]. Even when experimenting with 16 threads,
we sometimes observed a small slowdown due to worse cache and memory bus
performance, and due to the overhead of modifying the code to allow for work
stealing (recall that we are benchmarking against a dedicated sequential imple-
mentation).
In a small number of cases, we observe low speedups for non-trivial unsatisfi-
able instances. These are from cases where backjumping has a substantial effect
on search, making much of our speculative parallelism wasted effort. (Addition-
ally, if cancellation were not to be used, some of these instances would exhibit
large absolute slowdowns.)

4.3 Effects of Backjumping

In Fig. 4 we show the benefits of backjumping: points below the diagonal line
indicate an improvement to runtimes from backjumping. Close inspection shows
that backjumping usually at least pays for itself, or gives a slight improvement.
(This was not the case when we maintained conflict sets explicitly: there, the
overheads lead to a small average slowdown.)
For a few instances, backjumping makes an improvement of several orders of
magnitude. The effects are most visible for some of the LV benchmarks, which
consist of highly structured graphs. This mirrors the conclusions of Chen and
Van Beek [5], who saw that “adding CBJ to a backtracking algorithm . . . can
 A Parallel, Backjumping Subgraph Isomorphism Algorithm 307

108

107
LV (sat)
106 LV (unsat)
BVG / BVGm
Parallel Runtime (ms)

105 M4D / M4Dr

SF (sat)
SF (unsat)
104
r
football
103 images (sat)
images (unsat)
102 meshes (sat)
meshes (unsat)
101

100
100 101 102 103 104 105 106 107 108
Sequential Runtime (ms)

Fig. 3. The effects of parallelism, using 32 threads on a 16 core hyper-threaded system.

Each point is a problem instance; points below the diagonal line indicate a speedup.

108

107
LV (sat)
Runtime with Backjumping (ms)

106 LV (unsat)
BVG / BVGm
105 M4D / M4Dr
SF (sat)
SF (unsat)
104
r
football
103 images (sat)
images (unsat)
102 meshes (sat)
meshes (unsat)
101

100
100 101 102 103 104 105 106 107 108
Runtime without Backjumping (ms)

Fig. 4. The effects of backjumping. Each point is one benchmark instance; points below
the diagonal line indicate a speedup.
308 C. McCreesh and P. Prosser

109
Recursive calls with matching-based all-different

108

LV (sat)
107
LV (unsat)
BVG / BVGm
106
M4D / M4Dr
SF (sat)
105
SF (unsat)
r
104
football
images (sat)
103
images (unsat)
meshes (sat)
102
meshes (unsat)

101

100
100 101 102 103 104 105 106 107 108 109
Recursive calls with counting-based all-different
9
10
Recursive calls with counting-based all-different

108

107

106

105

104

103

102

101

100
100 101 102 103 104 105 106 107 108 109
Recursive calls with value-deleting all-different

Fig. 5. Above, the improvement to the search space size which would be given by
Régin’s matching-based all-different propagator. Below, the improvement given by
using counting all-different rather than simple deletion. Each point is one benchmark
instance; the point style shows the benchmark family. Points below the diagonal line
indicate a reduction in the search space size.
 A Parallel, Backjumping Subgraph Isomorphism Algorithm 309

(still) speed up the algorithm by several orders of magnitude on hard, structured

problems”. Real-world graphs often have unexpected structural properties which
are not present in random instances [22,34], so we consider backjumping to be
worthwhile.

4.4 Comparing All-Different Propagators

We now justify our use of the counting all-different propagator. In the top half of
Fig. 5 we show the benefits to the size of the search space that would be gained
if we used Régin’s algorithm at every step instead of our counting propagator
(cutting search off after 109 search nodes). We see that for most graph families,
there would be little to no benefit even if there was no additional performance
cost. Only in a small portion of the LV graphs do we see a gain (and in one case,
due to dynamic variable ordering, there is a penalty).
Thus, either our counting propagator is nearly always as good as matching,
or neither propagator does very much in this domain. In the bottom half of
Fig. 5 we show the benefits to the size of the search space that are gained from
using counting, rather than simply deleting a value from every other domain on
assignment. The large number of points below the diagonal line confirm that
going beyond simple value deletion for all-different propagation is worthwhile.

5 Conclusion

Going against conventional wisdom, we saw that replacing strong inference with
cheaper surrogates could pay off, and that backjumping could be implemented
cheaply enough to be beneficial. We also saw parallelism give a substantial bene-
fit. This was true even for relatively low runtimes, due to us exploiting parallelism
for pre-processing as well as for search. Parallel backjumping has only been given
limited attention [8,10,17]. However, a simple approach has worked reasonably
well here (in contrast to stronger clause-learning systems, where successes in
parallelism appear to be rare).
There is also plenty of scope for extensions of and improvement to our algo-
rithm. We have yet to deeply investigate the possibility of constructing domain-
or instance-specific supplemental graphs. Nor did we discuss directed graphs
or induced isomorphisms: supplemental graphs can be taken further for these
variations of the problem. In particular, composing transformations for induced
isomorphisms would allow us to reason about “paths of non-edges”, which may
be very helpful. Finally, we have yet to consider exploiting the symmetries and
dominance relations which we know are present in many graph instances.

Acknowledgments. The authors wish to thank Christine Solnon for discussions, pro-
viding the graph instances and the LAD implementation, Christophe Lecoutre for dis-
cussion and the SND implementation, and Lars Kotthoff and Alice Miller for their
comments.
310 C. McCreesh and P. Prosser

References
1. Audemard, G., Lecoutre, C., Samy-Modeliar, M., Goncalves, G., Porumbel, D.:
Scoring-based neighborhood dominance for the subgraph isomorphism problem. In:
O’Sullivan, B. (ed.) CP 2014. LNCS, vol. 8656, pp. 125–141. Springer, Heidelberg
(2014). http://dx.doi.org/10.1007/978-3-319-10428-7 12
2. Bessière, C., Régin, J.: MAC and combined heuristics: two reasons to forsake
FC (and CBJ?) on hard problems. In: Proceedings of the Second International
Conference on Principles and Practice of Constraint Programming, Cambridge,
Massachusetts, USA, August 19–22, 1996, pp. 61–75 (1996). http://dx.doi.org/10.
1007/3-540-61551-2 66
3. Bonnici, V., Giugno, R., Pulvirenti, A., Shasha, D., Ferro, A.: A subgraph iso-
morphism algorithm and its application to biochemical data. BMC Bioinformatics
14(Suppl 7), S13 (2013). http://www.biomedcentral.com/1471-2105/14/S7/S13
4. de Bruin, A., Kindervater, G.A.P., Trienekens, H.W.J.M.: Asynchronous par-
allel branch and bound and anomalies. In: Ferreira, A., Rolim, J.D.P. (eds.)
IRREGULAR 1995. LNCS, vol. 980, pp. 363–377. Springer, Heidelberg (1995).
http://dx.doi.org/10.1007/3-540-60321-2 29
5. Chen, X., van Beek, P.: Conflict-directed backjumping revisited. J. Artif. Intell.
Res. (JAIR) 14, 53–81 (2001). http://dx.doi.org/10.1613/jair.788
6. Chu, G., Schulte, C., Stuckey, P.J.: Confidence-based work stealing in parallel con-
straint programming. In: Gent, I.P. (ed.) CP 2009. LNCS, vol. 5732, pp. 226–241.
Springer, Heidelberg (2009). http://dx.doi.org/10.1007/978-3-642-04244-7 20
7. Coffman, T., Greenblatt, S., Marcus, S.: Graph-based technologies for intel-
ligence analysis. Commun. ACM 47(3), 45–47 (2004). http://doi.acm.org/
10.1145/971617.971643
8. Conrad, J., Mathew, J.: A backjumping search algorithm for a distributed memory
multicomputer. In: International Conference on Parallel Processing, ICPP 1994,
vol. 3, pp. 243–246, August 1994
9. Conte, D., Foggia, P., Sansone, C., Vento, M.: Thirty years of graph match-
ing in pattern recognition. International Journal of Pattern Recognition and
Artificial Intelligence 18(03), 265–298 (2004). http://www.worldscientific.com/
doi/abs/10.1142/S0218001404003228
10. Cope, M., Gent, I.P., Hammond, K.: Parallel heuristic search in Haskell. In:
Selected Papers from the 2nd Scottish Functional Programming Workshop
(SFP00), University of St Andrews, Scotland, July 26–28, 2000, pp. 65–76 (2000)
11. Cordella, L.P., Foggia, P., Sansone, C., Vento, M.: A (sub)graph isomorphism algo-
rithm for matching large graphs. IEEE Trans. Pattern Anal. Mach. Intell. 26(10),
1367–1372 (2004). http://doi.ieeecomputersociety.org/10.1109/TPAMI.2004.75
12. Damiand, G., Solnon, C., de la Higuera, C., Janodet, J.C., Samuel, É.:
Polynomial algorithms for subisomorphism of nD open combinatorial maps.
Computer Vision and Image Understanding 115(7), 996–1010 (2011).
http://www.sciencedirect.com/science/article/pii/S1077314211000816, special
issue on Graph-Based Representations in Computer Vision
13. Geelen, P.A.: Dual viewpoint heuristics for binary constraint satisfaction problems.
In: ECAI, pp. 31–35 (1992)
14. Gent, I.P., Miguel, I., Moore, N.C.A.: Lazy explanations for constraint propagators.
In: Carro, M., Peña, R. (eds.) PADL 2010. LNCS, vol. 5937, pp. 217–233. Springer,
Heidelberg (2010). http://dx.doi.org/10.1007/978-3-642-11503-5 19
 A Parallel, Backjumping Subgraph Isomorphism Algorithm 311

15. Gent, I.P., Miguel, I., Nightingale, P.: Generalised arc consistency for the alldif-
ferent constraint: An empirical survey. Artificial Intelligence 172(18), 1973–2000
(2008). http://www.sciencedirect.com/science/article/pii/S0004370208001410,
special Review Issue
16. Giugno, R., Bonnici, V., Bombieri, N., Pulvirenti, A., Ferro, A., Shasha, D.: Grapes:
A software for parallel searching on biological graphs targeting multi-core architec-
tures. PLoS ONE 8(10), e76911 (2013) http://dx.doi.org/10.1371%2Fjournal.pone.
0076911
17. Habbas, Z., Herrmann, F., Merel, P.P., Singer, D.: Load balancing strategies for
parallel forward search algorithm with conflict based backjumping. In: Proceed-
ings of the 1997 International Conference on Parallel and Distributed Systems,
pp. 376–381, December 1997
18. Harvey, W.D., Ginsberg, M.L.: Limited discrepancy search. In: IJCAI (1),
pp. 607–615. Morgan Kaufmann, San Francisco (1995)
19. Kessel, P.V., Quimper, C.: Filtering algorithms based on the word-RAM model. In:
Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence, July
22–26, 2012, Toronto, Ontario, Canada. (2012). http://www.aaai.org/ocs/index.
php/AAAI/AAAI12/paper/view/5135
20. Larrosa, J., Valiente, G.: Constraint satisfaction algorithms for graph pattern
matching. Mathematical Structures in Computer Science 12(4), 403–422 (2002).
http://dx.doi.org/10.1017/S0960129501003577
21. Lobachev, O.: Parallel computation skeletons with premature termination property.
In: Schrijvers, T., Thiemann, P. (eds.) FLOPS 2012. LNCS, vol. 7294, pp. 197–212.
Springer, Heidelberg (2012). http://dx.doi.org/10.1007/978-3-642-29822-6 17
22. MacIntyre, E., Prosser, P., Smith, B.M., Walsh, T.: Random constraint satis-
faction: theory meets practice. In: Maher, M.J., Puget, J.-F. (eds.) CP 1998.
LNCS, vol. 1520, p. 325. Springer, Heidelberg (1998). http://dx.doi.org/10.1007/
3-540-49481-2 24
23. Malitsky, Y.: Instance-Specific Algorithm Configuration. Springer (2014). http://
dx.doi.org/10.1007/978-3-319-11230-5
24. McCreesh, C., Prosser, P.: The shape of the search tree for the maximum clique
problem and the implications for parallel branch and bound. ACM Trans. Parallel
Comput. 2(1), 8:1–8:27 (2015). http://doi.acm.org/10.1145/2742359
25. McGregor, J.J.: Relational consistency algorithms and their application in
finding subgraph and graph isomorphisms. Inf. Sci. 19(3), 229–250 (1979).
http://dx.doi.org/10.1016/0020-0255(79)90023–9
26. Prosser, P.: Domain filtering can degrade intelligent backtracking search. In: Pro-
ceedings of the 13th International Joint Conference on Artifical Intelligence, IJCAI
1993 ,vol. 1, pp. 262–267. Morgan Kaufmann Publishers Inc., San Francisco (1993).
http://dl.acm.org/citation.cfm?id=1624025.1624062
27. Prosser, P.: Hybrid algorithms for the constraint satisfaction problem. Com-
putational Intelligence 9, 268–299 (1993). http://dx.doi.org/10.1111/j.1467-8640.
1993.tb00310.x
28. Puget, J.: A fast algorithm for the bound consistency of alldiff constraints. In: Pro-
ceedings of the Fifteenth National Conference on Artificial Intelligence and Tenth
Innovative Applications of Artificial Intelligence Conference, AAAI 1998, IAAI
1998, July 26–30, 1998, Madison, Wisconsin, USA, pp. 359–366 (1998). http://
www.aaai.org/Library/AAAI/1998/aaai98-051.php
312 C. McCreesh and P. Prosser

29. Quimper, C.-G., Walsh, T.: The all different and global cardinality constraints
on set, multiset and tuple variables. In: Hnich, B., Carlsson, M., Fages, F., Rossi,
F. (eds.) CSCLP 2005. LNCS (LNAI), vol. 3978, pp. 1–13. Springer, Heidelberg
(2006). http://dx.doi.org/10.1007/11754602 1
30. Régin, J.: A filtering algorithm for constraints of difference in CSPs. In: Proceed-
ings of the 12th National Conference on Artificial Intelligence, Seattle, WA, USA,
July 31– August 4, 1994, vol. 1, pp. 362–367 (1994). http://www.aaai.org/Library/
AAAI/1994/aaai94-055.php
31. Régin, J.C.: Développement d’outils algorithmiques pour l’Intelligence Artificielle.
Application à la chimie organique. Ph.D. thesis, Université Montpellier 2 (1995)
32. San Segundo, P., Rodriguez-Losada, D., Galan, R., Matia, F., Jimenez, A.: Exploit-
ing CPU bit parallel operations to improve efficiency in search. In: 19th IEEE
International Conference on Tools with Artificial Intelligence, ICTAI 2007, vol. 1,
pp. 53–59, October 2007
33. Sevegnani, M., Calder, M.: Bigraphs with sharing. Theoretical Computer
Science 577, 43–73 (2015). http://www.sciencedirect.com/science/article/pii/
S0304397515001085
34. Slater, N., Itzchack, R., Louzoun, Y.: Mid size cliques are more common
in real world networks than triangles. Network Science 2, 387–402 (2014).
http://journals.cambridge.org/article S2050124214000228
35. Smith, B.M., Grant, S.A.: Modelling exceptionally hard constraint satisfaction
problems. In: Smolka, G. (ed.) CP 1997. LNCS, vol. 1330, pp. 182–195. Springer,
Heidelberg (1997). http://dx.doi.org/10.1007/BFb0017439
36. Solnon, C.: Alldifferent-based filtering for subgraph isomorphism. Artif. Intell.
174(12–13), 850–864 (2010). http://dx.doi.org/10.1016/j.artint.2010.05.002
37. Solnon, C., Damiand, G., de la Higuera, C., Janodet, J.C.: On the com-
plexity of submap isomorphism and maximum common submap problems.
Pattern Recognition 48(2), 302–316 (2015). http://www.sciencedirect.com/
science/article/pii/S0031320314002192
38. Trienekens, H.W.: Parallel Branch and Bound Algorithms. Ph.D. thesis, Erasmus
University Rotterdam (1990)
39. Ullmann, J.R.: An algorithm for subgraph isomorphism. Journal of the ACM
(JACM) 23(1), 31–42 (1976)
40. Ullmann, J.R.: Bit-vector algorithms for binary constraint satisfaction and
subgraph isomorphism. J. Exp. Algorithmics 15, 1.6:1.1–1.6:1.64 (2011).
http://doi.acm.org/10.1145/1671970.1921702
41. Zampelli, S., Deville, Y., Solnon, C.: Solving subgraph isomorphism prob-
lems with constraint programming. Constraints 15(3), 327–353 (2010).
http://dx.doi.org/10.1007/s10601-009-9074-3
 Automated Auxiliary Variable Elimination
Through On-the-Fly Propagator Generation

Jean-Noël Monette(B) , Pierre Flener, and Justin Pearson

Department of Information Technology, Uppsala University, Uppsala, Sweden

{jean-noel.monette,pierre.flener,justin.pearson}@it.uu.se

Abstract. Model flattening often introduces many auxiliary variables.

We provide a way to eliminate some of the auxiliary variables occur-
ring in exactly two constraints by replacing those two constraints by a
new equivalent constraint for which a propagator is automatically gen-
erated on the fly. Experiments show that, despite the overhead of the
preprocessing and of using machine-generated propagators, eliminating
auxiliary variables often reduces the solving time.

1 Introduction
Constraint-based modelling languages such as Essence [6] and MiniZinc [12]
enable the modelling of problems at a higher level of abstraction than is sup-
ported by most constraint solvers. The transformation from a high-level model
to a low-level model supported by a solver is often called flattening and has
been the subject of intense research in order to produce good low-level, or flat-
tened, models (see, e.g., [14]). By decomposing complex expressions into a form
accepted by a solver, flattening often introduces many auxiliary variables into
the flattened model. Those auxiliary variables and the propagators of the con-
straints in which they appear may have a large negative impact on the efficiency
of solving (as we will show in Section 5).
In this paper, we propose a fully automated way to address this problem
by removing from the flattened model some auxiliary variables that appear in
exactly two constraints and replacing those two constraints by a new equivalent
constraint for which a propagator is generated. Given a flattened model, our
approach is fully automated and online. It can be summarised as follows:
1. Identify frequent patterns in the flattened model consisting of two constraints
sharing an auxiliary variable.
2. For a pattern, define a new constraint predicate that involves all variables
appearing in the pattern except for the shared auxiliary variable.
3. For a pattern, replace all its occurrences in the flattened model by instanti-
ating the new constraint predicate.
4. For each new predicate, generate a propagator description in the indexical
language of [11] and compile it for the targeted constraint solver.
5. Solve the modified flattened model using the constraint solver extended with
the new indexical-based propagators.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 313–329, 2015.
DOI: 10.1007/978-3-319-23219-5 22
314 J.-N. Monette et al.

Our experiments in Section 5 show that our approach is useful for instances that
are hard to solve, reducing the average time by 9% for those taking more than
one minute and sometimes more than doubling the search speed.
The rest of the paper assumes that the low-level language is FlatZinc [1] and
that the constraint solver is Gecode [7], as our current implementation is based
on FlatZinc and Gecode. However, the ideas presented here are applicable to
other low-level modelling languages and other constraint solvers.
The paper is organised as follows. We start with some preliminaries in
Section 2. Then we describe our approach in Section 3 and present a complete
example in Section 4. In Section 5, we show that this approach is effective in
practice. In Section 6, we discuss the merits and limitations of our approach as
well as some alternatives. Finally, we conclude in Section 7.

2 Preliminaries
A constraint-based model is a formal way to describe a combinatorial problem.
It is composed of a set of variables, each with a finite domain in which it must
take its value, a set of constraints, and an objective. A solution is an assignment
to all variables so that all constraints are satisfied. The objective is either to find
a solution, or to find a solution in which a given variable is minimised, or to find
a solution in which a given variable is maximised.
A constraint predicate, or simply predicate, is defined by a name and a sig-
nature that lists the formal arguments of the predicate. The arguments can be
constants, variables, or arrays thereof. For simplicity, we identify a predicate
with its name and we often do not give the types of its arguments. We specify
the semantics of a predicate P by a logic formula involving the arguments of
P and constants. For example, one could write PLUS(X, Y, Z)  X = Y + Z,
for the predicate PLUS with formal arguments X, Y , and Z. A predicate whose
signature has arrays of variables is said to be of unfixed arity, or n-ary. Unfixed-
arity predicates are usually referred to as global constraints [2] in constraint
programming. Each constraint in a constraint-based model is the instantiation
of a constraint predicate to some actual arguments. We denote formal arguments
in upper case and actual arguments inside a model in lower case.

2.1 MiniZinc and FlatZinc

MiniZinc is a solver-independent constraint-based modelling language for com-

binatorial problems. Before being presented to a solver, a MiniZinc model is
transformed into a FlatZinc model by a process called flattening. The Mini-
Zinc flattener inlines function and predicate calls, decomposes expressions, and
unrolls loops to provide the solver with a model that is a conjunction of primi-
tive constraints (i.e., constraints that are recognised by the targeted solver) over
simple arguments (i.e., only variables, constants, or arrays thereof). To do so, the
flattener may introduce auxiliary variables, which are annotated in the FlatZinc
model with is introduced var.
 Automated Auxiliary Variable Elimination 315

1 var int: w; 1 var int: w;

2 var int: y; 2 var int: y;
3 var int: z; 3 var int: z;
4 4 var int: x1 ::var_is_introduced;
5 5 var bool: x2 ::var_is_introduced;
6 6 var bool: x3 ::var_is_introduced;
7 constraint w = max(y,z) ∨ 1 ≤ z; 7 constraint x1 = max(y,z);
8 8 constraint x2 ≡ w = x1 ;
9 9 constraint x3 ≡ 1 ≤ z;
10 10 constraint x2 ∨ x3 ;
11 solve satisfy; 11 solve satisfy;

(a) MiniZinc model (b) FlatZinc model

Fig. 1. MiniZinc model (1a) and FlatZinc model resulting from its flattening (1b)

For example, the MiniZinc model of Figure 1a is flattened into the FlatZinc
model of Figure 1b, modulo editorial changes. The single constraint expression
of the MiniZinc model (line 7 in Figure 1a) is flattened by introducing three aux-
iliary variables (lines 4–6 in Figure 1b) and posting four primitive constraints:
the constraints in lines 7–9 functionally define the auxiliary variables represent-
ing parts of the original constraint expression, and the disjunction of line 10
corresponds to the top-level expression.
For 75 of the 488 benchmark instances we used (see Section 5 for details),
no auxiliary variables are introduced by flattening. For 264 instances, flattening
multiplies the number of variables in the model by more than 5. For 7 instances,
flattening even multiplies the number of variables by more than 100.

2.2 Patterns, Occurrences, and Extensions

A pattern is here a new constraint predicate with signature Y1 , ..., Yn , K1 , ..., Km ,
where the Yi are n ≥ 2 variable identifiers and the Ki are m ≥ 0 constant
identifiers. Its semantics is specified by the conjunction of two existing predicates
PA and PB applied to arguments A1 , . . . , Ap and B1 , . . . , Bq , for some p, q ≥ 2,
such that each Ai and each Bi is either one of Y1 , . . . , Yn , K1 , . . . , Km , or a
unique local and existentially quantified variable X, or a constant, and such
that X appears in both A1 , . . . , Ap and B1 , . . . , Bq . We reserve the identifier X
for the local and existentially quantified variable. Hence, for simplicity, we will
often omit writing the existential quantification of X in predicate semantics.
An occurrence of a pattern in a model is made of two constraints C1 and
C2 sharing a variable x that appears only in C1 and C2 such that PA can be
instantiated to C1 and PB to C2 with X being instantiated to x.
For example, the pattern P (Y1 , Y2 , Y3 , Y4 )  ∃X : X = max(Y1 , Y2 ) ∧ Y3 ≡
Y4 = X occurs in the model of Figure 1b, with C1 being x1 = max(y, z) in line 7,
C2 being x2 ≡ w = x1 in line 8, and X being x1 . Hence this occurrence of P is
equivalent to the constraint P (y, z, x2 , w).
316 J.-N. Monette et al.

1 def PLUS(vint X, vint Y, vint Z){

2 propagator{
3 X in (min(Y)+min(Z)) .. (max(Y)+max(Z)) ;
4 Y in (min(X)-max(Z)) .. (max(X)-min(Z)) ;
5 Z in (min(X)-max(Y)) .. (max(X)-min(Y)) ;
6 }
7 checker{ X == Y + Z }
8 }

Fig. 2. Indexical propagator description for X = Y + Z

A pattern P1 is said to be a specialisation of another pattern P2 if one can

define P1 in terms of P2 by properly instantiating some of the arguments of P2 .
For example, consider the patterns P1 (Y1 , Y2 )  (X ∨ Y1 ) ∧ X ≡ 1 ≤ Y2
and P2 (Y1 , Y2 , K1 )  (X ∨ Y1 ) ∧ X ≡ K1 ≤ Y2 , both of which occur on lines 9
and 10 of Figure 1b with X being x3 . Then P1 is a specialisation of P2 because
P1 (Y1 , Y2 ) ⇔ P2 (Y1 , Y2 , 1).

2.3 Indexicals

Indexicals [20] are used to describe concisely propagation in propagator-based

constraint solvers. The core indexical expression takes the form ‘X in σ’ and
restricts the domain of variable X to be a subset of the set-valued expression σ,
which depends on the domains of other variables. Figure 2 presents an index-
ical propagator description in the language of [11] for the constraint predicate
PLUS(X, Y, Z)  X = Y + Z; it involves three indexical expressions, one for
each variable (lines 3–5), the min and max operators referring to the domain of
the argument variable. Figure 2 also contains a checker (line 7), which is used
to test whether the constraint holds on ground instances and can be seen as a
specification of the predicate semantics.
Our indexical compiler [11] takes an indexical propagator description like
the one in Figure 2 and compiles it into an executable propagator for a number
of constraint solvers, including Gecode. The compiled propagator is said to be
indexical-based. The experimental results in [11] show that an indexical-based
propagator uses between 1.2 and 2.7 times the time spent by a hand-crafted
propagator on a selection of n-ary constraint predicates.

3 Our Approach
Given a FlatZinc model, our approach, whose implementation is referred to
as var-elim-idxs, adds a preprocessing step before solving the model. This
preprocessing is summarised in Algorithm 1. We do not describe lines 1, 5, and 6,
as they involve purely mechanical and well-understood aspects. The core of the
algorithm iteratively identifies frequent patterns (line 2), replaces them in the
 Automated Auxiliary Variable Elimination 317

Algorithm 1. Main preprocessing algorithm of var-elim-idxs

Input: a flattened model
Output: updated flattened model and extended solver
1: Parse the flattened model
2: while there is a most frequent pattern P in the model (Section 3.1) do
3: Replace each occurrence of P in the model by instantiating P (Section 3.2)
4: Generate and compile an indexical propagator description for P (Section 3.3)
5: Output the updated model
6: Link the compiled indexical-based propagators with the solver

model (line 3), and generates propagators (line 4). Sections 3.1 to 3.3 describe
this core.
The loop of lines 2 to 4 ends when no pattern occurs often enough, which
is when the number of occurrences of a most frequent pattern (if any) is less
than 10 or less than 5% of the number of variables in the model: under those
thresholds, the effort of preprocessing is not worth it. To save further efforts, if
the number of auxiliary variables is less than 10 or less than 5% of the number
of variables in the model (this criterion can be checked very fast with Unix
utilities such as grep), then the preprocessing is not performed at all. The two
thresholds, 10 and 5%, have been set arbitrarily after some initial experiments.

3.1 Identification of Frequent Patterns

We now show how we identify patterns that occur frequently in a given model.
First, we collect all auxiliary variables that appear in exactly two constraints.
Indeed, to be in an occurrence of a pattern, a shared variable must occur only in
those two constraints. Most auxiliary variables appear in exactly two constraints:
this is the case of 84% of the auxiliary variables appearing in the 488 benchmark
instances we used. For example, this is also the case for all auxiliary variables
in Figure 1b. However, due to common subexpression elimination [13,14], some
auxiliary variables appear in more than two constraints. We do not consider such
variables for elimination as common subexpression elimination usually increases
the amount of filtering and eliminating those variables might cancel this benefit.
Then, for each collected variable x, we create a pattern as follows. Let
PA (a1 , . . . , ap ) and PB (b1 , . . . , bq ) be the two constraints in which x appears. The
pattern is such that the two predicates are PA (A1 , . . . , Ap ) and PB (B1 , . . . , Bq ),
where each Ai is defined by the following rules, and similarly for each Bi :
– If ai is x, then Ai is X.
– If ai is a variable other than x, then Ai is Yk for the next unused k.
– If ai ∈ {−1, 0, 1, true, false}, then Ai is ai .
– If ai is a constant not in {−1, 0, 1, true, false}, then Ai is Kk for the next
unused k.
– If ai is an array, then Ai is an array of the same length where each element
is defined by applying the previous rules to the element of ai at the same
position.
318 J.-N. Monette et al.

The purpose of the third rule is to allow a better handling of some special
constants, which may simplify the generated propagators. For example, linear
(in)equalities can be propagated more efficiently with unit coefficients.
In general, there might be other shared variables between PA (a1 , . . . , ap ) and
PB (b1 , . . . , bq ) besides x but, to keep things simple, we consider them separately,
i.e., a new argument Yk with an unused k is created for each occurrence of another
shared variable. Ignoring other shared variables does not affect the correctness
of our approach, but the generated propagators may achieve less filtering than
possible otherwise. We will revisit this issue in Section 6.
In order to avoid the creation of symmetric patterns, we sort the elements of
an array when their order is not relevant: we do this currently only for the n-ary
Boolean disjunction and conjunction constraints.
For example, both (x∨y)∧(x ≡ 1 ≤ z) and (w∨u)∧(u ≡ 1 ≤ v) are considered
occurrences of P (Y1 , Y2 )  (X ∨ Y1 ) ∧ (X ≡ 1 ≤ Y2 ), although, syntactically,
the shared variables (x and u, respectively) occur in different positions in their
respective disjunctions.
Finally, we count the occurrences of each created pattern. In doing so, we
ignore the following patterns, expressed here in terms of criteria that are specific
to FlatZinc and Gecode:
– Patterns involving an n-ary predicate with at least five variables among
its arguments: currently, our approach considers only fixed-arity patterns.
Hence n-ary constraints with different numbers of variables are considered
as fixed-arity constraints with different arities. We only want to keep those
of small arities for efficiency reasons. The threshold is set to 5 because no
fixed-arity predicate in FlatZinc has more than four variable arguments.
– Patterns involving a predicate for which an indexical-based propagator is
expected to perform poorly with respect to a hand-written propagator: this
includes all the global constraint predicates from the MiniZinc standard
library, as well as the element, absolute value, division, modulo, and multi-
plication predicates of FlatZinc.
– Patterns involving a bool2int predicate in which the Boolean variable is the
shared variable, and another predicate that is not a pattern itself: this case
is ignored as, in some cases, the Gecode-FlatZinc parser takes advantage of
bool2int constraints and we do not want to lose that optimisation.
The two first criteria partially but not completely overlap: for example, the
constraint all different([X, Y, Z, W ]) is not ruled out by the first criterion as
it has fewer than five variables, but it is ruled out by the second one; conversely,
int lin eq(coeffs, [X, Y, Z, W, U ], k) is ruled out by the first criterion but not
by the second one.
The result of pattern identification is a pattern that occurs most in the model.

3.2 Pattern Instantiation

Having identified a most frequent pattern P , we replace each occurrence of P by
an instantiation of P . More precisely, for each auxiliary variable x that appears
 Automated Auxiliary Variable Elimination 319

in exactly two constraints, if the pattern created for x in Section 3.1 is P or a

specialisation of P (detected by simple pattern matching), then we replace in
the model the declaration of x and the two constraints in which x appears by
an instantiation of P , obtained by replacing each formal argument of P by the
actual arguments of the two constraints in which x appears. To achieve this,
each argument Ai of PA in the semantics of P is considered together with the
argument ai of the instantiation of PA in which x appears, in order to apply the
following rules, and similarly for each Bi :

– If Ai is X, then variable ai , which is x, is not in the instantiation of P .

– If Ai is Yk , then Yk is instantiated to the variable ai .
– If Ai ∈ {−1, 0, 1, true, false}, then ai is not in the instantiation of P .
– If Ai is Kk , then Kk is instantiated to the constant ai .
– If Ai is an array, then the previous rules are applied to each element of Ai .

For example, consider the pattern P1 (Z1 , Z2 )  (X ≡ 1 ≤ Z1 ) ∧ (Z2 ∨ X).

Variable x3 of Figure 1b appears in x3 ≡ 1 ≤ z (line 9) and x2 ∨ x3 (line 10).
Then P1 can be instantiated to P1 (z, x2 ) and this constraint replaces lines 6, 9,
and 10 in Figure 1b.
Due to the sorting of the elements for the n-ary Boolean conjunction and
disjunction predicates, some occurrences of a pattern may disappear before their
instantiation. Consider the MiniZinc-level expression z1 > 0 ∨ z2 > 0. Flattening
introduces two auxiliary Boolean variables, say b1 and b2 , together with the three
constraints b1 ≡ z1 > 0, b2 ≡ z2 > 0, and b1 ∨b2 . Hence there are two occurrences
of the pattern P (Y1 , Y2 )  (X ≡ Y1 > 0) ∧ (X ∨ Y2 ) but only one of them will
actually be replaced, say the one in which X is b1 . After replacing it, the model
contains the two constraints P (z1 , b2 ) and b2 ≡ z2 > 0, changing the pattern to
which b2 belongs. This new pattern might be replaced in a later iteration of the
loop in Algorithm 1.

3.3 Indexical Propagator Description Generation and Compilation

The generation of a propagator for a pattern uses our indexical compiler [11]
and performs the following steps:

1. Translation of the pattern into a checker in the indexical syntax.

2. Elimination of the shared variable from the checker.
3. Generation of an indexical propagator description, based on the checker.
4. Compilation [11] of the indexical propagator description into an actual prop-
agator, written in C++ in the case of Gecode.

Step 1 only involves a change of syntax, and Step 4 has already been described
in [11]. Hence, we will focus here on Steps 2 and 3.
Let X be the variable to eliminate in Step 2. Here, variable elimination can
take two forms. First, if X is constrained to be equal to some expression φ in one
of the two conjuncts, i.e., if the checker can be written as P (. . . , X, . . . ) ∧ X = φ
for some predicate P , then all occurrences of X in the checker are replaced by
320 J.-N. Monette et al.

φ, i.e., the checker becomes P (. . . , φ, . . . ). As Boolean variables are considered

a special case of integer variables in the indexical compiler, this rule also covers
the case of X being a Boolean variable constrained to be equivalent to a Boolean
expression. Second, if both conjuncts are disjunctions, one involving a Boolean
variable X and the other ¬X, i.e., if the checker can be written as (δ1 ∨ X) ∧
(δ2 ∨ ¬X) for some Boolean expressions δ1 and δ2 , then applying the resolution
rule yields a single disjunction without X, i.e., the checker becomes δ1 ∨ δ2 .
The generation in Step 3 of an indexical propagator description from a
checker works by syntactic transformation: rewriting rules are recursively applied
to the checker expression and its subexpressions in order to create progressively
a collection of indexical expressions. The whole transformation has more than
250 rules. We limit our presentation to the most representative ones.
The rule for a conjunction γ1 ∧ γ2 concatenates the rewriting of γ1 and γ2 .
The rule for a disjunction δ1 ∨ δ2 is such that δ2 is propagated only once δ1 can
be shown to be unsatisfiable, and conversely. The rule for an equality φ1 = φ2
creates expressions that force the value of φ1 to belong to the possible values
for φ2 , and conversely. If φ1 is a variable, say Y , then the rule creates the
indexical Y in UB(φ2 ), where UB(φ2 ) is possibly an over-approximation of the
set containing all possible values of the expression φ2 , based on the domains of
the variables appearing in φ2 . If φ1 is not a variable but a compound expression,
then it must be recursively decomposed. Consider the case of φ1 being Y1 + Y2 :
two indexical expressions are created, namely Y1 in UB(φ2 − Y2 ) and Y2 in
UB(φ2 − Y1 ). The other rules cover all the other expressions that can appear in a
checker. The function UB(.) is also defined by rules. As an example, UB(φ1 −φ2 )
is rewritten as UB(φ1 )  UB(φ2 ), where  is pointwise integer set substraction:
S  T = {s − t | s ∈ S ∧ t ∈ T }.
This generation mechanism packaged with our indexical compiler has been
used in [9,17] to prototype propagators rapidly for newly identified constraints,
but it has never been described before. It is very similar to the compilation of
projection constraints in Nicolog [18], but generalised to n-ary constraints.

4 Example: Ship Schedule

To illustrate our approach, we now consider the ship-schedule.cp.mzn model

from the MiniZinc Challenge 2012. The objective of the problem is to find which
boats are sailing and when, in order to satisfy several port constraints, e.g.,
tides, tugboat availability, and berth availability, as well as to maximise the
total weight that can be transported. The FlatZinc model produced by flatten-
ing the MiniZinc model with the 7ShipsMixed.dzn data file, which represents an
instance with 7 boats and 74 time slots, contains 7,187 constraints and 5,978 vari-
ables, among which 4,848 are auxiliary, i.e., 81% of the variables.
When given this flattened model, var-elim-idxs iteratively identifies the
patterns reported in Table 1. Note that pattern P0 is used in pattern P1 . The loop
of Algorithm 1 ends upon identifying pattern P4 , which occurs less often than
5% of the number of variables and is not instantiated. In total, 4 new constraint
 Automated Auxiliary Variable Elimination 321

Table 1. Patterns found in ship-schedule.cp.mzn with 7ShipsMixed.dzn

Predicate Definition Frequency
P0 (Y1 , Y2 )  (X ∨ Y1 ) ∧ X ≡ (Y2 = 0) 892
P1 (Y1 , Y2 , Y3 , Y4 )  P0 (X, Y4 ) ∧ X ≡ (Y1 ∧ Y2 ∧ Y3 ) 612
P2 (Y1 , Y2 , Y3 , Y4 , K1 )  (X ∨ ¬(Y1 ∧ Y2 ∧ Y3 )) ∧ X ≡ (Y4 = K1 ) 612
P3 (Y1 , Y2 , Y3 , K1 )  (X ∨ ¬(Y1 ∧ Y2 )) ∧ X ≡ (Y3 = K1 ) 276
P4 (Y1 , Y2 , Y3 , Y4 )  (X ∨ Y1 ) ∧ X ≡ (Y2 ∧ Y3 ∧ Y4 ) 146

1 def P_1(vint Y_1::Bool, vint Y_2::Bool, vint Y_3, vint Y_4){

2 checker{
3 0==Y_4 or (1 <= Y_1 and 1 <= Y_2 and 1 <= Y_3)
4 }
5 propagator(gen)::DR{
6 once(not 0 memberof dom(Y_4)){
7 Y_1 in 1 .. sup;
8 Y_2 in 1 .. sup;
9 Y_3 in 1 .. sup;
10 }
11 (max(Y_1) < 1 or max(Y_2) < 1 or max(Y_3) < 1) -> Y_4 in {0};
12 }
13 }

Fig. 3. Indexical propagator description generated for pattern P1 of Table 1 from

ship-schedule.cp.mzn with 7ShipsMixed.dzn. Note that the indexical compiler treats
Boolean variables as a special case of integer variables with domain {0, 1}, where 1
represents true. For example, 1 <= Y 1 means Y 1 is true.

predicates (P0 to P3 ) are introduced in the model, eliminating 2,392 variables,

i.e., 40% of the variables. Among those, 222 variables are eliminated thanks to
pattern specialisation, i.e., because the pattern created for such a variable is not
exactly one of the patterns shown in Table 1 but a specialisation of one of those,
for example a specialisation of P2 with K1 = 1.
The final model contains 3,586 variables, of which 2,456 are auxiliary, and
4,795 constraints, of which 1,780 use the generated propagators. Many auxil-
iary variables are not eliminated because either they appear in more than two
constraints, or they appear in too infrequent patterns, such as P4 in Table 1,
or they appear in constraints that are not handled by var-elim-idxs, such as
n-ary constraints with n ≥ 5, the multiplication constraint, and the bool2int
constraint in the case explained in Section 3.1.
The new constraint predicates are handled by the indexical compiler to spec-
ify their checkers, generate indexical propagator descriptions, and compile the
latter into Gecode propagators. Figure 3 shows the generated indexical propaga-
tor description for pattern P1 of Table 1. It is interesting to note that the auxil-
iary variable X eliminated by instantiating P1 represents the truth of Y1 ∧Y2 ∧Y3 .
322 J.-N. Monette et al.

Hence the propagator for X ≡ (Y1 ∧Y2 ∧Y3 ) in the original flattened model needs
to watch both when all conjuncts become true to set X to true and when some
conjunct becomes false to set X to false. In contrast, the propagator of Figure 3
is only interested in the falsity of the conjuncts to restrict Y4 to 0. The generated
C++ code of the indexical-based propagators for Gecode is 579 lines long.
While the unmodified Gecode-FlatZinc interpreter solves the flattened model
in 111.5 seconds, var-elim-idxs solves it in 89.1 seconds, divided into 4.9 sec-
onds of preprocessing and 84.2 seconds of actual solving.

5 Experimental Evaluation

We implemented the preprocessing step of var-elim-idxs1 in Scala, using our

indexical compiler2 and the FlatZinc parser of the OscaR project3 as that parser
is written in Scala as well. For the experiments, we used Gecode 4.4.04 and the
MiniZinc flattener 2.0.1.5 The experiments were carried out inside a VirtualBox
virtual machine running Ubuntu 14.04 LTS 32-bit, with access to one core of a
64-bit Intel Core i7 at 3 GHz and 1 GB of RAM.
We tested var-elim-idxs on 489 of the 500 FlatZinc instances from the
MiniZinc Challenges 2010 to 2014. We excluded 11 instances that could not be
flattened, for lack of memory or because of a syntax error.6 We ran both Gecode
and var-elim-idxs once on each instance with a time-out of 10 minutes per
instance. Unless otherwise noted, ‘Gecode’ refers here to running an unmodified
version of Gecode and ‘var-elim-idxs’ refers to running our preprocessing step
followed by running the extended version of Gecode.
For 94 instances, the preprocessing is not run at all because the grep com-
mand detects that there are too few auxiliary variables: the behaviour of Gecode
and var-elim-idxs is identical as the time spent by grep is negligible. For 172
instances, the preprocessing is run but does not identify any frequent enough
pattern: the behaviour of Gecode and var-elim-idxs is identical except for the
extra time spent on preprocessing, discussed in the last paragraph of this section.
Table 2 reports the results, aggregated per MiniZinc model, on the 223
instances in which the preprocessing identifies frequent patterns. We refer to
those instances as modified instances. The bottom of the table presents the
aggregated results over all modified instances for which the total time of Gecode
is respectively more than 1 and 60 seconds. The node rate ratio for an instance
is computed as rv /rg , where rx = nx /tx with nx being the number of nodes of
the search tree visited before time-out, and tx being the total time in the column
ratio total and the search time in the column ratio search; the subscript x = ‘g’
1
https://bitbucket.org/jmonette/var-elim-idxs
2
https://bitbucket.org/jmonette/indexicals
3
http://www.oscarlib.org
4
http://www.gecode.org/
5
http://www.minizinc.org/2.0/
6
The sugiyama model could not be parsed by MiniZinc 2.0.1. It could be parsed by
MiniZinc 1.6, and the problem is corrected in the development version of MiniZinc.
 Automated Auxiliary Variable Elimination 323

Table 2. Results for the 223 modified instances, aggregated per MiniZinc model, with
the following columns: name of the model; number of instances; mean and standard
deviation of the percentage of auxiliary variables; mean and standard deviation of the
percentage of variables eliminated by var-elim-idxs (over all variables of the model);
geometric mean and geometric standard deviation of the node rate ratio including
preprocessing; as well as geometric mean and geometric standard deviation of the
node rate ratio excluding preprocessing. The models are ordered by decreasing ratio
excluding preprocessing (column ‘ratio search’).

name inst. % aux. % elim. ratio total ratio search

l2p 5 93 (1) 73 (4) 1.27 (1.92) 2.30 (1.12)
amaze3 5 92 (1) 10 (1) 0.71 (1.98) 1.71 (2.83)
league 11 94 (4) 30 (13) 1.37 (1.71) 1.45 (1.64)
openshop 5 96 (1) 96 (1) 0.92 (2.15) 1.35 (1.13)
ship-schedule 15 82 (1) 37 (4) 0.52 (3.07) 1.32 (1.07)
wwtpp-real 10 75 (1) 70 (2) 0.27(14.75) 1.32 (1.39)
radiation 10 64 (1) 32 (1) 1.15 (1.15) 1.22 (1.07)
wwtpp-random 5 75 (0) 62 (0) 0.46 (8.31) 1.21 (1.29)
javarouting 5 88 (1) 82 (1) 1.16 (1.03) 1.18 (1.02)
solbat 30 98 (0) 16 (0) 0.65 (2.56) 1.17 (1.34)
amaze 6 55 (1) 36 (1) 1.12 (1.05) 1.16 (1.01)
project-planning 6 66 (0) 31 (2) 1.12 (1.04) 1.13 (1.04)
open-stacks 5 82 (0) 43 (2) 0.63 (2.69) 1.08 (1.04)
traveling-tppv 5 83 (0) 28 (0) 1.04 (1.01) 1.05 (1.02)
fjsp 3 75 (8) 21 (3) 0.92 (1.23) 1.04 (1.08)
tpp 6 88 (1) 24 (2) 0.97 (1.05) 1.03 (1.01)
smelt 4 72 (3) 9 (2) 1.03 (1.08) 1.03 (1.08)
train 9 53 (1) 23 (2) 1.03 (1.06) 1.03 (1.06)
pattern-set-mining 1 66 (0) 29 (0) 1.02 (1.00) 1.03 (1.00)
mspsp 2 72 (2) 54 (5) 1.02 (1.02) 1.02 (1.02)
on-call-rostering 4 66 (6) 24 (4) 1.01 (1.01) 1.02 (1.02)
carpet-cutting 2 57 (0) 37 (0) 1.01 (1.02) 1.02 (1.02)
jp-encoding 5 92 (0) 20 (0) 1.00 (1.01) 1.01 (1.01)
cyclic-rcpsp 10 91 (3) 75 (5) 1.00 (1.02) 1.00 (1.02)
rcpsp-max 6 97 (1) 96 (1) 0.99 (1.02) 1.00 (1.02)
rcpsp 4 96 (3) 94 (4) 0.99 (1.02) 0.99 (1.02)
elitserien 5 71 (0) 24 (0) 0.94 (1.07) 0.99 (1.02)
liner-sf-repositioning 4 85 (0) 10 (0) 0.97 (1.02) 0.99 (1.02)
rectangle-packing 5 88 (0) 49 (1) 0.97 (1.04) 0.98 (1.05)
stochastic-fjsp 2 81 (0) 5 (0) 0.94 (1.00) 0.98 (1.03)
still-life-wastage 5 87 (1) 5 (0) 0.88 (1.14) 0.97 (1.01)
amaze2 6 93 (0) 28 (10) 0.86 (1.17) 0.87 (1.17)
fillomino 2 94 (0) 6 (0) 0.40 (2.27) 0.86 (1.07)
roster 5 82 (0) 63 (0) 0.78 (1.42) 0.79 (1.41)
mario 10 92 (0) 22 (1) 0.56 (1.87) 0.78 (1.33)
Total 223 83 (13) 38 (26) 0.81 (2.54) 1.12 (1.38)
Total (> 1 s.) 207 83 (14) 38 (26) 1.00 (1.45) 1.12 (1.31)
Total (> 60 s.) 174 83 (14) 40 (26) 1.09 (1.27) 1.11 (1.27)
324 J.-N. Monette et al.

Table 3. Results for the 172 unmodified instances, with the columns of Table 2
inst. % aux. % elim. ratio total ratio search
Total 172 80 (20) 0 (0) 0.82 (1.80) 1.00 (1.08)
Total (> 1 s.) 159 79 (20) 0 (0) 0.95 (1.13) 1.00 (1.01)
Total (> 60 s.) 110 76 (20) 0 (0) 1.00 (1.02) 1.00 (1.01)

refers to Gecode and ‘v’ to var-elim-idxs. We use node rates in order to have
a meaningful measure for both the instances that are solved before time-out and
those that are not. This assumes that the explored search trees are the same,
which is discussed in Section 6. A ratio larger than 1 means that var-elim-idxs
is faster than Gecode. We do not consider ratios between 0.97 and 1.03 to rep-
resent a significant change.
The percentage of auxiliary variables is generally very high, with an average
of 83%, but on average only 70% of all the variables are auxiliary and appear
in exactly two constraints. The percentage of eliminated variables varies a lot,
from as little as 5% to 96%, effectively eliminating all auxiliary variables in the
case of the openshop model. On average, var-elim-idxs extends Gecode with
2.3 new propagators, with a maximum of 9 propagators for an instance of the
cyclic-rcpsp model.
The column ratio search shows that preprocessing generally either improves
the node rate during search (ratio larger than 1.03) or leaves it almost unchanged
(ratio between 0.97 and 1.03). The node rate can be more than doubled: see the
l2p model. For the four models at the bottom of the table, however, the perfor-
mance is worse after preprocessing. On average, the node rate during search is
1.12 times higher. The geometric standard deviation is generally low, i.e., close to
1.0, for instances of the same MiniZinc model, except when some of the instances
are solved very fast, partly due to measurement errors.
The column ratio total shows that, when also counting the time for prepro-
cessing, the results are still promising. On average, the node rate is 0.81 times
lower using preprocessing. This number is strongly affected by instances that are
solved very fast. If we take into account only the 207 instances that originally
take more than one second to solve, then the node rate of var-elim-idxs is
on average identical to the one of Gecode. If we take into account only the 174
instances that originally take more than one minute to solve, then the node rate
of var-elim-idxs is on average 1.09 times higher.
Interestingly, Table 2 also shows that there is no strong correlation between
the number of eliminated variables and the node rate ratio. For instance, nearly
all auxiliary variables of the rcpsp model are eliminated but the node rate ratio
is close to 1, despite the fact that the number of propagator calls is divided
by two. This probably indicates that the generated indexical-based propagator
suffers from some inefficiencies.
The median preprocessing time for the 223 modified instances is 4.4 seconds,
roughly equally divided between the time spent by our code in Scala and the
time spent by the g++ compiler. The minimum time is 2.9 seconds, of which a
 Automated Auxiliary Variable Elimination 325

closer analysis shows that more than 2.5 seconds are actually spent in the set-up,
such as loading classes or parsing header files, independently of the size of the
instance or the number of identified patterns. It is important to note that neither
the g++ compiler nor our indexical compiler were developed for such a use-case,
as compilation is usually performed offline. The median preprocessing time for
the 172 instances unmodified by preprocessing is 0.9 seconds. The minimum time
is 0.7 seconds, again mostly spent in loading classes. The largest preprocessing
time observed is 30 seconds, in the case of a very large nmseq instance. Table 3
reports aggregated results for the 172 unmodified instances: the cost of uselessly
running the preprocessing is largely unnoticeable for unmodified instances that
take more than 60 seconds to be solved.

6 Discussion

In the light of the experimental results, this section discusses more thoroughly
the merits and limitations of our approach.

6.1 Related Work

Dealing with the introduction of auxiliary variables is an important challenge

for developers of both solvers and modelling languages.

Variable Views. The initial purpose of variable views [3,16] and domain
views [19] is to reduce the number of implemented propagators in a solver, but
they can also be used to eliminate some auxiliary variables. A view allows one
to post a constraint on an argument that is a function of a variable instead of
a variable. If a constraint is of the form x = f (y), where x and y are variables,
then one can introduce a view v = f (y) and replace the variable x by the view v.
Compared with our approach, views have the benefits that they are not limited
to variables appearing in two constraints and that they do not require generat-
ing new propagators, hence that they can eliminate variables that appear, for
example, in global constraints. Views are however in general limited to unary
functions, except in [3]. More importantly, to the best of our knowledge, no solver
automatically transforms a flattened constraint into a view.

Flattening and Modelling Techniques. Common subexpression elimina-

tion [13,14] reduces the number of auxiliary variables by merging into a single
variable all variables that represent the same expression. This also has the effect
of increasing the amount of filtering. Hence, as explained in Section 3.1, we do
not eliminate such variables.
Half-reification [5] is a modelling technique that replaces constraints of the
form B ≡ φ by B =⇒ φ, where B is a Boolean variable and φ a Boolean
expression. Although this does not reduce the number of variables, it can reduce
solving time by having simpler constraints. However, there are no half-reified
326 J.-N. Monette et al.

constraint predicates in FlatZinc. Our approach enables some optimisation in

the spirit of half-reification, as shown in the example of Section 4.
Model globalisation [10] aims at replacing a collection of constraints at the
MiniZinc level by an equivalent global constraint. Such a replacement usually
reduces the number of auxiliary variables and increases the amount of filtering,
provided the global constraint is not decomposed during flattening. Globalisation
may improve solving time much more than our approach but it is an offline and
interactive process, hence orthogonal to our online and automated approach.

Propagator Generation. Our approach uses our indexical compiler to gen-

erate propagators. The generation of stateless propagators [8] is an alternative
that can yield much faster propagators. It is however limited by the size of the
domains, as the constraint is essentially represented extensionally. It is meant to
be used offline, as are other approaches to propagator generation, such as [4].

6.2 Properties and Extensions

Unlike most of the approaches in Section 6.1, our approach is entirely online and
automated. We review here some of its properties and discuss possible extensions.

Search Tree Shape. Given a search strategy and enough time, the search
trees explored by Gecode on the original model and by var-elim-idxs on the
modified model are the same if all the following conditions are respected:
– The search strategy does not depend on the propagation order.
– The search strategy does not need to branch on the auxiliary variables.
– The generated propagators do the same filtering as the replaced propagators.
Except in the case of the roster model, where the search strategy is incompletely
specified, the two first conditions are respected for all the instances we used
in Section 5. The third condition is more difficult to check, but seems generally
respected: out of the 84 modified instances that did not time out, only 7 instances
had a different and always larger search tree, namely one fjsp instance, two league
instances, one roster instance, and three still-life-wastage instances.

Domains. Our approach assumes that the domains of the eliminated variables
are non-constraining because the shared variable X is existentially quantified
without specifying a domain. This is why we restricted ourselves to variables
introduced by the MiniZinc flattener, annotated with var is introduced, as
the domains proved to be non-constraining for those variables. However, auxil-
iary variables may also be introduced manually to simplify a model. A sound
way to extend our approach to such variables while retaining correctness is to
verify that the domain is non-constraining before considering a variable x for
replacement, by only considering how the propagators of the two constraints in
which x appears reduce the domain of x given the domains of their other vari-
ables. This would also let us apply our approach to other modelling languages
that do not have the var is introduced annotation, such as, e.g., XCSP [15].
 Automated Auxiliary Variable Elimination 327

Instances and Problems. We made a deliberate choice to work at the Flat-

Zinc, rather than MiniZinc, level for two reasons. First, it is much simpler to work
with a flat format than with a rich modelling language. Second, it might not be
clear before or during flattening what the frequent patterns are. This choice led
us to work with individual instances. However, instances from the same Mini-
Zinc model share the same frequent patterns. Hence, when several instances
of the same MiniZinc model must be solved successively, most of the results
of the preprocessing of the first instance can actually be reused to reduce the
preprocessing time of the following ones. In particular, when the preprocessing
does not modify the FlatZinc model, detecting this on small instances saves the
potentially high cost of unnecessarily parsing large instances.

Improved Propagator Generation. As seen in Table 2, var-elim-idxs does

not remove all the auxiliary variables. Partly, this is not a limitation of our app-
roach but of its implementation. Increasing the reach of our approach amounts to
improving the generation of the propagators in order to handle efficiently more
constraints, including n-ary ones. This can be done by improving our indexical
compiler [11] or by using other techniques such as [8] or [3], but such improve-
ments are orthogonal to this paper. Our experiments show that our approach is
already practical as it is.

Increased Filtering. When identifying patterns, if more than one shared vari-
able is identified, then it is possible to generate propagators achieving more
filtering. Our approach can be extended to multiple shared variables. However,
for it to be worthwhile, it is necessary to ensure that the propagator genera-
tion takes advantage of multiple occurrences of a variable other than X in the
checker. This is currently not the case but an interesting line of future work.

7 Conclusion
We presented a new approach to eliminate many of the auxiliary variables
introduced into a flattened constraint-based model. Our approach adds a
preprocessing step that modifies the flattened model and extends the solver
with propagators generated on the fly for new constraint predicates. This is
made possible through the generation of indexical-based propagators from log-
ical formulas. Experiments with our prototype implementation show that our
approach makes a solver about 9% faster on average, and sometimes more than
2 times faster, for instances that take more than one minute to solve. This indi-
cates that our preprocessing should be activated for instances that are difficult
to solve, which are the ones for which it is important to decrease solving time.

Acknowledgements. This work is supported by grants 2011-6133 and 2012-

4908 of the Swedish Research Council (VR). We thank the anonymous reviewers
for their constructive and insightful comments.
328 J.-N. Monette et al.

References
1. Becket, R.: Specification of FlatZinc. http://www.minizinc.org/downloads/doc-1.
6/flatzinc-spec.pdf
2. Beldiceanu, N., Carlsson, M., Demassey, S., Petit, T.: Global constraint catalogue:
Past, present, and future. Constraints 12(1), 21–62 (2007). The catalogue is at
http://sofdem.github.io/gccat
3. Correia, M., Barahona, P.: View-based propagation of decomposable constraints.
Constraints 18(4), 579–608 (2013)
4. Dao, T.B.H., Lallouet, A., Legtchenko, A., Martin, L.: Indexical-based solver learn-
ing. In: Van Hentenryck, P. (ed.) CP 2002. LNCS, vol. 2470, pp. 541–555. Springer,
Heidelberg (2002)
5. Feydy, T., Somogyi, Z., Stuckey, P.J.: Half reification and flattening. In: Lee, J.
(ed.) CP 2011. LNCS, vol. 6876, pp. 286–301. Springer, Heidelberg (2011)
6. Frisch, A.M., Grum, M., Jefferson, C., Martinez Hernandez, B., Miguel, I.: The
design of ESSENCE: a constraint language for specifying combinatorial problems.
In: IJCAI 2007, pp. 80–87. Morgan Kaufmann (2007)
7. Gecode Team: Gecode: A generic constraint development environment (2006).
http://www.gecode.org
8. Gent, I.P., Jefferson, C., Linton, S., Miguel, I., Nightingale, P.: Generating custom
propagators for arbitrary constraints. Artificial Intelligence 211, 1–33 (2014)
9. Hassani Bijarbooneh, F.: Constraint Programming for Wireless Sensor Networks.
Ph.D. thesis, Department of Information Technology, Uppsala University, Sweden
(2015). http://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-241378
10. Leo, K., Mears, C., Tack, G., Garcia de la Banda, M.: Globalizing constraint
models. In: Schulte, C. (ed.) CP 2013. LNCS, vol. 8124, pp. 432–447. Springer,
Heidelberg (2013)
11. Monette, J.-N., Flener, P., Pearson, J.: Towards solver-independent propagators.
In: Milano, M. (ed.) CP 2012. LNCS, vol. 7514, pp. 544–560. Springer, Heidelberg
(2012)
12. Nethercote, N., Stuckey, P.J., Becket, R., Brand, S., Duck, G.J., Tack, G.: MiniZinc:
towards a standard CP modelling language. In: Bessière, C. (ed.) CP 2007. LNCS,
vol. 4741, pp. 529–543. Springer, Heidelberg (2007). The MiniZinc toolchain is
available at http://www.minizinc.org
13. Nightingale, P., Akgün, Ö., Gent, I.P., Jefferson, C., Miguel, I.: Automatically
improving constraint models in savile row through associative-commutative com-
mon subexpression elimination. In: O’Sullivan, B. (ed.) CP 2014. LNCS, vol. 8656,
pp. 590–605. Springer, Heidelberg (2014)
14. Rendl, A., Miguel, I., Gent, I.P., Jefferson, C.: Automatically enhancing constraint
model instances during tailoring. In: Bulitko, V., Beck, J.C. (eds.) SARA 2009.
AAAI Press (2009)
15. Roussel, O., Lecoutre, C.: XML representation of constraint networks: Format
XCSP 2.1. CoRR abs/0902.2362 (2009). http://arxiv.org/abs/0902.2362
16. Schulte, C., Tack, G.: View-based propagator derivation. Constraints 18(1), 75–107
(2013)
17. Scott, J.D.: Rapid prototyping of a structured domain through indexical compila-
tion. In: Schaus, P., Monette, J.N. (eds.) Domain Specific Languages in Combina-
torial Optimisation (CoSpeL workshop at CP 2013) (2013). http://cp2013.a4cp.
org/workshops/cospel
 Automated Auxiliary Variable Elimination 329

18. Sidebottom, G., Havens, W.S.: Nicolog: A simple yet powerful cc(FD) language.
Journal of Automated Reasoning 17, 371–403 (1996)
19. Van Hentenryck, P., Michel, L.: Domain views for constraint programming. In:
O’Sullivan, B. (ed.) CP 2014. LNCS, vol. 8656, pp. 705–720. Springer, Heidelberg
(2014)
20. Van Hentenryck, P., Saraswat, V., Deville, Y.: Design, implementation, and eval-
uation of the constraint language cc(FD). Tech. Rep. CS-93-02, Brown University,
Providence, USA (January 1993), revised version in Journal of Logic Program-
ming 37(1–3), 293–316 (1998). Based on the unpublished manuscript Constraint
Processing in cc(FD) (1991)
 Automatically Improving SAT Encoding
of Constraint Problems Through Common
Subexpression Elimination in Savile Row

Peter Nightingale(B) , Patrick Spracklen, and Ian Miguel

School of Computer Science, University of St Andrews, St Andrews, UK

{pwn1,jlps,ijm}@st-andrews.ac.uk

Abstract. The formulation of a Propositional Satisfiability (SAT) prob-

lem instance is vital to efficient solving. This has motivated research on
preprocessing, and inprocessing techniques where reformulation of a SAT
instance is interleaved with solving. Preprocessing and inprocessing are
highly effective in extending the reach of SAT solvers, however they nec-
essarily operate on the lowest level representation of the problem, the raw
SAT clauses, where higher-level patterns are difficult and/or costly to
identify. Our approach is different: rather than reformulate the SAT rep-
resentation directly, we apply automated reformulations to a higher level
representation (a constraint model) of the original problem. Common
Subexpression Elimination (CSE) is a family of techniques to improve
automatically the formulation of constraint satisfaction problems, which
are often highly beneficial when using a conventional constraint solver.
In this work we demonstrate that CSE has similar benefits when the
reformulated constraint model is encoded to SAT and solved using a
state-of-the-art SAT solver. In some cases we observe speed improve-
ments of over 100 times.

1 Introduction

The Propositional Satisfiability Problem (SAT) is to find an assignment to a set

of Boolean variables so as to satisfy a given Boolean formula, typically expressed
in conjunctive normal form [4]. SAT has many important applications, such
as hardware design and verification, planning, and combinatorial design [14].
Powerful, robust solvers have been developed for SAT employing techniques
such as conflict-driven learning, watched literals, restarts and dynamic heuristics
for backtracking solvers [15], and sophisticated incomplete techniques such as
stochastic local search [22].
The formulation of a SAT problem instance is vital to efficient solving. This
has motivated research on preprocessing [7,27], and inprocessing [12] where refor-
mulation of the SAT instance is interleaved with solving. Both techniques are
highly effective in extending the reach of SAT solvers, however they necessarily
operate on the lowest level representation of the problem, the raw SAT clauses,
where higher-level patterns are difficult and/or costly to identify.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 330–340, 2015.
DOI: 10.1007/978-3-319-23219-5 23
 Automatically Improving SAT Encoding of Constraint Problems 331

Our approach is different: rather than reformulate the SAT representation

directly, we apply automated reformulations to a higher level representation of
the original problem. An increasingly popular means of deriving SAT formula-
tions is by taking a constraint model and employing a set of automated encoding
steps to produce an equivalent SAT formulation [28]. Constraint satisfaction is
a formalism closely related to SAT in which we seek an assignment of values to
decision variables so as to satisfy a set of constraints [21]. Constraint modelling
languages typically support decision variables with richer domains and a richer
set of constraints than the CNF used with SAT. Hence, an input problem can
be expressed conveniently in a higher level constraint language, while employing
efficient SAT solvers to find solutions.
Common Subexpression Elimination (CSE) is a very well established tech-
nique in compiler construction [5]. In that context the value of a previously-
computed expression is used to avoid computing the same expression again.
Shlyakhter et al [23] exploited identical subformulae during grounding out of
quantified Boolean formulae. Similarly, it is a useful technique in the automatic
improvement of constraint models, where it acts to reduce the size of a con-
straint model by removing redundant variables and constraints [1,10,19,20]. This
in turn can create a stronger connection between different parts of the model,
resulting in stronger inference and reduced search during constraint solving.
Earlier work applied CSE directly to SAT formulations, with limited suc-
cess [29]. Herein we establish the success of an alternative approach in which
CSE is applied to a constraint model prior to SAT encoding. We apply CSE
to a constraint problem instance expressed in the constraint modelling language
Essence , which includes integer (as well as Boolean) decision variables, a set of
infix operators on integer and Boolean expressions, and various global constraints
and functions [20]. The reformulated constraint model is automatically encoded
to SAT using the Savile Row system, yielding substantial improvements in
SAT solver runtime over encoding without CSE.
Our method has the advantage of allowing us to exploit patterns present in
the high level description of the problem that are obscured in the SAT formula-
tion, and so very difficult to detect using SAT pre/inprocessing approaches. As
a simple example, a decision variable in a constraint model typically requires a
collection of SAT variables and clauses to encode. If, via CSE, we are able to
reduce two such variables to one then the set of Boolean variables and clauses
to encode the second variable will never be added to the SAT formulation. Per-
forming the equivalent step directly on the SAT formulation would require the
potentially very costly step of identifying the structure representing the second
variable then proving its equivalence to the structure encoding the first.
In performing CSE on a constraint model preparatory to SAT encoding,
we have modified and enhanced existing constraint model CSE approaches to
take into account that the eventual target is a SAT formulation. Firstly the
set of candidate expressions for CSE differs when the target is SAT. Secondly,
implied constraints that are added to elicit common subexpressions are removed
following CSE if they are unchanged. In addition we describe for the first time
332 P. Nightingale et al.

Algorithm 1. Identical-CSE(AST, ST)

Require: AST: Abstract syntax tree representing the model
Require: ST: Symbol table containing CSP decision variables
1: newcons ← empty list {Collect new constraints}
2: map ← empty hash table mapping expressions to lists
3: populateMap(AST, map)
4: for all key in map in decreasing size order do
5: ls ← map(key) {ls is a list of identical AST nodes}
6: ls ← filter(isAttached, ls) {Remove AST nodes no longer contained in AST or
newcons}
7: if length(ls) > 1 then
8: e ← head(ls)
9: bnds ← bounds(e)
10: aux ← ST.newAuxVar(bnds)
11: newc ← ( e = aux ) {New constraint defining aux }
12: newcons.append(newc)
13: for all a ∈ ls do
14: Replace a with copy(aux ) within AST or newcons
15: AST ← AST ∧ fold(∧, newcons)

Algorithm 2. populateMap(A, map)

Require: A: Reference to an abstract syntax tree
Require: map: Hash table mapping expressions to lists
1: if A is a candidate for CSE then
2: Add A to list map[A]
3: for all child ∈ A.Children() do
4: populateMap(child, map)

an identical CSE algorithm that is independent of general flattening, allowing

flexibility to extract common subexpressions would not ordinarily be flattened
and to control the order of CSE.

2 CSE for SAT Encoding

The simplest form of CSE that we consider is Identical CSE, which extracts
sets of identical expressions. Suppose x × y occurs three times in a model. Identi-
cal CSE would introduce a new decision variable a and new constraint x × y = a.
The three original occurrences of x × y would be replaced by a. In Savile Row,
Identical CSE is implemented with Algorithm 1. Andrea Rendl’s Tailor [10,20]
and MiniZinc [13,26] also implement Identical CSE, however (in contrast to Tai-
lor and MiniZinc) our algorithm is not tied to the process of flattening nested
expressions into primitive expressions supported directly by the constraint solver.
This is advantageous because it allows us to identify and exploit common subex-
pressions in expressions that do not need to be flattened. The SMT solver CVC4
merges identical subtrees in its abstract syntax tree [3]. It is not clear whether
this affects the search or is simply a memory saving feature.
 Automatically Improving SAT Encoding of Constraint Problems 333

The first step is to recursively traverse the model (by calling Algorithm 2)
to collect sets of identical expressions. Algorithm 2 collects only expressions
that are candidates for CSE. Atomic variables and constants are not candidates.
Compound expressions are CSE candidates by default, however when the target
is a SAT encoding we exclude all compound expressions that can be encoded as
a single SAT literal. This avoids creating a redundant SAT variable that is equal
to (or the negation of) another SAT variable, thus improving the encoding. The
following expressions are not candidates: x = c, x = c, x ≤ c, x < c, x ≥ c,
x > c, ¬x (where x is a decision variable and c is a constant).
The second step of Identical CSE is to iterate through sets of expressions
in decreasing size order (line 4). When an expression e is eliminated by CSE,
the number of occurrences of any expressions contained in e is reduced. There-
fore eliminating long expressions first may obviate the need to eliminate short
expressions. For each set (of size greater than one) of identical expressions a new
decision variable aux is created, and each of the expressions is replaced with aux.
One of the expressions e in the set is used to create a new constraint e = aux .
Crucially the new constraint contains the original object e so it is possible to
extract further CSEs from within e.
Prior to running Identical CSE the model is simplified by evaluating all con-
stant expressions and placing it into negation normal form. In addition some
type-specific simplifications are performed (eg x ↔ true rewrites to x). Commu-
tative expressions (such as sums) are sorted to make some equivalent expressions
syntactically identical.
In our previous work we investigated Associative-Commutative CSE (AC-
CSE) for constraint solvers [19] and in that context Identical CSE was always
enabled. Identical CSE is complementary to AC-CSE.
Active CSE. Active CSE extends Identical CSE by allowing non-identical
expressions to be extracted using a single auxiliary variable. For example, sup-
pose we have x = y and x = y in the model. We can introduce a single Boolean
variable a and a new constraint a ↔ (x = y), then replace x = y with a and
x = y with ¬a. For solvers that support negation (such as SAT solvers) ¬a can
be expressed in the solver input language with no further rewriting, so we have
avoided encoding both x = y and x = y.
The Active CSE algorithm implemented in Savile Row is an extension
of Algorithm 1. The algorithm works as follows: for each candidate expres-
sion e a simple transformation is applied to it (for example producing ¬e). The
transformed expression is placed into the normal form and commutative subex-
pressions are sorted. The algorithm then queries map to discover expressions
matching the transformed expression.
Active CSE as implemented in Savile Row 1.6.3 applies four transforma-
tions: Boolean negation, arithmetic negation, multiply by 2, and multiply by -2.
Rendl implemented Boolean negation active CSE in her Tailor system, along
with active reformulations based upon De Morgan’s laws and Horn clauses [20].
In Savile Row, the use of negation normal form obviates the use of the latter
two. To our knowledge MiniZinc [13,26] does not implement Active CSE.
334 P. Nightingale et al.

Associative-Commutative CSE (AC-CSE). Nightingale et al [19] (for finite

domains) and Araya et al [1] (for numerical CSP) established the use of AC-
CSE for constraint models. To our knowledge neither Tailor [10,20] nor MiniZ-
inc [13,26] implement AC-CSE. It exploits the properties of associativity and
commutativity of binary operators, such as in sum constraints. For SAT encod-
ing, our approach refines the procedure for AC-CSE given in Nightingale et al.
In that procedure, implied sum constraints are added, which are deduced from
global constraints in the model, such as all-different and global cardinality. These
implied sums are used to trigger AC-CSE. Since large sum constraints are cum-
bersome to encode in SAT, and can therefore degrade performance, we add a
test to check whether the implied sums are modified following AC-CSE. If not,
they are deemed not to be useful and removed prior to SAT encoding.
Extended resolution [2] is gaining interest and can be viewed as AC-CSE
applied directly to the disjunctive clauses of a SAT formula.
Effects of CSE on the Output Formula. We give a short example of a
constraint reformulation and its effect on the SAT encoding. Suppose we have
two occurrences of x × y, both are contained in sums, and x, y ∈ {1 . . . 10}.
Ordinarily we would create a new auxiliary variable (a1 , a2 ∈ {1 . . . 100}) for
each occurrence, and add two new constraints: x×y = a1 and x×y = a2 . Both a1
and a2 would be encoded using just under 200 SAT variables and approximately
400 clauses each. Also, both new constraints would be encoded using 100 clauses
each. In contrast, Identical CSE would create a single auxiliary variable for both
occurrences of x × y, and there would be one new constraint, saving hundreds of
SAT variables and clauses. It is difficult to see how SAT pre/inprocessing rules
could identify the structure that was exploited by Identical CSE.

3 Experimental Evaluation

Our goal is to investigate whether reformulations performed on a constraint

problem instance are beneficial when the problem instance is solved by encoding
to SAT and using a state-of-the-art SAT solver. To achieve this we need to
ensure that the baseline encoding to SAT is sensible. Therefore we have used
standard encodings from the literature such as the order encoding for sums [28]
and support encoding [8] for binary constraints. Also we do not attempt to
encode all constraints in the language: several constraint types are decomposed
before encoding to SAT. Details are given in the Savile Row tutorial 1.6.3
appendix A [18].
In our experiments we compare four configurations of Savile Row: Basic,
which includes the default options of unifying equal variables, filtering domains
and aggregation; Identical CSE, which is Basic plus Identical CSE; Identical &
Active CSE, which is Basic plus the two named CSE algorithms, and Identical,
Active & AC-CSE, which is Basic plus all three CSE algorithms. Our benchmark
set is the set of example problems included with Savile Row 1.6.3 [18]. There
are 49 problem classes including common benchmark problems such as EFPA
 Automatically Improving SAT Encoding of Constraint Problems 335

Fig. 1. Identical CSE vs Basic (upper left), Active & Identical CSE vs Identical CSE
(upper right), Identical, Active & AC-CSE vs Identical & Active CSE (lower left), same
vs Basic (lower right).

[11] and car sequencing [6] as well as less common problems such as Black Hole
solitaire [9]. In total there are 492 problem instances.
Experiments were run with 32 processes in parallel on a machine with two
16-core AMD Opteron 6272 CPUs at 2.1 GHz and 256 GB RAM. We used
the SAT solver Lingeling [12] which was winner of the Sequential, Application
SAT+UNSAT track of the SAT competition 2014. We downloaded lingeling-ayv-
86bf266-140429.zip from http://fmv.jku.at/lingeling/. We used default options
for Lingeling so inprocessing was switched on. All times reported include Savile
Row time and Lingeling’s reported time, and are a median of 10 runs with 10
different random seeds given to Lingeling. A time limit of one hour was applied.
We used a clause limit of 100 million, and for instances that exceeded the clause
limit we treated them as if they timed out at 3600s (to allow comparison with
others). Of 492 instances, 7 reached the clause limit with Basic and 6 with the
other configurations.
Summary Plots for Full Set of Benchmarks. In Figures 1–2 we present a
summary view of the performance of our CSE methods over our full set of 49
benchmark problem classes. Figure 1 (upper left) compares the basic encoding
with identical CSE. On easier problem instances CSE has a limited effect, but
as problem difficulty increases so does the potential of identical CSE to reduce
search effort very significantly - in some cases by over 20 times. There are a small
number of outliers among the harder instances where identical CSE degrades
overall performance. We conjecture that this is due to the change in problem
336 P. Nightingale et al.

Fig. 2. Identical, Active & AC-CSE vs Identical & Active CSE plotted against reduc-
tion in SAT variables. The plot on the right is a subset of the left.

structure affecting the heuristics of Lingeling. The degradation effect is limited

compared with the potential for a large speedup and the number of outliers is
small. The geometric mean speed-up is 1.24.
In Figure 1 (upper right) we compare identical CSE alone with identical CSE
combined with active CSE. The results show that this additional step is largely
neutral or incurs a very small overhead of performing the active CSE checks, but
that there are a number of occasions where active CSE significantly enhances
identical CSE. Again, there are a small number of outliers where performance is
significantly degraded, which we again believe to be due to a bad interaction with
the SAT solver search strategy. The geometric mean speed-up is 0.98, indicating
a very small average slow-down.
Figure 1 (lower left and right) plots the utility of AC-CSE. In some cases we
see a very considerable improvement in performance, however there are also cases
where performance is degraded. Five notable problem classes have been sepa-
rated in the plots. Of these, Killer Sudoku is the most ambiguous, with clusters
of instances both above and below the break-even line. For Car Sequencing and
SONET, some of the easier instances are below the break-even line, but the
more difficult instances exhibit a speed-up. Peg Solitaire Action is degraded on
all seven instances. Molnars exhibits a speed up with one exception. Over all
instances the geometric mean speed-up is 1.24.
To partly explain these results, we measured the size of the formula produced
with and without AC-CSE. Figure 2 has the same y-axis as Figure 1 (lower left)
but with a different x-axis: the ratio of the number of variables in the SAT
formula. Values of x above 1 indicate that applying AC-CSE has reduced the
number of SAT variables. For Killer Sudoku, there is a clear link between the
number of SAT variables in the formula and the speed up quotient. It is also
worth noting that for all instances of Peg Solitaire Action applying AC-CSE
both increases the number of SAT variables and degrades performance. On the
other hand Car Sequencing and SONET show no correlation between speed up
 Automatically Improving SAT Encoding of Constraint Problems 337

Fig. 3. Identical & Active CSE vs Basic: Car Sequencing (left) and SONET (right).

quotient and reduction of SAT variables, indicating that the number of SAT
variables alone is a very coarse measure of difficulty.
Case Study 1: Car Sequencing Our Essence model of Car Sequencing [24]
uses a sequence of integer variables x[1 . . . n] to represent the sequence of cars on
a production line. For each option (to be fitted at a station on the production
line) we have a limit on the proportion of cars: at most p of any q adjacent
cars may have the option installed so as not to overload the station. To model
this we employ overlapping sums of length q containing x[i] ∈ S, where S is the
set of car classes that have the option installed, and i is an index within the
subsequence of length q.
The number of each car class to build is enforced with a global cardinality
constraint [17] on x. Also, for each option we know how many  cars require
n
that option in total (t) thus we add a further implied constraint: i=1 (x[i] ∈
S) = t. We experiment with the 80 instances used in Nightingale [17]. Identical
and Active CSE are both able to extract the expressions x[i] ∈ S that appear
in several sum constraints, avoiding multiple SAT encodings of the same set-
inclusion constraint. Figure 3 (left) plots the time improvement of Active CSE
compared with the Basic encoding. The improvement is substantial and increases
with the difficulty of the problem instances.
AC-CSE is able to extract common subexpressions among the sum con-
straints for a given option. The p of q constraints overlap with each other and also
with the implied constraint for the option. Figure 1 (lower left) plots the time
improvement of adding AC-CSE to identical and active CSE. The additional
improvement is substantial, with many instances becoming solvable within one
hour and a peak speed up of over 100 times. With the Basic encoding 13 of
the 80 instances time out at one hour. In contrast, when combining Identical,
Active and AC-CSE we found that only two instances timed out. The other 11
are solved within one hour, most with very substantial speed-ups.
Case Study 2: SONET The SONET problem [16,25] is a network design
problem where each node is installed on a set of rings (fibre-optic connections).
If two nodes are required to be connected, there must exist a ring on which
they are both installed. We use the simplified SONET problem where each ring
338 P. Nightingale et al.

has unlimited data capacity (Section 3 of [25]). Rings are indistinguishable so

we use lexicographic ordering constraints to order the rings in non-decreasing
order. This is an optimisation problem: the number of node-ring connections
is minimised. The problem formulation and set of 24 instances are exactly as
described in Nightingale et al [19].
Figure 3 (right) compares Identical and Active CSE to the Basic encoding
on this problem class. The initial formulation of SONET contains no identical or
active common subexpressions, however each decomposition of a lexicographic
ordering constraint has identical subexpressions that are exploited by Identical
CSE, causing the modest gains seen in the plot. There are four groups of con-
straints in SONET: the objective function, the constraints ensuring nodes are
connected when required, a constraint for each ring limiting the number of nodes,
and the symmetry breaking constraints. Apart from symmetry breaking all con-
straints are sums and all three groups overlap, therefore AC-CSE is successful
on this problem as shown in Figure 1 (lower left).

4 Conclusion
Common Subexpression Elimination has proven to be a valuable tool in the
armoury of reformulations applied to constraint models, however hitherto there
has only been limited success in applying CSE to SAT formulations [29]. We
have shown how CSE can be used to improve SAT formulations derived through
an automated encoding process from constraint models. Our approach has the
advantage that it can identify and exploit structure present in a constraint model
that is subsequently obscured by the encoding process, while still taking advan-
tage of powerful SAT solvers. The result is a method that, when applicable, can
produce a very significant reduction in search effort.
We have evaluated our approach on a wide range of benchmark problems. On
some instances we observed improvements of SAT solver speed of over 50 times.
On the car sequencing problem, for example, the peak speed increase is over 100
times. With the basic approach, 13 of 80 car sequencing instances could not be
solved in one hour, whereas with the full CSE approach only two instances could
not be solved.

Acknowledgments. We wish to thank the EPSRC for funding this work through
grants EP/H004092/1 and EP/M003728/1, and Christopher Jefferson for helpful dis-
cussions.

References
1. Araya, I., Neveu, B., Trombettoni, G.: Exploiting common subexpressions in
numerical CSPs. In: Stuckey, P.J. (ed.) CP 2008. LNCS, vol. 5202, pp. 342–357.
Springer, Heidelberg (2008)
2. Audemard, G., Katsirelos, G., Simon, L.: A restriction of extended resolution for
clause learning sat solvers. In: Proceedings of the Twenty-Fourth AAAI Conference
on Artificial Intelligence (2010)
 Automatically Improving SAT Encoding of Constraint Problems 339

3. Barrett, C., Conway, C.L., Deters, M., Hadarean, L., Jovanović, D., King, T.,
Reynolds, A., Tinelli, C.: CVC4. In: Gopalakrishnan, G., Qadeer, S. (eds.) CAV
2011. LNCS, vol. 6806, pp. 171–177. Springer, Heidelberg (2011)
4. Biere, A., Heule, M., van Maaren, H.: Handbook of Satisfiability, vol. 185. IOS
Press (2009)
5. Cocke, J.: Global common subexpression elimination. ACM Sigplan Notices 5(7),
20–24 (1970)
6. Dincbas, M., Simonis, H., Van Hentenryck, P.: Solving the car-sequencing problem
in constraint logic programming. In: Proceedings of the 8th European Conference
on Artificial Intelligence (ECAI 1988), pp. 290–295 (1988)
7. Eén, N., Biere, A.: Effective preprocessing in SAT through variable and clause
elimination. In: Bacchus, F., Walsh, T. (eds.) SAT 2005. LNCS, vol. 3569,
pp. 61–75. Springer, Heidelberg (2005)
8. Gent, I.P.: Arc consistency in SAT. In: Proceedings of the 15th European Confer-
ence on Artificial Intelligence (ECAI 2002), pp. 121–125 (2002)
9. Gent, I.P., Jefferson, C., Kelsey, T., Lynce, I., Miguel, I., Nightingale, P., Smith,
B.M., Tarim, S.A.: Search in the patience game ‘black hole’. AI Communications
20(3), 211–226 (2007)
10. Gent, I.P., Miguel, I., Rendl, A.: Tailoring solver-independent constraint models:
a case study with Essence and Minion. In: Miguel, I., Ruml, W. (eds.) SARA
2007. LNCS (LNAI), vol. 4612, pp. 184–199. Springer, Heidelberg (2007)
11. Huczynska, S., McKay, P., Miguel, I., Nightingale, P.: Modelling equidistant fre-
quency permutation arrays: an application of constraints to mathematics. In: Gent,
I.P. (ed.) CP 2009. LNCS, vol. 5732, pp. 50–64. Springer, Heidelberg (2009)
12. Järvisalo, M., Heule, M.J.H., Biere, A.: Inprocessing rules. In: Gramlich, B.,
Miller, D., Sattler, U. (eds.) IJCAR 2012. LNCS, vol. 7364, pp. 355–370. Springer,
Heidelberg (2012)
13. Leo, K., Tack, G.: Multi-pass high-level presolving. In: Proceedings of the 24th
International Joint Conference on Artificial Intelligence (IJCAI) (to appear, 2015)
14. Marques-Silva, J.: Practical applications of boolean satisfiability. In: 9th Interna-
tional Workshop on Discrete Event Systems (WODES 2008), pp. 74–80 (2008)
15. Moskewicz, M.W., Madigan, C.F., Zhao, Y., Zhang, L., Malik, S.: Chaff: engineer-
ing an efficient SAT solver. In: Proceedings of the 38th Annual Design Automation
Conference, pp. 530–535. ACM (2001)
16. Nightingale, P.: CSPLib problem 056: Synchronous optical networking (SONET)
problem. http://www.csplib.org/Problems/prob056
17. Nightingale, P.: The extended global cardinality constraint: An empirical survey.
Artificial Intelligence 175(2), 586–614 (2011)
18. Nightingale, P.: Savile Row, a constraint modelling assistant (2015). http://
savilerow.cs.st-andrews.ac.uk/
19. Nightingale, P., Akgün, Ö., Gent, I.P., Jefferson, C., Miguel, I.: Automatically
improving constraint models in Savile Row through associative-commutative com-
mon subexpression elimination. In: O’Sullivan, B. (ed.) CP 2014. LNCS, vol. 8656,
pp. 590–605. Springer, Heidelberg (2014)
20. Rendl, A.: Effective Compilation of Constraint Models. Ph.D. thesis, University of
St Andrews (2010)
21. Rossi, F., van Beek, P., Walsh, T. (eds.) Handbook of Constraint Programming.
Elsevier (2006)
22. Shang, Y., Wah, B.W.: A discrete lagrangian-based global-search method for solv-
ing satisfiability problems. Journal of Global Optimization 12(1), 61–99 (1998)
340 P. Nightingale et al.

23. Shlyakhter, I., Sridharan, M., Seater, R., Jackson, D.: Exploiting subformula shar-
ing in automatic analysis of quantified formulas. In: Sixth International Conference
on Theory and Applications of Satisfiability Testing (SAT 2003) (2003), poster
24. Smith, B.: CSPLib problem 001: Car sequencing. http://www.csplib.org/
Problems/prob001
25. Smith, B.M.: Symmetry and search in a network design problem. In: Barták,
R., Milano, M. (eds.) CPAIOR 2005. LNCS, vol. 3524, pp. 336–350. Springer,
Heidelberg (2005)
26. Stuckey, P.J., Tack, G.: MiniZinc with functions. In: Gomes, C., Sellmann, M.
(eds.) CPAIOR 2013. LNCS, vol. 7874, pp. 268–283. Springer, Heidelberg (2013)
27. Subbarayan, S., Pradhan, D.K.: NiVER: non-increasing variable elimination reso-
lution for preprocessing SAT instances. In: Hoos, H.H., Mitchell, D.G. (eds.) SAT
2004. LNCS, vol. 3542, pp. 276–291. Springer, Heidelberg (2005)
28. Tamura, N., Taga, A., Kitagawa, S., Banbara, M.: Compiling finite linear CSP into
SAT. Constraints 14(2), 254–272 (2009)
29. Yan, Y., Gutierrez, C., Jeriah, J.C., Bao, F.S., Zhang, Y.: Accelerating SAT solving
by common subclause elimination. In: Proceedings of the Twenty-Ninth AAAI
Conference on Artificial Intelligence (AAAI 2015), pp. 4224–4225 (2015)
 Exact Sampling for Regular and Markov
Constraints with Belief Propagation

Alexandre Papadopoulos1,2(B) , François Pachet1,2 ,

Pierre Roy1 , and Jason Sakellariou1,2
1
Sony CSL, 6 Rue Amyot, 75005 Paris, France
roy@csl.sony.fr
2
Sorbonne Universités, UPMC University Paris 06, UMR 7606, LIP6,
75005 Paris, France
{alexandre.papadopoulos,jason.sakellariou}@lip6.fr, pachet@csl.sony.fr

Abstract. Sampling random sequences from a statistical model, subject

to hard constraints, is generally a difficult task. In this paper, we show
that for Markov models and a set of Regular global constraints and
unary constraints, we can perform perfect sampling. This is achieved
by defining a factor graph, composed of binary factors that combine
a Markov chain and an automaton. We apply a simplified version of
belief propagation to sample random sequences satisfying the global con-
straints, with their correct probability. Since the factor graph is linear,
this procedure is efficient and exact. We illustrate this approach to the
generation of sequences of text or music, imitating the style of a corpus,
and verifying validity constraints, such as syntax or meter.

Keywords: Global constraints · Unary constraints · Markov

constraints · Belief propagation · Sampling

1 Introduction
Generating novel sequences, such as text or music, that imitate a given style is
usually achieved by replicating statistical properties of a corpus. This inherently
stochastic process can be typically performed by sampling a probability distri-
bution. In practice, we often need to impose additional properties on sequences,
such as syntactic patterns for text, or meter for music, that are conveniently
stated using constraint satisfaction approaches. However, typical constraint sat-
isfaction procedures are not concerned with the distribution of their solutions.
On the other hand, traditional sampling algorithms are generally not suited to
satisfy hard constraints, since they can suffer from high rejection rates or lack
coverage of the solution space. Both issues can be avoided, in some cases, by
taking advantage of constraint programming techniques.
In this paper, we show how to sample Markov sequences subject to a conjunc-
tion of Regular constraints [22], i.e., constraints stated with an automaton,
as well as additional unary constraints. Regular grammars can express parts-
of-speech patterns on text. In music, Meter [26] constrains Markov temporal

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 341–350, 2015.
DOI: 10.1007/978-3-319-23219-5 24
342 A. Papadopoulos et al.

sequences to be metrically correct. Meter can also be expressed as an automa-

ton, as we explain further in this paper. We achieve this result by defining a tree-
structured factor graph composed of unary and binary factors. The variables of
this graphical model represent the elements of the sequence, and binary factors
encode a type of conjunction between the Markov model and the automaton.
We apply belief propagation to sample sequences with their right probability.

1.1 Related Work

The combination of statistical and logical methods has been an active research
direction in artificial intelligence in the last few years. In constraint program-
ming, stochastic techniques are often used for guiding search, but less for charac-
terising solutions. Some work studies the impact of search heuristics on solution
diversity [27], but such endeavours tend to use optimisation techniques [11,13].
Conversely, introducing constraints to probabilistic graphical models is prob-
lematic since hard constraints introduce many zero probabilities, and this causes
typical sampling algorithms to suffer from high rejection rates. To overcome such
issues, Gogate and Dechter proposed SampleSearch [10], with a guaranteed uni-
form sampling of the solutions of a CSP, using a complete solver to reduce rejec-
tion rates. Likewise, Ermon et. al [8] use a constraint solver in a blackbox scheme,
and sample the solution space uniformly, often with better performance. In SAT,
Markov logic networks is a well established formalism that unifies probabilistic
and deterministic properties [7,25]. MC-SAT [24] samples from a non-uniform
distribution of the satisfying assignments of a SAT formula. Such methods, with
applications in verification, model checking, or counting problems, are general
but expensive. The solution we propose, which focuses on a specific setting, is
not derivable from such general methods, and is both tractable and exact.

2 Sequence Generation with Markov constraints

A Markov chain is a stochastic process, where the probability for state Xi , a

random variable, depends only on the last state Xi−1 . Each random variable
Xi takes values amongst an alphabet, denoted X . Seen as a generative pro-
cess, a Markov chain produces sequence X1 , . . . , Xn with a probability P (X1 ) ·
P (X2 |X1 ) · · · P (Xn |Xn−1 ). Order k Markov chains have a longer memory: the
Markov property states that P (Xi |X1 , . . . , Xi−1 ) = P (Xi |Xi−k , . . . , Xi−1 ). They
are equivalent to order 1 Markov chains on an alphabet composed of k-grams,
and therefore we assume only order 1 Markov chains.
Markov chains have been classically used for generating sequences that imi-
tate a given style [5,14,23]. A Markov chain is trained by learning the transition
probabilities on a corpus. For example, a musical piece can be represented as a
sequence of complex objects, constituted of pitch, duration, metrical position,
and more [17]. A Markov chain trained on this corpus will produce musical
sequences in the style of the composer. With text, we can use a Markov chain
 Exact Sampling for Regular and Markov Constraints with BP 343

whose alphabet is the set of words of the corpus, to generate new sentences in
the style of its author.
Markov generation can be controlled using Markov constraints. This allows
us to specify additional properties that a sequence should verify. For example,
Meter [26] imposes that sequences of notes are metrically correct. Often, such
constraints can be conveniently stated using a Regular constraint [22], defined
with an automaton A = Q, Σ, δ, q0 , F , where Q is a set of states, Σ an alphabet
defining labels on transitions, δ the transition function linking a state q ∈ Q and
a label a ∈ Σ to the successor state q  = δ(q, a), q0 ∈ Q the initial state,
and F ⊆ Q the set of accepting states. In this case, we have Σ = X , i.e.
transitions are labelled using states of the Markov chain, so that the automaton
recognises admissible Markov sequences. Combining Markov constraints with
other constraints, we can restrict the solution space in any desirable way [15],
but without any guarantee that the generated sequences will reflect the original
distribution in any way. In the specific case of unary constraints, we can have this
guarantee [2]. The result presented here can be seen as a further generalisation
of this result to a set of Regular constraints. A specific implementation of this
idea was used to generate non-plagiaristic sequences [18].

3 Background on Belief Propagation

Let X1 , . . . , Xn be n discrete random variables, and let p(X1 , . . . , Xn ) be a distri-

bution of the random sequence X1 , . . . , Xn . A graphical model [21] is a compact
representation of p as the product
m of m factors holding on a subset of the vari-
ables, i.e. p(X1 , . . . , Xn ) = j=1 fj (Sj ), where the factor fj is a function holding
on a subset Sj ⊆ {X1 , . . . , Xn } of the variables. CSPs can be seen as graphical
models, where solutions are uniformly distributed.
Belief propagation, specifically the sum-product algorithm [20] is an algo-
rithm for performing statistical inference, based on a factor graph representa-
tion. A factor graph is a bipartite undirected graph G = (X, F, E), representing
the factorisation of a probability function. Nodes represent either variables or
factors, and edges connect factors to the variables to which that factor applies:
X = {X1 , . . . , Xn }, F = {f1 , . . . , fm }, and an edge (Xi , fj ) is in E iff Xi ∈ Sj .

Example 1. Consider a probability function holding on three variables

X1 , X2 , X3 , defined as the product of four factors p(X1 , X2 , X3 ) = f1 (X1 , X2 ) ·
f2 (X2 , X3 ) · f3 (X1 , X3 ) · f4 (X3 ). The corresponding factor graph is shown on
Figure 1.

The main use of factor graph in statistical inference

 is to compute marginals.
Marginals are defined for each variable: pi (Xi ) = {Xj |j=i} p(X1 , . . . , Xn ). Once
marginals have been computed, sampling can be performed easily. When the
factor graph is a tree, computing marginals is polynomial. Tree factor graphs
correspond to Berge-acyclic constraint networks, and such results generalise the
well-known results in constraints [3,6,9].
344 A. Papadopoulos et al.

X1 f1 X2 f2 X3 f4
f3

Fig. 1. The factor graph for the function p(X1 , X2 , X3 ) = f1 (X1 , X2 ) · f2 (X2 , X3 ) ·
f3 (X1 , X3 ) · f4 (X3 ).

4 Belief Propagation for Markov and Regular

We apply those techniques to the problem of sampling constrained Markov

sequences, and describe belief propagation in the case where we impose sequences
to be recognised by an automaton A, i.e. to belong to the language L(A) of words
recognised by A. This is equivalent to sampling the target distribution ptarget
defined as:

⎧
⎨P (X2 |X1 ) · · · P (Xn |Xn−1 ) · if X1 · · · Xn ∈ L(A)
ptarget (X1 , . . . , Xn ) ∝ P (X ) · · · Pn (Xn )
⎩ 1 1
0 otherwise

We use the symbol ∝ to indicate that the equality holds after normalisation,
so that ptarget defines a probability function. P (X2 |X1 ) · · · P (Xn |Xn−1 ) gives
the typical order 1 Markov probability of the sequences X1 , . . . , Xn , provided
it is accepted by the automaton. Additionally, we add unary constraints Pi , i.e.
factors biasing each variable Xi individually. Implicitly, there is a big factor
holding on the full sequence X1 , . . . , Xn taking value 1 when X1 · · · Xn ∈ L(A),
and value 0 otherwise, corresponding to a hard global constraint. Consequently,
the factor graph of ptarget is not a tree.
We propose a reformulation of ptarget (X1 , . . . , Xn ) into a new function preg of
Y1 , . . . , Yn , where the new Yi variables take values (a, q) ∈ X ×Q, where a ∈ X is
a state of the Markov chain, and q ∈ Q is a state of the automaton. Recall that
transitions of the automaton are also labelled with elements of X . This function
preg is composed of simple binary factors, and its factor graph, which is tree
structured, is shown on Figure 2.

g1 g2 gn−1 gn

Y1 f1 Y2 Yn−1 fn−1 Yn

Fig. 2. The factor graph of the distribution on Markov sequences accepted by an

automaton A, defined by preg (Y1 , . . . , Yn )
 Exact Sampling for Regular and Markov Constraints with BP 345

We define a binary factor combining the Markov transition probabilities with

the valid transitions from the automaton, as follows:

  P (a |a), if q  = δ(q, a ),
f ((a, q), (a , q )) ∝
0 otherwise
This factor gives the probability for choosing, from state q, the transition
labelled with a , which reaches q  (denoted by q  = δ(q, a )). This probability
depends on the label a of the transition that was used to reach q, and is given
by the Markov transition probability from a to a . This factor is applied along
the sequence, i.e. fi = f, ∀1 ≤ i < n,. The binary factors imply that non-zero
probability sequences correspond to a walk in the automaton. Unary factors gi
additionally impose that such walks start from the initial state (enforced by g1 )
and end at an accepting state (enforced by gn ), while taking into account the
unary constraints of ptarget (enforced by all gi ):

 Pn (a), if q ∈ F
P1 (a), if q = δ(q0 , a) gn ((a, q)) ∝
g1 ((a, q)) ∝ 0, otherwise.
0, otherwise.

Other unary factors are simply defined as gi ((a, q)) ∝ Pi (a).

Theorem 1. Sampling ptarget is equivalent to sampling preg , and projecting each

resulting sequence (a1 , q1 ), . . . , (an , qn ) to a1 , . . . , an .

Proof. We prove there is a one-to-one correspondence between non-zero proba-

bility sequences of preg and ptarget , and that corresponding sequences have the
same probability.
Let
(a1 , q1 ), . . . , (an , qn ) be a sequence such that preg ((a1 , q1 ), . . . , (an , qn )) ≥ 0.
This means that q1 is the successor of the initial state q0 for a1 (from the def-
inition of g1 ), qi is the successor of state qi−1 for ai , for each i > 1 (from the
definition of f ), and qn is an accepting state (from the definition of gn ). In other
words, a1 , . . . , an is accepted by the automaton, and, according to the definitions
of the factors, with probability exactly equal to ptarget .
Conversely, suppose that a1 , . . . , an is a sequence with a non-zero ptarget
probability. Since A is deterministic, there exists a unique sequence of states
q0 , q1 , . . . , qn , with qn ∈ F , that recognises a1 , . . . , an , and therefore a unique
sequence (a1 , q1 ), . . . , (an , qn ) with a preg probability equal to ptarget (a1 , . . . , an ).

In order to sample sequences from this factor graph, we adapt the general
sum-product algorithm [21], and simplify it for the following reasons: the factor
graph has no cycle (removing any issue for converging to a fixed point), the
factor graph is almost a linear graph (induced by the sequence), factors are
only unary and binary, and the procedure is used only for sampling individual
sequences. This algorithm is shown on Algorithm 1 for self-containedness. It
computes the backward messages mi← , the forward messages mi→ , and the
346 A. Papadopoulos et al.

actual sequence y1 , . . . , yn , all highlighted in blue in the algorithm. The exact

justification of the algorithm is a well-established result [12,20], and we only
give an intuitive explanation. During the backward phase, mi← contains the
marginal of Yi of the product of all factors of preg holding on Yi , . . . , Yn . This
represents the impact on Yi of the sub-factor graph “to the right” of Yi , in the
same way that arc-consistency guarantees that a value can be extended to a full
instantiation. Eventually, m1← is the marginal of Y1 of all preg , and a value is
drawn randomly according to this distribution. The full sequence is generated
during the forward phase. At each iteration, pi (Yi ) is the marginal over Yi of
preg given the partial instantiation. In order to sample several sequences, the
backward phase needs to be performed only once, and the forward phase will
sample a new random sequence every time, with its correct probability. From a
constraint programming point of view, computing the marginals at each step is
a generalisation to random variables of computing arc-consistent domains. The
time for sampling one sequence is bounded by O(n · (|X ||Q|)2 ).

Algorithm 1. Sum-product algorithm for sampling Markov with Regular

Data: Function preg (Y1 , . . . , Yn ) and its factor graph
Result: A sequence y1 , . . . , yn , with probability preg (y1 , . . . , yn )
// Backward phase
mn← ← gn
for i ← n − 1 to 1 do
foreach y ∈ X × Q do
mi← (y) ← y ∈X ×Q gi (y) · fi (y, y  ) · mi+1← (y  )
Normalise mi←
// Forward phase
p1 ← m1←
y1 ← Draw with probability p1 (y1 )
for i ← 2 to n do
foreach y ∈ Q do
mi→ (y) ← fi−1 (yi−1 , y)
Normalise mi→
foreach y ∈ X × Q do pi (y) ← mi→ (y) · mi← (y)
yi ← Draw with probability pi (yi )
return (y1 , . . . , yn )

5 Examples
If no automaton is imposed, our model, which imposes only unary constraints,
is equivalent to the model in [16]. We compared the new model with our old
model, and observed it behaves equivalently, with the benefit of an improved
efficiency. We generated sequences of 16 notes with a Markov order 1 in the
style of Bill Evans, with two unary constraints constraining the first and last
 Exact Sampling for Regular and Markov Constraints with BP 347

note. Our old model could sample an average of 450 sequences per second, while
our new model produces an average of 1200 sequences per second, almost three
times more.
Automata can be handy for expressing patterns on text or music. For exam-
ple, in music, a semiotic structure is a symbolic description of higher level pat-
terns, from manual annotation or using pattern discovery techniques [4]. Semiotic
structures can be easily stated using automata. In text, automata can be used
to state syntactic rules over sequences of words.
Meter [26], in its most basic form, imposes that a sequence of variables
have a fixed total duration D, assuming each value has a specific duration, and
assuming the existence of a special padding value with a null duration, which
is used only at the end of the sequence. Meter can also be encoded using
Regular. We build an automaton Q, X , δ, q0 , F  where each state represents
a partial duration between 0 and D, i.e. Q = {q0 , . . . , qD }. For every element
e ∈ X of the Markov chain, we add a transition from qo1 to qo2 labelled by e iff
o2 = o1 + d(e), where d(e) is the duration of e. Finally, we set F = {qD }. This
ensures that any accepting sequence will have a total duration of D exactly.
By imposing this automaton to the Markov model, we can sample metrically
correct Markov sequences with their correct probabilities. We tested this with a
toy problem: produce sequences of a variable number of words, but with fixed
number of syllables equal to 36, i.e. the duration of a word is its number of
syllables. We are able to sample around in average 1100 sequences per second,
against 230 sequences per second produced by a CP model with a single Meter
constraint, almost five times more.
In previous work, we introduced MaxOrder, which limits the maximum
order of generated sequences, i.e. the length of exact copies made from the input
corpus [19]. This constraint was filtered by computing a particular automaton
and propagating it using Regular. We can use this automaton with the model
of this paper, in order to sample Markov sequences with a maximum order
guarantee. Furthermore, by computing the intersection between the Meter and
the MaxOrder automaton, we can also impose meter on such sequences.

6 Evaluation

We compare our fixed-length belief propagation model with a random walk in

the automaton. The purpose of this experiment is to show that a random walk
in the automaton does not sample sequences correctly, and confirm empirically
that our belief propagation-based model is correct, with a limited time penalty.
We used each method to sample sequences of words based on Pushkin’s
Eugene Onegin, of length 8, of Markov order 1 and with a max order less than 4,
imposed using a max order automaton [19]. We implemented our experiments in
Oracle Java 7, and ran them on an iMac with a 3.4GHz Intel Core i7 CPU, and
16GB RAM. The automaton was computed in about 200ms. For the random walk
method, we imposed the length by rejecting shorter sequences. In total, we sam-
pled over 20 million sequences. Of those, 5 million were unique sequences. The
348 A. Papadopoulos et al.

baseline random walk procedure generated an average of 5500 sequences per sec-
ond (counting only non-rejected sequences), while the belief propagation-based
method generated an average of 3500 sequences per second. For comparison, our
Regular-based CP model produced only about 50 sequences per second. We
filtered those that were generated over 50 times, of which there were about 47000
with random walk, and about 35000 with belief propagation. We estimated the
probability of a sequence by computing the sum of the probability of all unique
sequences found by either method, and use this for normalising.
We plot our results on Figure 3. Each point on either graph corresponds
to a sequence. Its value on the x-axis is its probability, estimated as described
previously, while the values on the y-axis is the empirical probability, i.e. the
frequency at which the specific sequence has been sampled compared to the
total number of sequences. Figure 3(a) shows that the baseline sampling app-
roach performs poorly: many sequences, even of similar probability, are over or
under-represented. On the other hand, Figure 3(b) provides a striking empirical
confirmation of the correctness of the belief propagation model.

0.0005 0.0005
probability (random walk automaton)

0.0004 0.0004
probability (belief propagation)

0.0003 0.0003

0.0002 0.0002

0.0001 0.0001

0 0
01 02 03 04 05 01 02 03 04 05
0 0.00 0.00 0.00 0.00 0.00 0 0.00 0.00 0.00 0.00 0.00
probability (estimated) probability (estimated)

(a) Random walk in the automaton (b) Belief propagation

Fig. 3. Sampling with random walk in the automaton compared to belief propagation.

7 Conclusion
We defined a belief propagation model for sampling Markov sequences that are
accepted by a given automaton. To this aim, we introduced a tree-structured
factor graph, on which belief propagation is polynomial and exact. This fac-
tor graph uses binary factors, which encode a type of conjunction between the
underlying Markov model and the given automaton. We showed that this pro-
cedure allows us to sample sequences faster than equivalent CP models, and
demonstrated that such sequences are sampled with their exact probabilities.
This result can be used for sequence generation problems in which users want
a set of solutions that are both probable in a given statistical model, and satisfy
hard regular constraints. More generally, we believe that this approach offers an
interesting bridge between statistical inference and constraint satisfaction.
 Exact Sampling for Regular and Markov Constraints with BP 349

Acknowledgments. This research is conducted within the Flow Machines project

funded by the ERC under the European Unions 7th Framework Programme (FP/2007-
2013) / ERC Grant Agreement n. 291156. We thank Ricardo de Aldama, funded by the
Lrn2Cre8 project (FET grant agreement no. 610859) for discussions about sampling
and CSP.

References
1. Proceedings, The Twenty-First National Conference on Artificial Intelligence and
the Eighteenth Innovative Applications of Artificial Intelligence Conference, July
16–20, 2006, Boston, Massachusetts, USA. AAAI Press (2006)
2. Barbieri, G., Pachet, F., Roy, P., Esposti, M.D.: Markov constraints for generating
lyrics with style. In: Raedt, L.D., Bessière, C., Dubois, D., Doherty, P., Frasconi, P.,
Heintz, F., Lucas, P.J.F. (eds.) Frontiers in Artificial Intelligence and Applications,
ECAI, vol. 242, pp. 115–120. IOS Press (2012)
3. Beeri, C., Fagin, R., Maier, D., Yannakakis, M.: On the Desirability of Acyclic
Database Schemes. J. ACM 30(3), 479–513 (1983)
4. Bimbot, F., Deruty, E., Sargent, G., Vincent, E.: Semiotic structure labeling of
music pieces: Concepts, methods and annotation conventions. In: Gouyon, F.,
Herrera, P., Martins, L.G., Müller, M. (eds.) Proceedings of the 13th Interna-
tional Society for Music Information Retrieval Conference, ISMIR 2012, Mosteiro
S.Bento Da Vitória, Porto, Portugal, October 8–12, 2012, pp. 235–240. FEUP
Edições (2012). http://ismir2012.ismir.net/event/papers/235-ismir-2012.pdf
5. Brooks, F.P., Hopkins, A., Neumann, P.G., Wright, W.: An experiment in musical
composition. IRE Transactions on Electronic Computers 6(3), 175–182 (1957)
6. Dechter, R., Pearl, J.: Tree Clustering for Constraint Networks. Artif. Intell. 38(3),
353–366 (1989)
7. Domingos, P.M., Kok, S., Poon, H., Richardson, M., Singla, P.: Unifying logical and
statistical AI. In: Proceedings, The Twenty-First National Conference on Artificial
Intelligence and the Eighteenth Innovative Applications of Artificial Intelligence
Conference, July 16–20, 2006, Boston, Massachusetts, USA [1], pp. 2–9. http://
www.aaai.org/Library/AAAI/2006/aaai06-001.php
8. Ermon, S., Gomes, C.P., Selman, B.: Uniform solution sampling using a
constraint solver as an oracle. In: de Freitas, N., Murphy, K.P. (eds.)
Proceedings of the Twenty-Eighth Conference on Uncertainty in Artificial
Intelligence, Catalina Island, CA, USA, August 14–18, 2012, pp. 255–264.
AUAI Press (2012). https://dslpitt.org/uai/displayArticleDetails.jsp?mmnu=1&
smnu=2&article id=2288&proceeding id=28
9. Freuder, E.C.: A Sufficient Condition for Backtrack-Free Search. J. ACM 29(1),
24–32 (1982)
10. Gogate, V., Dechter, R.: Studies in solution sampling. In: Fox, D., Gomes, C.P.
(eds.) Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence,
AAAI 2008, Chicago, Illinois, USA, July 13–17, 2008, pp. 271–276. AAAI Press
(2008). http://www.aaai.org/Library/AAAI/2008/aaai08-043.php
11. Hebrard, E., Hnich, B., O’Sullivan, B., Walsh, T.: Finding diverse and similar
solutions in constraint programming. In: Veloso, M.M., Kambhampati, S. (eds.)
Proceedings, The Twentieth National Conference on Artificial Intelligence and the
Seventeenth Innovative Applications of Artificial Intelligence Conference, July 9–
13, 2005, Pittsburgh, Pennsylvania, USA, pp. 372–377. AAAI Press / The MIT
Press (2005). http://www.aaai.org/Library/AAAI/2005/aaai05-059.php
350 A. Papadopoulos et al.

12. Mezard, M., Montanari, A.: Information, physics, and computation. Oxford
University Press (2009)
13. Nadel, A.: Generating diverse solutions in SAT. In: Sakallah, K.A., Simon, L. (eds.)
SAT 2011. LNCS, vol. 6695, pp. 287–301. Springer, Heidelberg (2011)
14. Nierhaus, G.: Algorithmic composition: paradigms of automated music generation.
Springer (2009)
15. Pachet, F., Roy, P.: Markov constraints: steerable generation of markov sequences.
Constraints 16(2), 148–172 (2011)
16. Pachet, F., Roy, P., Barbieri, G.: Finite-length markov processes with constraints.
In: Walsh, T. (ed.) IJCAI, pp. 635–642. IJCAI/AAAI (2011)
17. Pachet, F., Roy, P.: Imitative leadsheet generation with user constraints. In:
Schaub, T., Friedrich, G., O’Sullivan, B. (eds.) ECAI 2014–21st European Con-
ference on Artificial Intelligence, 18–22 August 2014, Prague, Czech Republic -
Including Prestigious Applications of Intelligent Systems (PAIS 2014). Frontiers in
Artificial Intelligence and Applications, vol. 263, pp. 1077–1078. IOS Press (2014).
http://dx.doi.org/10.3233/978-1-61499-419-0-1077
18. Papadopoulos, A., Pachet, F., Roy, P.: Generating non-plagiaristic markov
sequences with max order sampling. In: Degli Esposti, M., Altmann, E., Pachet,
F. (eds.) Universality and Creativity in Language (forthcoming). Lecture Notes in
Morphogenesis. Springer (2015)
19. Papadopoulos, A., Roy, P., Pachet, F.: Avoiding plagiarism in markov sequence
generation. In: Brodley, C.E., Stone, P. (eds.) Proceedings of the Twenty-Eighth
AAAI Conference on Artificial Intelligence, July 27–31, 2014, Québec City, Québec,
Canada, pp. 2731–2737. AAAI Press (2014). http://www.aaai.org/ocs/index.php/
AAAI/AAAI14/paper/view/8574
20. Pearl, J.: Reverend bayes on inference engines: A distributed hierarchical approach.
In: Waltz, D.L. (ed.) Proceedings of the National Conference on Artificial Intel-
ligence. Pittsburgh, PA, August 18–20, 1982, pp. 133–136. AAAI Press (1982).
http://www.aaai.org/Library/AAAI/1982/aaai82-032.php
21. Pearl, J.: Probabilistic reasoning in intelligent systems - networks of plausible infer-
ence. Morgan Kaufmann series in representation and reasoning. Morgan Kaufmann
(1989)
22. Pesant, G.: A regular language membership constraint for finite sequences of vari-
ables. In: Wallace, M. (ed.) CP 2004. LNCS, vol. 3258, pp. 482–495. Springer,
Heidelberg (2004)
23. Pinkerton, R.C.: Information theory and melody. Scientific American (1956)
24. Poon, H., Domingos, P.M.: Sound and efficient inference with probabilistic and
deterministic dependencies. In: Proceedings, The Twenty-First National Confer-
ence on Artificial Intelligence and the Eighteenth Innovative Applications of Arti-
ficial Intelligence Conference, July 16–20, 2006, Boston, Massachusetts, USA [1],
pp. 458–463. http://www.aaai.org/Library/AAAI/2006/aaai06-073.php
25. Richardson, M., Domingos, P.: Markov logic networks. Machine Learning 62(1–2),
107–136 (2006). http://dx.doi.org/10.1007/s10994-006-5833-1
26. Roy, P., Pachet, F.: Enforcing meter in finite-length markov sequences. In: des-
Jardins, M., Littman, M.L. (eds.) Proceedings of the Twenty-Seventh AAAI Con-
ference on Artificial Intelligence, July 14–18, 2013, Bellevue, Washington, USA.
AAAI Press (2013). http://www.aaai.org/ocs/index.php/AAAI/AAAI13/paper/
view/6422
27. Schreiber, Y.: Value-ordering heuristics: search performance vs. solution diversity.
In: Cohen, D. (ed.) CP 2010. LNCS, vol. 6308, pp. 429–444. Springer, Heidelberg
(2010)
 Randomness as a Constraint

Steven D. Prestwich1(B) , Roberto Rossi2 , and S. Armagan Tarim3

1
Insight Centre for Data Analytics, University College Cork, Cork, Ireland
s.prestwitch@cs.ucc.ie
2
University of Edinburgh Business School, Edinburgh, UK
3
Department of Management, Cankaya University, Ankara, Turkey

Abstract. Some optimisation problems require a random-looking solu-

tion with no apparent patterns, for reasons of fairness, anonymity,
undetectability or unpredictability. Randomised search is not a good
general approach because problem constraints and objective functions
may lead to solutions that are far from random. We propose a constraint-
based approach to finding pseudo-random solutions, inspired by the Kol-
mogorov complexity definition of randomness and by data compression
methods. Our “entropy constraints” can be implemented in constraint
programming systems using well-known global constraints. We apply
them to a problem from experimental psychology and to a factory inspec-
tion problem.

1 Introduction
For some applications we require a list of numbers, or some other data struc-
ture, that is (or appears to be) random, while also satisfying certain constraints.
Examples include the design of randomised experiments to avoid statistical bias
[13], the generation of random phylogenetic trees [14], quasirandom (low dis-
crepancy) sequences for efficient numerical integration and global optimisation
[24], randomised lists without repetition for use in experimental psychology [10],
random programs for compiler verification [6], and the random scheduling of
inspections for the sake of unpredictability [30].
An obvious approach to obtaining a random-looking solution is simply to
use a randomised search strategy, such as stochastic local search or backtrack
search with a randomised value ordering. In some cases this works, for example
it can generate a random permutation of a list, but in general there are several
drawbacks with the randomised search approach:
– If only random-looking solutions are acceptable then the constraint model
is not correct, as it permits solutions that are unacceptable. The correctness
of a constraint model should be independent of the search strategy used to
solve it.
– Randomised search can not prove that a random-looking solution does not
exist, or prove that a solution is as random-looking as possible.
– Unless the randomised search is carefully designed (see [9] for example) it is
likely to make a biased choice of solution.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 351–366, 2015.
DOI: 10.1007/978-3-319-23219-5 25
352 S.D. Prestwich et al.

– Even if we sample solutions in an unbiased way, there is no guarantee that

such a solution will look random. Although randomly sampled unconstrained
sequences are almost certain to appear random (almost all long sequences
have high algorithmic entropy [2]) a constrained problem might have mostly
regular-looking solutions. Similarly, optimising an objective function might
lead to regular-looking solutions. In Section 3.2 we give examples of both
phenomena.
Instead it would be useful to have available a constraint israndom(v) that forces
a vector v of variables to be random, which could simply be added to a constraint
model. First, however, we must define what we mean by random.
In information theory, randomness is a property of the data source used to
generate a data sequence, not of a single sequence. The Shannon entropy of
the source can be computed from its symbol probabilities using Shannon’s well-
known formula [32]. But in algorithmic information theory, randomness can be
viewed as a property of a specific data sequence, and its Kolmogorov complexity,
or algorithmic entropy, is defined as the length of the smallest algorithm that can
describe it. For example the sequence 1111111111 may have the same probability
of occurring as 1010110001 but it has lower algorithmic entropy because it can be
described more simply (write 1 ten times). Algorithmic entropy formally captures
the intuitive notion of whether a list of numbers “looks random”, making it useful
for our purposes. We shall refer to algorithmic entropy simply as entropy by a
slight abuse of language.
Having chosen (algorithmic) entropy as our measure of randomness, we would
like to have available a constraint of the form entropy(v, e) to ensure that the
entropy of v is at least e. Unfortunately, defining such a constraint is impossible
because algorithmic entropy is uncomputable [5]. Instead we take a pragmatic
approach by defining constraints that eliminate patterns exploited by well-known
data compression algorithms, which can be combined as needed for specific appli-
cations. There is a close relationship between algorithmic entropy and com-
pressibility: applying a compression algorithm to a sequence of numbers, and
measuring the length of the compressed sequence, gives an upper bound on
the algorithmic entropy of the original sequence. Thus by excluding readily-
compressible solutions we hope to exclude low-entropy (non-random) solutions.
The paper is organised as follows. Section 2 presents constraint-based
approaches to limiting the search to high-entropy solutions. Section 3 applies
these ideas to two problems. Section 4 discusses related work. Section 5 con-
cludes the paper and discusses future work.

2 Entropy Constraints
We require constraints to exclude low-entropy solutions, which we shall call
entropy constraints. This raises several practical questions: what types of pattern
can be excluded by constraints, how the constraints can be implemented, and
what filtering algorithms are available? We address these problems below. In our
experiments we use the Eclipse constraint logic programming system [1].
 Randomness as a Constraint 353

2.1 Non-uniform Distributions

Data in which some symbols occur more often than others have non-uniform
probability distributions. Huffman coding [15] and arithmetic coding [29] are
compression methods that exploit this feature by encoding symbols with a vari-
able number of bits (a prefix code with many bits for rare symbols and few bits
for common symbols).
To eliminate this type of pattern from our solutions we define some sim-
ple entropy constraints. We assume that all variables have the same domain
{0, . . . , m − 1} but it is easy to generalise to different domains. Given lower
bounds  = 0 , . . . , m−1  and upper bounds u = u0 , . . . , um−1  on the
frequencies of symbols 0, . . . , m − 1 we define a frequency entropy constraint
freq(v, m, , u). It can be directly implemented by a global cardinality con-
straint [27] GCC(v, 0, . . . , m − 1, , u).
To see the effect of these constraints, consider a CSP with 100 finite domain
variables v0 , . . . , v99 ∈ {0, 1, . . . , 9} and no problem constraints. We use the
default backtrack search heuristic to find the lex-least solution which we take
as a proxy for the lowest-entropy solution. The lex-least solution is the solution
that is least under the obvious lexicographical ordering (order by increasing value
of the first variable, breaking ties using the second variable and so on, under the
static variable ordering v1 , v2 , . . .). If even the lex-least solution has high entropy
then we believe that other solutions will too. To test this we applied randomised
search to a problem with entropy constraints, and never obtained a solution that
could be compressed further than the lex-least solution.
If we add freq(v, 10, 0, 14)1 to this problem we obtain the lex-least solution
00000000000000111111111111112222222222222233333333
33333344444444444444555555555555556666666666666677
We shall estimate the entropy of this lex-least solution by compressing it using
the well-known gzip2 compression algorithm. Though gzip does not necessarily
give an accurate entropy estimate it has been successfully used for this purpose
(see for example [4]). Denoting the entropy of solutions by  and measuring it
in bytes, the lex-least solution has  = 43. For a random sequence of digits in
the range 0–9 we typically obtain  values of 80–83 so the lex-least solution
is far from random, as can be observed. With no entropy constraints the lex-
least solution is simply 0 repeated 100 times, which gzip compresses from 100
bytes to 26 bytes. There is no paradox in compressing these random numbers:
integers in the range 0–9 can be represented by fewer than 8 bits (1 byte). Note
that gzip compresses an empty text file to 22 bytes, so most of the 26 bytes is
decompression metadata.

2.2 Repeated Strings

For the compression of discrete data probably the best-known methods are based
on adaptive dictionaries. These underlie compression algorithms such as Linux
1
We use i with integer i to denote a vector i, . . . , i of appropriate length.
2
http://www.gzip.org/
354 S.D. Prestwich et al.

compress, V.24 bis, GIF, PKZip, Zip, LHarc, PNG, gzip and ARJ [31] and use
algorithms described in [34–36]. These methods detect repeated k-grams (blocks
of k adjacent symbols) and replace them by pointers to dictionary entries. For
example the string 011101011100 contains two occurrences of the 5-gram 01110,
which can be stored in a dictionary and both its occurrences replaced by a
pointer to that entry.
We could generalise the freq constraint to limit the number of occurrences
of every possible k-gram, but as there are mk of them this is impractical unless
k is small. A more scalable approach is as follows. Given an integer k ≥ 2 and
an upper bound t on the number of occurrences of all k-grams over symbols
{0, . . . , m − 1} in a vector v of variables, we define a constraint dict(v, m, k, t).
We shall call this a dictionary entropy constraint. It can be implemented by
the Multi-Inter-Distance(x, t, p) global constraint [20] with p = n − k + 1
k−1
(the number of x variables), x = x0 , . . . , xn−k , and the xi = j=0 mj vi+j are
auxiliary integer variables representing k-grams, where n is the sequence length.
This global constraint enforces an upper bound t on the number of occurrences
of each value within any p consecutive x variables. In the special case t = 1 we
can instead use alldifferent(x) [28].
To test this idea on the artificial problem of Section 2.1 we add
dict(v, 10, k, 1) dictionary constraints for various k-values (but not the fre-
quency constraints of Section 2.1). The results in Table 1 show that as we reduce
k the solution contains fewer obvious patterns, and k = 2 gives a solution that
(to gzip) is indistinguishable from a purely random sequence, as 80 bytes is
within the range of  for random sequences of this form.

Table 1. Lex-least solutions with dictionary entropy constraints

k lex-least solution 
50 00000000000000000000000000000000000000000000000001
00000000000000000000000000000000000000000000000000 29
25 00000000000000000000000001000000000000000000000000
20000000000000000000000003000000000000000000000000 35
12 00000000000010000000000020000000000030000000000040
00000000005000000000006000000000007000000000008000 48
6 00000010000020000030000040000050000060000070000080
00009000011000012000013000014000015000016000017000 60
3 00010020030040050060070080090110120130140150160170
18019021022023024025026027028029031032033034035036 71
2 00102030405060708091121314151617181922324252627282
93343536373839445464748495565758596676869778798890 80

In this example there are sufficient digrams (2-grams) to avoid any repetition,
but for smaller m this will not be true, and in the general case we might need
to use larger t or larger k. Note that we must choose tk ≥ n/mk  otherwise the
problem will be unsatisfiable.
 Randomness as a Constraint 355

2.3 Correlated Sources

Though gzip and related compression algorithms often do a very good job, they
are not designed to detect all patterns. A solution with no repeated k-grams
might nevertheless have low entropy and noticeable patterns. For example the
following sequence of 100 integers in the range 0–9 compresses to 80 bytes, and
is therefore indistinguishable from a random sequence to gzip:
01234567890246813579036914725804815926370516273849
94837261507362951840852741963097531864209876543210
Yet it was written by hand following a simple pattern and is certainly not a
random sequence, as becomes apparent if we examine the differences between
adjacent symbols:
1 1 1 1 1 1 1 1 1 -9 2 2 2 2 -7 2 2 2 2 -9 3 3 3 -8 3 3 -5 3 3 -8 4 4 -7 4 4 -7 4
-3 4 -7 5 -4 5 -4 5 -4 5 -4 5 0 -5 4 -5 4 -5 4 -5 4 -5 7 -4 3 -4 7 -4 -4 7 -4 -4 8
-3 -3 5 -3 -3 8 -3 -3 -3 9 -2 -2 -2 -2 7 -2 -2 -2 -2 9 -1 -1 -1 -1 -1 -1 -1 -1 -1
The same differences often occur together but gzip is not designed to detect this
type of pattern. As another example, the high-entropy (k = 2) solution found in
Section 2.2 has differences
jjkilhmgnfoepdqcrbsbjjkilhmgnfoepdqcrcjjkilhmgnfoe
pdqdjjkilhmgnfoepejjkilhmgnfofjjkilhmgngjjkilhmhj

where differences −9 . . . + 9 are represented by symbols a–s. The differences also

look quite random: gzip compresses this list of symbols to 92 bytes which is
typical of a random sequence of 99 symbols from a–s. Yet they have a non-
uniform distribution: a does not occur at all while j occurs 15 times.
In data compression an example of a correlated source of data is one in
which each symbol depends probabilistically on its predecessor. This pattern is
exploited in speech compression methods such as DPCM [7] and its variants.
Another application is in lossless image compression, where it is likely that some
regions of an image contain similar pixel values. This is exploited in the JPEG
lossless compression standard [33], which predicts the value of a pixel by con-
sidering the values of its neighbours. Greater compression can sometimes be
achieved by compressing the differences between adjacent samples instead of the
samples themselves. This is called differential encoding.
(1)
We can confound differential compressors by defining new variables vi =
vi − vi+1 + m − 1 with domains {0, . . . , 2(m − 1)} to represent the differences
between adjacent solution variables (shifted to obtain non-negative values), and
(1)
applying entropy constraints from Sections 2.1 and 2.2 to the vi . We shall call
these differential [frequency, dictionary] entropy constraints. We use the notation
v (1) because later on we shall consider differences of differences v (2) and so on.
Adding a differential frequency constraint freq(v (1) , 18, 0, 10) and a differ-
ential dictionary constraint dict(v (1) , 18, 3, 1) to the earlier constraints we get
differences
jkilhmgnfoepdqcrbsbjkjilikhmhlgngmfofnepeodqdpcrcq
djkkhliimgoenfpejlhnfmhjklglgofmiiingjmhkhkjljijk
356 S.D. Prestwich et al.

which has  = 90, and lex-least solution

00102030405060708091122132314241525162617271828192
93345354363837394464847556857495876596697867799889
which has  = 80: both  values indicate totally random sequences of the respec-
tive symbol sets. However, in some ways this solution still does not look very
random: its initial values are 0, and roughly the first third of the sequence has a
rather regular pattern. This is caused by our taking the lex-least solution, and
by there being no problem constraints to complicate matters. In such cases we
could use a SPREAD-style constraint [22] to prevent too many small values from
occurring at the start of the sequence, or perhaps resort to randomised search.
Note that on this artificial example randomised search usually finds high-entropy
solutions even without entropy constraints, but it is not guaranteed to do so.

2.4 Using Entropy Constraints

By expressing the randomness condition as constraints we ensure that all solu-
tions are incompressible by construction. Therefore the search method used to
find the sequences does not matter and we can use any convenient and efficient
search algorithm, such as backtrack search (pruned by constraint programming
or mathematical programming methods) or metaheuristics (such as tabu search,
simulated annealing or a genetic algorithm). As long as the method is able to
find a solution it does not matter if the search is biased, unless we require sev-
eral evenly distributed solutions. But in the latter case we could define a new
problem P  whose solution is a set of solutions to the original problem P, with
constraints ensuring that the P-solutions are sufficiently distinct.
All our entropy constraints are of only two types: freq and dict. Both can
be implemented via well-known Constraint Programming global constraints, or
in integer linear programs via reified binary variables, or in SAT via suitable
clauses. (In our experiments of Section 3 we use the second method because
our constraint solver does not provide the necessary global constraints.) We can
relate them by a few properties (proofs omitted):
dict(v, m, k, t) ⇒ dict(v, m, k + 1, t)
dict(v (i) , m, k, t) ⇒ dict(v (i−1) , m, k + 1, t)
freq(v (i) , m, 0, t) ⇒ dict(v (i−1) , m, 2, t)
From these we can deduce
freq(v (i) , m, 0, t) ⇒ dict(v, m, i + 1, t)
which provides an alternative way of limiting k-gram occurrences. But higher-
order differential constraints should be used with caution. For example we could
use freq(v (2) , m, 0, 1) instead of dict(v, m, 3, 1) as both prevent the trigram
125 from occurring more than once. But the former is stronger as it also pre-
vents trigrams 125 and 668 from both occurring: the trigrams have the order-1
differences (1,3) and (0,2) respectively, hence the same order-2 difference (2).
If we do not consider 125 and 668 to be similar in any relevant way for our
application then using the freq constraint is unnecessarily restrictive.
 Randomness as a Constraint 357

3 Applications
We consider two applications. Section 3.1 describes a known problem from exper-
imental psychology, and Section 3.2 describes an artificial factory inspection
problem where the inspection schedule must be unpredictable.

3.1 Experimental Psychology

Experimental psychologists often need to generate randomised lists under con-
straints [10]. An example of such an application is word segmentation studies
with a continuous speech stream. The problem discussed in [10] has a multiset W
of 45 As, 45 Bs, 90 Cs and 90 Ds, and no two adjacent symbols can be identical.
There is an additional constraint: that the number of CD digrams must be equal
to the number of As. From this multiset must be generated a randomised list.
Generating such a list was long thought to be a simple task and a standard
list randomisation algorithm was used: randomly draw an item from W and add
it to the list, unless it is identical to the previous list item in which case replace it
and randomly draw another item; halt when W is empty. But it was shown in [10]
that this algorithm can lead to a large bias when the frequencies of the items
are different, as in the problem considered. The bias is that the less-frequent
items A and B appear too often early in the list, and not often enough later
in the list. The bias effect is particularly bad for short lists generated by such
randomisation-without-replacement methods. The bias can ruin the results of
an experiment by confounding frequency effects with primacy or recency effects.
The solution proposed by [10] is to create transition frequency and transition
probability tables, and to use these tables to guide sequence generation (we do
not give details here). This is therefore a solved problem, but the authors state
that the correct ... table corresponding to the constraints of a given problem can
be notoriously hard to construct, and it would be harder to extend their method
to problems with more complex constraints.
Generating such lists is quite easy using our method. For the above example
we create a CSP with 270 variables vi each with domain {0, 1, 2, 3} represent-
ing A, B, C and D respectively. To ensure the correct ratio of items we use
frequency constraints freq(v, 4, 45, 45, 90, 90, 45, 45, 90, 90). We add a con-
straint to ensure that there are 45 CD digrams. To ensure a reasonably even
spread of values we add constraints freq(v i , 4, 0, 11, 11, 22, 22) to each fifth
v 1 , . . . , v 5 of v. Finally we add constraints dict(v i , 4, 5, 1) to each fifth.
Backtrack search turned out to be very slow so we use a local search algo-
rithm, implemented in the constraint solver to take advantage of its filtering
algorithms. The algorithm is not intended to be particularly competitive or
interesting, but we describe it here for the sake of reproducibility. In an ini-
tial iteration it attempts to assign a value to each variable in turn using the
ordering v1 , v2 , . . ., and halts when domain wipeout occurs. We refer to this ini-
tial choice of value for each variable the default. In the next iteration it makes a
few random assignment choices, but in most cases it choose the default. If any
choice leads to domain wipeout then all alternatives are tried for that variable
358 S.D. Prestwich et al.

(but backtracking to other variables does not occur). If the new set of choices
assign no fewer variables than the defaults then they become the new defaults. If
all variables are assigned in an iteration then the algorithm has found a feasible
solution and it halts.
The solution we found after a few tens of seconds is shown in Figure 1, with
 = 118. In further tests local search achieved  in the range 114–124, whereas
without entropy constraints it achieved  of 100–115.
CBDCDCADCBDBCBCABCDCBDCDBCABADCBCDCABDCBCADCDCDCACADAD
CDADCBDABDCDCBCDCDCDABCABDCDBCDBADCABDBDCACDCACDBCACDC
ADCDCACDCBDACDCADCDCBDCABCDBDBDACACDCDCDBDCACDBCDCADBD
CDCDADCACDACDCDCDCBDCDBDBDBDCDCACDBDBCDBCADCADADCDCADC
BCDBCDCDACACBCDCBCDBDCDCDCBCADCDCADCBDADCADACABDADBDAD

Fig. 1. A solution to the experimental psychology problem

[10] noted a systematic bias in the A:C ratio between fifths of the sequence:
approximately 0.59, 0.58, 0.53, 0.49, 0.32. In ours we have 0.47, 0.50, 0.47, 0.44
and 0.59. Our approach has much in common with the Mix system used by
psychologists but also important differences (see the discussion in Section 4).

3.2 Factory Inspection

Next we consider an artificial factory inspection problem. Suppose an inspector

must schedule 20 factory inspections over 200 days. We represent this by finite
domain variables v1 , . . . , v20 ∈ {1, . . . , 200}. The inspection plan should appear
as random as possible so that the owners can not predict when they will be
inspected.
To make the problem more realistic we could add constraints preventing
different factories from being inspected on certain days, or introduce an objective
function. We do this below, but first we shall simply restrict all inspections to
certain available days which are unevenly distributed through the year: 1–40,
70–100, 130–150 and 180–190. We order the vi in strictly ascending order using
a global constraint ordered(v).
The lex-least solution is
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

with differences
1111111111111111111

All the inspections take place in the first 20 days, as is clearer if we visualise the
schedule using two rows of 100 days each, the dark colour indicating scheduled
inspections, the light colour indicating days in which an inspection could be
scheduled, and white indicating days in which no inspection could be scheduled:
 Randomness as a Constraint 359

We could simply randomise the value ordering in each variable assignment during
backtrack search, but this is unlikely to find the highest-entropy solutions. A
typical solution found in this way is
73 139 142 144 146 147 148 149 150 180
181 182 183 184 185 186 187 188 189 190

with differences
66 3 2 2 1 1 1 1 30 1 1 1 1 1 1 1 1 1 1

and the schedule:

This solution also looks non-random. Because the variables are ordered, choos-
ing each value randomly in turn causes clustering in the high values. We could
randomise the search in a more clever way, for example by biasing earlier assign-
ments to lower values, or branching under a different variable ordering. For
such a simple problem this would not be hard to do, but the more complicated
the constraint network is the harder this task becomes. Stochastic local search
might typically find a high-entropy solution but, as pointed out in Section 1,
randomised search alone is not enough so we shall apply entropy constraints.
As the vi are all different we use difference variables. Applying entropy con-
straints freq(v (1) , 198, 0, 2) and freq(v (2) , 395, 0, 1) gives the lex-least solution
1 2 3 5 9 11 16 20 28 31 39 70 73 82 87 99 130 136 150 180

and the schedule

This is much more random-looking, indicating that an appropriate use of entropy

constraints can yield higher-entropy solutions than random search, as mentioned
in Section 1.
To confirm our intuition that this solution is sufficiently random we apply
gzip. However, we do not compress the solutions in integer form as it contains
spaces between numbers and uses multiple characters to represent integers: this
helps to hide patterns from gzip so that its entropy estimate is poor. Instead we
compress the binary schedule representations (as used above) of:

(i) the lex-least solution without entropy constraints

(ii) a randomised backtrack search solution
(iii) the lex-least solution with entropy constraints
(iv) mean results for random binary sequences of length 200 containing exactly
20 ones, without the restriction to certain ranges of days
We also estimate their Approximate Entropy (ApEn), a measure of the regularity
and predictability of a sequence of numbers. ApEn was originally developed to
360 S.D. Prestwich et al.

analyse medical data [23] but has since found many applications. We use the
definition given in [2] for a sequence n symbols from an alphabet of size m:

H(k) − H(k − 1) if k > 1
ApEn(k) =
H(1) if k = 1

where
k

m
H(k) = − pi log2 pi
i=1

and pi is the probability of k-gram i occurring in the sequence, estimated as

its observed frequency fi divided by n − k + 1. H(1) is simply the Shannon
entropy measured in bits, and ApEn(k) is a measure of the entropy of a block of
k symbols conditional on knowing the preceding block of k − 1 symbols. ApEn is
useful for estimating the regularity of sequences, it can be applied to quite short
sequences, and it often suffices to check up to ApEn(3) to detect regularity [2].
Results are shown in Table 2. The compression and ApEn results for (iii) are
almost as good as those of (iv). This indicates not only that the inspections are
unpredictable by the factory owners, but that the owners can not even detect in
hindsight (by gzip and ApEn) the fact that the inspector had time constraints.
Note that 0.47 is the theoretical ApEn(k) for all k ≥ 1, for a binary source with
probabilities 0.1 and 0.9 as in this example.

Table 2. Inspection schedule entropies

ApEn(k)
solution  k=1k=2k=3
(i) 29 0.47 0.03 0.03
(ii) 39 0.47 0.28 0.24
(iii) 54 0.47 0.46 0.42
(iv) 56 0.47 0.46 0.45

We apply a further statistical test of randomness to the binary representation:

the Wald-Wolfowitz runs test. A run is an unbroken substring of 0s (or 1s) with
preceding and subsequent 1s (or 0s) except at the start and end of the string. For
example the string 01100010 contains five runs: 0, 11, 000, 1 and 0. A randomly
chosen string is unlikely to have a very low or very high number of runs, and
this can be used as a test of randomness. For a random binary sequence with
180 zeroes and 20 ones, we can be 95% confident that the sequence is random
if there are between 33 and 41 runs. The lex-least solution has 36 runs so it
passes the test. Thus the lex-least solution passes several tests of randomness.
(One might ask why we did not also generate constraints to enforce ApEn and
the runs test: we discuss this point in Section 5.)
This example is perhaps too simple as it contains only one constraint
and no objective function. In fact if we replace the ordered constraint by
 Randomness as a Constraint 361

an alldifferent constraint then local search without entropy constraints

finds a high-entropy solution with  = 55, ApEn(1)=0.47, ApEn(2)=0.45 and
ApEn(3)=0.45 — though backtrack search with randomised value ordering
only achieves  = 48, ApEn(1)=0.47, ApEn(2)=0.46 and ApEn(3)=0.35. We
should not have to manipulate the constraint model or choose the right solver
to find high-entropy solutions, and there might be a good reason for using
ordered instead of alldifferent (for example it breaks permutation symme-
try), but to make the problem more interesting we now consider two alternative
complications.
Firstly, suppose that we have some additional constraints:

– no two inspections may occur on consecutive days: strengthen the ordered

constraint to vi+1 > vi + 1 (i = 1 . . . 19)
– there must be at least 13 inspections in the first 100 days: v13 ≤ 100
– there must be at least 2 inspections in the fourth time period: v19 ≥ 180

Without entropy constraints local search now finds a lower-entropy solution with
 = 44, ApEn(1)=0.47, ApEn(2)=0.46 and ApEn(3)=0.31, but with entropy
constraints it finds this schedule:

with high entropy:  = 57, ApEn(1)=0.49, ApEn(2)=0.47 and ApEn(3)=0.47.

Secondly, instead of additional constraints suppose we have an objective func-
tion. Consider a scenario in which the factories are all in country A, the inspector
lives in a distant country B, and flights between A and B are expensive. We might
aim to minimise the number of flights while still preserving the illusion of ran-
domness. To do this we could maximise the number of inspections on consecutive
days. If we require an objective value of at least 10 then we are unable to find a
solution under the above entropy constraints. We are forced to use less restrictive
entropy constraints such as freq(v (1) , 198, 0, 10) and dict(v (1) , m, 2, 1) yielding
the following schedule by local search:

with 10 nights in a hotel,  = 50, ApEn(1)=0.47, ApEn(2)=0.38 and

ApEn(3)=0.32. This illustrates the fact that imposing additional criteria might
force us to accept lower-entropy solutions.

4 Related Work

[30] proposed statistical constraints to enforce certain types of randomness on a

solution: solutions should pass statistical tests such as the t-test or Kolmogorov-
Smirnov. An inspection plan found in using the latter test is shown in Figure 2.
If this is the only test that might be applied by an observer then we are done.
However, the schedule exhibits a visible pattern that could be used to predict
362 S.D. Prestwich et al.

the next inspection with reasonable certainty. The pattern is caused by the
deterministic search strategy used. It might be important to find an inspection
schedule that is not predictable by visual inspection, or by a machine learning
algorithm. In this case statistical tests are not enough and we must also enforce
randomness.

Fig. 2. A partially-random inspection plan that passes a statistical test

In the field of hardware verification SAT solvers have been induced to gener-
ate random stimuli: see for example [16] for a survey of methods such as adding
randomly-chosen XOR constraints. Constraint solvers have also been used for
the same purpose, a recent example being [19]. These approaches aim to gener-
ate an unbiased set of solutions, as do the methods of [6,8,9,11,12], whereas we
aim to maximise the algorithmic entropy of a single solution. But (as pointed
out in Section 2.4) we could obtain a similar result by defining a new problem P 
whose solution is a set of solutions to the original problem P, and add entropy
constraints to P  .
[3] describes a software system called Mix for generating constrained ran-
domised number sequences. It implements a hand-coded local search algorithm
with several types of constraint that are useful for psychologists, including con-
straints that are very similar to our freq and dict constraints (maximum repe-
tition and pattern constraints respectively). However, no connection is made to
Kolmogorov complexity or data compression, Mix does not use a generic con-
straint solver or metaheuristic, it does not use differential constraints (though
it has other constraints we do not have), and it is designed for a special class of
problem.
The SPREAD constraint [22] has something in common with our frequency
constraints but with a different motivation. It balances distributions of values,
for example spreading the load between periods in a timetable. It has efficient
filtering algorithms but it does not aim to pass compression-based randomness
tests.
Markov Constraints [21] express the Markov condition as constraints, so that
constraint solvers can generate Markovian sequences. They have been applied to
the generation of musical chord sequences.

5 Discussion
We proposed several types of entropy constraint to eliminate different types
of pattern in a solution, leading to high-entropy solutions as estimated by
 Randomness as a Constraint 363

compression algorithms and the Approximate Entropy function. These are com-
plementary to statistical tests of the kind explored in [30]. All our constraints are
based on well-known global constraints and can also be implemented in MIP or
SAT. Note that instead of specifying bounds on the occurrences of symbols and
k-grams we could allow the user to specify bounds on the Approximate Entropy
ApEn(k) for various k. However, we believe that the former approach is more
intuitive.
Using constraints to represent randomness makes it easy to generate random-
looking solutions with special properties: we simply post constraints for random-
ness and for the desired properties, then any solution is guaranteed to satisfy
both. However, applying entropy constraints is something of an art involving a
compromise between achieving high entropy, satisfying the problems constraints
and possibly optimising an objective function. Even with few or no problem
constraints we must take care not to exclude so many patterns that no solutions
remain, as Ramsey theory [25] shows that any sufficiently large object must
contain some structure. In fact adding entropy constraints does not necessarily
preserve satisfiability. If a problem has no sufficiently random-looking solutions
then entropy constraints might eliminate all solutions. However, an advantage
of this is that (as mentioned in Section 1) we can prove that no such solutions
exist: this cannot be done with the randomised search approach. Alternatively
we could take an optimisation approach by treating entropy constraints as soft
constraints, and searching for the most random-looking solution.
Of course our solutions are only pseudorandom, not truly random. They were
generated by restricting repeated symbols and k-grams in order to be incompress-
ible to a certain class of data compression algorithms. It could be objected that
they might fail other tests of randomness that are important to applications. Our
response to this argument is: we can turn these other tests into additional con-
straints. For example if our solution in Section 3.2 had failed the Wald-Wolfowitz
runs test, we could have added a constraint to ensure that it passed the test, as
follows. Suppose we have a sequence of n binary numbers, with n0 zeroes and
n1 ones (n0 + n1 = n). Under a normal approximation (valid for n0 , n1 ≥ 10)
the expected number of runs is
2n0 n1
μ=1+
n
and the variance of this number is
2n0 n1 (2n0 n1 − n) (μ − 1)(μ − 2)
σ2 = =
n (n − 1)
2 n−1

To test for randomness with 95% confidence we require that the observed number
of runs R is within μ ± 1.96σ. To implement this test as a constraint on binary
variables v1 , . . . , vn we define new binary variables bi = reify(vi = vi+1 ) and post
a constraint  

n−2
μ − 1.96σ ≤ 1 + bi ≤ μ + 1.96σ
i=0
364 S.D. Prestwich et al.

If we do not know the values of n0 and n1 in advance, the constraint implemen-

tation can create auxiliary integer variables n0 , n1 and real variables μ, σ, and
post additional constraints:

vi = n1 n0 + n1 = n
μn = n + n0 n1 σ 2 (n − 1) = (μ − 1)(μ − 2)

There is another possible objection to our approach — in fact to the whole

idea of eliminating patterns in solutions. It can be argued that a solution with a
visible pattern is statistically no less likely to occur than a solution with no such
pattern, and that patterns are merely psychological artefacts. For example if we
generate random binary sequences of length 6 then 111111 is no less random
than 010110 because both have the same probability of occurring. Considering
the latter to be “more random” than the former is a form of Gambler’s Fallacy,
in which (for example) gamblers assume that numbers with obvious patterns are
less likely to occur than random-looking numbers. But if we wish to convince
humans (or automated software agents designed by humans) that a solution was
randomly generated then we must reject patterns that appear non-random to
humans. This intuition is made concrete by the ideas of algorithmic information
theory [5]. We do not expect all readers to agree with our view: randomness is a
notoriously slippery concept [2] whose nature is beyond the scope of this paper.
There are several interesting directions for future work. We hope to find new
applications, possibly with other patterns to be eliminated. We could try to
devise randomness constraints using ideas from other literatures. One approach
is to take randomness tests applied to number sequences and turn them into
constraints. For example we might implement spectral tests [17] which are con-
sidered to be powerful. But they are complex and we conjecture that they are
unlikely to lead to efficient filtering (though this is far from certain and would
be an interesting research direction). Moreover, they seem better suited to very
long sequences of numbers: far longer than the size of typical solutions to opti-
misation problems. For evaluating pseudo-random number generators there is a
well-known set of simpler tests: the Die-Hard Tests,3 later extended to the Die
Harder Tests 4 and distilled down to three tests in [18]. However, these tests are
also aimed at very long sequences of numbers, and again it is not obvious how
to derive constraints from them.

Acknowledgement. This publication has emanated from research supported in part

by a research grant from Science Foundation Ireland (SFI) under Grant Number
SFI/12/RC/2289. S. Armagan Tarim is supported by the Scientific and Technologi-
cal Research Council of Turkey (TUBITAK).

3
http://www.stat.fsu.edu/pub/diehard/
4
http://www.phy.duke.edu/∼rgb/General/dieharder.php
 Randomness as a Constraint 365

References
1. Apt, K.R., Wallace, M.: Constraint Logic Programming Using Eclipse. Cambridge
University Press (2007)
2. Beltrami, E.: What Is Random? Chance and Order in Mathematics and Life.
Copernicus (1999)
3. van Casteren, M., Davis, M.H.: Mix, a Program for Pseudorandomization.
Behaviour Research Methods 38(4), 584–589 (2006)
4. Cilibrasi, R., Vitányi, P.M.B.: Clustering by Compression. IEEE Transactions on
Information Theory 51(4), 1523–1545 (2005)
5. Chaitin, G.J.: Algorithmic Information Theory. Cambridge University Press (1987)
6. Claessen, K., Duregård, J., Palka, M.H.: Generating constrained random data with
uniform distribution. In: Codish, M., Sumii, E. (eds.) FLOPS 2014. LNCS, vol.
8475, pp. 18–34. Springer, Heidelberg (2014)
7. Cutler, C.C.: Differential Quantization for Television Signals. U. S. Patent 2, 605,
361, July 1952
8. Dechter, R., Kask, K., Bin, E., Emek, R.: Generating random solutions for con-
straint satisfaction problems. In: Proceedings of the 18th National Conference on
Artificial Intelligence, pp. 15–21 (2002)
9. Ermon, S., Gomes, C.P., Selman, B.: Uniform solution sampling using a constraint
solver as an oracle. In: Proceedings of the Twenty-Eighth Conference on Uncer-
tainty in Artificial Intelligence, pp. 255–264. AUAI Press (2012)
10. French, R.M., Perruchet, P.: Generating Constrained Randomized Sequences: Item
Frequency Matters. Behaviour Research Methods 41(4), 1233–1241 (2009)
11. Hebrard, E., Hnich, B., O’Sullivan, B., Walsh, T.: Finding diverse and similar solu-
tions in constraint programming. In: Proceedings of the 20th National Conference
on Artificial Intelligence (2005)
12. Van Hentenryck, P., Coffrin, C., Gutkovich, B.: Constraint-based local search for
the automatic generation of architectural tests. In: Gent, I.P. (ed.) CP 2009. LNCS,
vol. 5732, pp. 787–801. Springer, Heidelberg (2009)
13. Hinkelmann, K., Kempthorne, O.: Design and Analysis of Experiments I and II.
Wiley (2008)
14. Housworth, E.A., Martins, E.P.: Random Sampling of Constrained Phylogenies:
Conducting Phylogenetic Analyses When the Philogeny is Partially Known. Syst.
Biol. 50(5), 628–639 (2001)
15. Huffman, D.A.: A Method for the Construction of Minimum Redundancy Codes.
Proceedings of the IRE 40, 1098–1101 (1951)
16. Kitchen, N., Kuehlmann, A.: Stimulus generation for constrained random sim-
ulation. In: Proceedings of the 2007 IEEE/ACM International Conference on
Computer-Aided Design, pp. 258–265. IEEE Press (2007)
17. Knuth, D.E.: The Art of Computer Programming. Seminumerical Algorithms, 2nd
edn., vol. 2, p. 89. Addison-Wesley (1981)
18. Marsaglia, G., Tsang, W.W.: Some Difficult-to-pass Tests of Randomness. Journal
of Statistical Software 7(3) (2002)
19. Naveh, R., Metodi, A.: Beyond feasibility: CP usage in constrained-random func-
tional hardware verification. In: Schulte, C. (ed.) CP 2013. LNCS, vol. 8124,
pp. 823–831. Springer, Heidelberg (2013)
20. Ouellet, P., Quimper, C.-G.: The multi-inter-distance constraint. In: Proceedings
of the 22nd International Joint Conference on Artificial Intelligence, pp. 629–634
(2011)
366 S.D. Prestwich et al.

21. Pachet, F., Roy, P.: Markov Constraints: Steerable Generation of Markov
Sequences. Constraints 16(2), 148–172 (2011)
22. Pesant, G., Régin, J.-C.: SPREAD: a balancing constraint based on statistics. In:
van Beek, P. (ed.) CP 2005. LNCS, vol. 3709, pp. 460–474. Springer, Heidelberg
(2005)
23. Pincus, S.M., Gladstone, I.M., Ehrenkranz, R.A.: A Regularity Statistic for Medical
Data Analysis. Journal of Clinical Monitoring and Computing 7(4), 335–345 (1991)
24. Press, W.H., Flannery, B.P., Teukolsky, S.A., Vetterling, W.T.: Numerical Recipes
in C, 2nd edn. Cambridge University Press, UK (1992)
25. Ramsey, F.P.: On a Problem of Formal Logic. Proceedings London Mathematical
Society s2 30(1), 264–286 (1930)
26. Refalo, P.: Impact-based search strategies for constraint programming. In: Wallace,
M. (ed.) CP 2004. LNCS, vol. 3258, pp. 557–571. Springer, Heidelberg (2004)
27. Régin, J.-C.: Generalized Arc Consistency for Global Cardinality Constraint.
In: 14th National Conference on Artificial Intelligence, pp. 209–215 (1996)
28. Régin, J.-C.: A filtering algorithm for constraints of difference in CSPs.
In: Proceedings of the 12th National Conference on Artificial Intelligence,
Vol. 1, pp. 362–367 (1994)
29. Rissanen, J.J., Langdon, G.G.: Arithmetic Coding. IBM Journal of Research and
Development 23(2), 149–162 (1979)
30. Rossi, R., Prestwich, S., Tarim, S.A.: Statistical constraints. In: 21st European
Conference on Artificial Intelligence (2014)
31. Sayood, K.: Introduction to Data Compression. Morgan Kaufmann (2012)
32. Shannon, C.E.: A Mathematical Theory of Communication. Bell System Technical
Journal 27(3), 379–423 (1948)
33. Wallace, G.K.: The JPEG Still Picture Compression Standard. Communications
of the ACM 34, 31–44 (1991)
34. Welch, T.A.: A Technique for High-Performance Data Compression. IEEE Com-
puter, 8–19, June 1984
35. Ziv, J., Lempel, A.: A Universal Algorithm for Data Compression. IEEE Transac-
tions on Information Theory IT 23(3), 337–343 (1977)
36. Ziv, J., Lempel, A.: Compression of Individual Sequences via Variable-Rate Coding.
IEEE Transactions on Information Theory IT 24(5), 530–536 (1978)
 Quasipolynomial Simulation of DNNF
by a Non-determinstic Read-Once
Branching Program

Igor Razgon(B)

Department of Computer Science and Information Systems,

Birkbeck, University of London, London, U.K.
igor@dcs.bbk.ac.uk

Abstract. We prove that dnnfs can be simulated by Non-deterministic

Read-Once Branching Programs (nrobps) of quasi-polynomial size. As
a result, all the exponential lower bounds for nrobps immediately apply
for dnnfs.

1 Introduction
Decomposable Negation Normal Forms (dnnfs) [3] is a well known formalism in
the area of propositional knowledge compilation notable for its efficient represen-
tation of cnfs with bounded structural parameters. The dnnfs lower bounds
are much less understood. For example, it has been only recently shown that
dnnfs can be exponentially large on (monotone 2-) cnfs [2]. Prior to that, it
was known that on monotone functions dnnfs are not better than monotone
dnnfs [6]. Hence all the lower bounds for monotone circuits apply for dnnfs.
However, using monotone circuits to obtain new dnnf lower bounds is hardly an
appropriate methodology because, in light of [2], on monotone functions, dnnfs
are much weaker than monotone circuits.
In this paper we show that dnnfs are strongly related to Non-deterministic
Read-Once Branching Programs (nrobps) that can be thought as Free Binary
Decision Diagrams (fbdds) with or-nodes. In particular, we show that a dnnf
can be transformed into a nrobp with a quasi-polynomial increase of size. That
is, all the exponential lower bounds known for nrobps (see e.g. [5,8]) apply for
dnnfs. As nrobps can be linearly simulated by dnnfs (using a modification
of the simulation of fbdds by dnnfs proposed in [4]), we believe that the pro-
posed result makes a significant progress in our understanding of complexity of
dnnfs. Indeed, instead of trying to establish exponential lower bounds directly
for dnnfs, we can now do this for nrobps , which are much better understood
from the lower bound perspective.
In the proposed simulation, we adapt to unrestricted dnnfs the approach
that was used in [1] for quasi-polynomial simulation of decision dnnfs by fbdds.
For the adaptation, we find it convenient to represent nrobps in a form where
variables carry no labels and edges are labelled with literals. In particular, each
input node u of the dnnf is represented in the resulting nrobp as an edge

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 367–375, 2015.
DOI: 10.1007/978-3-319-23219-5 26
368 I. Razgon

labelled with the literal of u and these are the only edges that are labelled (com-
pare with [1] where the labelling is ‘pertained’ to or nodes, which is impossible
for unrestricted dnnfs where the or nodes can have an arbitrary structure).
The most non-trivial aspect of the simulation is the need to transform an
and of two nrobps Z1 and Z2 into a single nrobp. Following [1], this is done by
putting Z1 ‘on top’ of Z2 . However, this creates the problem that Z1 becomes
unusable ‘outside’ this construction (see Section 4.1. and, in particular, Figure
2 of [1] for illustration of this phenomenon). Similarly to [1], we address this
problem by introducing multiple copies of Z1 .
Formal Statement of the Result. A dnnf Z ∗ is a directed acyclic graph
(dag) with many roots (nodes of in-degree 0) called input nodes and one leaf
(node of out-degree 0) called the output node. The input nodes are labelled
with literals, the rest are and, and or nodes such that each and node has the
decomposability property defined as follows. Let us define V ar(u) for a node u of
Z ∗ as the set of variables x such that Z ∗ has a path from a node labelled by x to
u. Then, if u is an and node of Z ∗ and v and w are two different in-neighbours
of u then V ar(v) ∩ V ar(w) = ∅. Let Zu∗ be the subgraph of Z ∗ induced by a
node u and all the nodes from which u can be reached. Then the function F [Zu∗ ]
computed by Zu∗ is defined as follows. If u is an input node then F [Zu∗ ] = x,
where x is the literal labelling u. If u is an or or an and node with in-neighbours
v1 , . . . vq then F [Zu∗ ] = F [Zv∗1 ] ∨ · · · ∨ F [Zv∗q ], or F [Zu∗ ] = F [Zv∗1 ] ∧ · · · ∧ F [Zv∗q ],
respectively. The function F [Z ∗ ] computed by Z ∗ is F [Zout ∗
], where out is the
∗
output node of Z . In the rest of the paper we assume that the and nodes of
Z ∗ are binary. This assumption does not restrict generality of the result since
an arbitrary dnnf can be transformed into one with binary and nodes with a
quadratic increase of size.
A Non-deterministic Read-once Branching Program (nrobp) is a dag Z with
one root (and possibly many leaves). Some edges of Z are labelled with literals of
variables in the way that each variable occurs at most once on each path P of Z.
We denote by A(P ) the set of literals labelling the edges of a path P of Z. To define
a function F [Z] computed by Z, let us make a few notational agreements. First, we
define a truth assignment to a set of variables as the set of literals of these variables
that become true as result of the assignment. For example, the assignment {x1 ←
true, x2 ← f alse, x3 ← true} is represented as {x1 , ¬x2 , x3 }. For a function F on a
set V ar of variables, we say that an assignment S of V ar satisfies F is F (S) = true.
Now, let S be an assignment of variables labelling the edges of Z. Then S satisfies
F [Z] if and only if there is a root-leaf path P of Z with A(P ) ⊆ S. A dnnf and a
nrobp for the same function are illustrated on Figure 1.
Remark. The above definition of nrobp is equivalent to fbdd with or-nodes in
the sense that each of them can simulate the other with a linear increase of size.
Our main result proved in the next section is the following.

Theorem 1. Let Z ∗ be a dnnf with m nodes computing a function F of n

variables. Then F can be computed by a nrobp of size O(mlog n+2 ).
 Quasipolynomial Simulation of DNNF by a NROBP 369

& & &

X1 X1 ~X1

~X1 ~X2 ~X3

X2 ~X2 X2
X1 X2 X3

X3 X3
~X3

Fig. 1. dnnf and nrobp for function (x1 ∧ x2 ∧ ¬x3 ) ∨ (x1 ∧ ¬x2 ∧ x3 ) ∨ (¬x1 ∧ x2 ∧ x3 )

2 Proof of Theorem 1

This section is organised as follows. We first present a transformation of a dnnf

Z ∗ into a graph Z, then state two auxiliary lemmas about properties of special
subgraphs of Z, their proofs postponed to Section 2.1, and then prove Theorem 1.
The first step of the transformation is to fix one in-coming edge of each and-
node u of Z ∗ as the light edge of u. This is done as follows. Let u1 and u2 be
two in-neighbours of u such that |V ar(u1 )| ≤ |V ar(u2 )|. Then (u1 , u) is the light
edge of u and (u2 , u) is the heavy edge of u. (Of course if both in-neighbours
of u depend on the same number of variables then u1 and u2 can be chosen
arbitrarily.) We say that an edge (v, w) of Z ∗ is a light edge if w is an and-node
and (v, w) is its light edge. Let P be a path from u to the output node out of
Z ∗ . Denote the set of light edges of P by le(P ). Denote by LE(u) the set of all
such le(P ) for a u − out path P .
Now we define a graph Z consisting of the following nodes.

– (u, le, in) for all u ∈ V (Z ∗ ) (recall that if G is a graph, V (G) denotes the
set of nodes of G) and le ∈ LE(u). The ‘in’ in the third coordinate stands
for ‘internal’ to distinguish from the ‘leaf’ nodes defined in the next item.
– For each input node u of Z ∗ and for each le ∈ LE(u), Z has a node (u, le, lf )
where ‘lf’ stands for ‘leaf’. We say that (u, le, lf ) is the leaf corresponding to
(u, le, in).

When we refer to a node of Z with a single letter, we use bold letters like u, v
to distinguish from nodes u, v of Z ∗ . We denote by mnode(u), coord(u), type(u),
the respective components of u, that is u = (mnode(u), coord(u), type(u)). We
also call the components the main node of u, the coordinate of u and the type of
u. The nodes of Z whose type is in are internal nodes and the nodes whose type
is lf are leaf nodes. The leaf nodes are not necessarily leaves of Z but rather
leaves of special subgraphs of Z that are important for the proof.
370 I. Razgon

Setting the Environment for Definition of Edges of Z. We explore the

nodes u of Z ∗ topologically sorted from the input to the output and process
each internal node u of Z with u = mnode(u). In particular, we introduce out-
neighbours of u, possibly, together with labelling of respective edges, the set of
nodes Leaves(u), and a subgraph Graph(u) of Z which will play a special role
in the proof. The detailed description of processing of u is provided below.

– Suppose that u is an input node. Let y be the literal labelling u in Z ∗ and

let u be the leaf corresponding to u.
1. Introduce an edge (u, u ) and label this edge with y.
2. Set Leaves(u) = {u }.
3. Define Graph(u) as having node set {u, u } and the edge (u, u ).
– Suppose that u is an or node. Let v1 , . . . vq be the in-neighbours of u in Z ∗ .
Let v1 , . . . , vq be the internal nodes of Z with v1 , . . . , vq being the respective
main nodes and with coord(vi ) = coord(u) for all 1 ≤ i ≤ q.
1. Introduce edges (u, v1 ), . . . , (u, vq ).
2. Set Leaves(u) = Leaves(v1 ) ∪ · · · ∪ Leaves(vq ).
3. Graph(u) is obtained from Graph(v1 ) ∪ · · · ∪ Graph(vq ) by adding node
u plus the edges (u, v1 ), . . . , (u, vq ).
– Suppose u is an and node. Let u1 , u2 be two in-neighbours of u in Z ∗ and
assume that the edge (u1 , u) is the light one. Let u1 , u2 be two internal
nodes of Z whose respective main nodes are u1 and u2 and coord(u1 ) =
coord(u) ∪ {(u1 , u)} and coord(u2 ) = coord(u).
1. Introduce edges (u, u1 ) and (w, u2 ) for each w ∈ Leaves(u1 ).
2. Set Leaves(u) = Leaves(u2 ).
3. Graph(u) is obtained from Graph(u1 ) ∪ Graph(u2 ) by adding node u
and the edges described in the first item.

Remark. Let us convince ourselves that the nodes v1 , . . . , vq , and u1 , u2 with

the specified coordinates indeed exist. Indeed, suppose that u is an or-node of
Z ∗ and let v be an in-neighbour of u. Let P be a path from u to the output
node of Z ∗ . Then le((v, u) + P ) = le(P ) confirming possibility of choice of nodes
v1 , . . . , vq . Suppose that u is an and-node and let (u1 , u) and (u2 , u) be the
light and heavy edges of u respectively. For a P as before, le((u1 , u) + P ) =
{(u1 , u)} ∪ le(P ) and le((u2 , u) + P )) = le(P ) confirming that the nodes u1 and
u2 exist. Thus the proposed processing is well-defined.

Lemma 1. Let u ∈ V (Z) with type(u) = in and let u = mnode(u). Then the
following statements hold.

1. Graph(u) is a dag.
2. u is the (only) root of Graph(u) and Leaves(u) is the set of leaves of
Graph(u).
3. If u is an or-node and v1 , . . . , vq are as in the description of processing of
u then each root-leaf path P of Graph(u) is of the form (u, vi ) + P  where
P  is a root-leaf path of Graph(vi ).
 Quasipolynomial Simulation of DNNF by a NROBP 371

4. Suppose u is an and node and let u1 , u2 be as in the description of processing

of u. Then each root-leaf path P of Graph(u) is of the form (u, u1 ) + P1 +
(w, u2 ) + P2 , where P1 , P2 are root-leaf paths of Graph(u1 ) and Graph(u2 ),
respectively and w is the last node of P1 .
5. V ar(u) = V ar(u) where V ar(u) is the set of all variables labelling the edges
of Graph(u).
6. Graph(u) is read-once (each variable occurs at most once on each path).

It follows from the first, second, and the last statements of Lemma 1 that
Graph(u) is a nrobp. Therefore, we can consider the function F [Graph(u)]
computed by Graph(u).

Lemma 2. For each u ∈ V (Z) with type(u) = in, F [Graph(u)] = F [Zu∗ ] where
u = mnode(u) and Zu∗ is as defined in the first paragraph of formal statement
part of the introduction.

Proof of Theorem 1. Let out be the output node of Z ∗ and out = (out, ∅, in)
be a node of Z. Graph(out) is a nrobp by Lemma 1 and, by Lemma 2, it
∗ ∗
computes function F (Zout ). By definition, Zout = Z ∗ and hence Graph(out)
∗
computes the same function as Z .
To upper-bound the size of Z ∗ , observe that for each u ∈ V (Z), |coord(u)| ≤
log n. Indeed, let us represent coord(u) as (u1 , u1 ), . . . , (uq , uq ), a sequence of
edges occurring in this order on a path of Z ∗ . Then each (ui , ui ) is the light
edge of an and-node ui . By the decomposability property of dnnf, |V ar(ui )| ≤
|V ar(ui )|/2. Also, since Z ∗ has a path from ui to ui+1 , |V ar(ui )| ≤ |V ar(ui+1 )|.
Applying this reasoning inductively, we conclude that |V ar(u1 )| ≤ |V ar(uq )|/2q .
Since |V ar(u1 )| ≥ 1 and |V ar(uq )| ≤ n, it follows that |coord(u)| = q ≤ log n.
Thus coord(u) is a set of light edges of Z ∗ of size at most log n. Since there
is at most one light edge per element of Z ∗ , there are at most m light edges
in total. Thus the number of possible second coordinates for a node of Z is
log n m
i ≤ m
log n+1
i=1 . As the number of distinct first and third coordinates is
at most m and 2, respectively, the result follows. 

2.1 Proofs of Auxiliary Lemmas for Theorem 1

Proof of Lemmas 1 and 2 requires two more auxiliary lemmas.

Lemma 3. Let u ∈ V (Z) with type(u) = in and let u = mnode(u). Then for
each v ∈ V (Graph(u)), coord(u) ⊆ coord(v). Moreover, if type(v) = lf then
coord(u) = coord(v) if and only if v ∈ Leaves(u).

Proof. By induction on nodes u of Z according to the topological sorting of

the nodes u = mnode(u) of Z ∗ from input to output nodes. That is if v is an
neighbour of u then for any node v with v = mnode(v) the lemma holds by the
induction assumption.
If u is an input node then V (Graph(u)) consists of u and the leaf cor-
responding to u, hence the first statement holds by construction. Otherwise,
372 I. Razgon

V (Graph(u)) consists of u itself and the union of all V (Graph(v)), where, fol-
lowing the description of processing of u, v is one of v1 , . . . , vq if u is an or-node
and v is either u1 or u2 if u is an and-node. For each such v it holds by defini-
tion that coord(u) ⊆ coord(v). That is, each node w = u of Graph(u) is in fact
a node of such a Graph(v). By the induction assumption, coord(v) ⊆ coord(w)
and hence coord(u) ⊆ coord(w) as required.
Using the same inductive reasoning, we show that for each w ∈ Leaves(u),
coord(w) = coord(u). This is true by construction if u is an input node.
Otherwise, Leaves(u) is defined as the union of one or more Leaves(v) such
that coord(v) = coord(u) by construction. Then, letting v be such that
w ∈ Leaves(v) we deduce, by the induction assumption that coord(w) =
coord(v) = coord(u).
It remains to prove that for each w ∈ V (Graph(u)) \ Leaves(u) such that
type(w) = lf , coord(u) ⊂ coord(w). This is vacuously true if u is an input node.
Otherwise, the induction assumption can be straightforwardly applied as above
if w ∈ Graph(v) for some v as above and w ∈ / Leaves(v). The only situation
where it is not the case is when u is an and node and v = u1 where u1 is as in
the description of processing of an and node. In this case coord(u) ⊂ coord(u1 )
by construction and, since coord(u1 ) ⊆ coord(w), by the first statement of this
lemma, coord(u) ⊂ coord(w). 

Lemma 4. 1. For each internal u ∈ V (Z), the out-going edges of u in Z are

exactly those that have been introduced during processing of u.
2. For each v ∈ V (Z) with type(v) = lf , the out-degree of v in Z is at
most 1. Moreover the out-degree of v is 0 in each Graph(u ) such that
v ∈ Leaves(u ).
3. Let u be an internal node and let (w1 , w2 ) be an edge where w1 ∈
V (Graph(u)) \ Leaves(u). Then (w1 , w2 ) is an edge of Graph(u).

Proof. The first statement follows by a direct inspection of the processing algo-
rithm. Indeed, the only case where an edge (u, v) might be introduced during
processing of a node u = u is where mnode(u ) is an and node. However, in
this case type(u) must be lf in contradiction to our assumption.
Consider the second statement. Consider an edge (v, w) such that type(v) =
lf . Suppose this edge has been created during processing of a node u. Then
u = mnode(u) is an and-node. Further, let u1 , u2 be as in the description of
processing of u. Then v ∈ Leaves(u1 ) and w = u2 . By construction, coord(w) =
coord(u) and by Lemma 3, coord(v) = coord(u1 ). Hence, by definition of u1 ,
coord(w) ⊂ coord(v). Suppose that v ∈ Leaves(u ) for some internal u ∈ V (Z).
Then, by Lemma 3, coord(v) = coord(u ) and hence coord(w) ⊂ coord(u ). It
follows from Lemma 3 that w is not a node of Graph(u ) and hence (v, w) is
not an edge of Graph(u ). Thus the out-degree of v in Graph(u ) is 0.
Continuing the reasoning about edge (v, w), we observe that coord(w) =
coord(v) \ {(u1 , u)} where u1 = mnode(u1 ). Notice that all the edges of
coord(v) = coord(u1 ) lie on a path from u1 to the output node of Z ∗ and
(u1 , u) occurs first of them. Due to the acyclicity of Z ∗ , (u1 , u) is uniquely
 Quasipolynomial Simulation of DNNF by a NROBP 373

defined (in terms of v) and hence so is coord(w). Furthermore, as specified

above, mnode(w) = mnode(u2 ) which is the neighbour u other than u1 . Since
(u1 , u) is uniquely defined, u and u1 are uniquely defined as its head and tail and
hence so is mnode(u2 ). Finally, by construction, we know that w is an internal
node. Thus all three components of w are uniquely defined and hence so is w
itself. That is, v can have at most one neighbour.
The third statement is proved by induction analogous to Lemma 3. The
statement is clearly true if u = mnode(u) is an input node of Z ∗ because then
Graph(u) has only one edge. Assume this is not so. If w1 = u then the statement
immediately follows from the first statement of this lemma and the definition
of Graph(u). Otherwise, w1 ∈ V (Graph(v)) where v is as defined in the proof
of Lemma 3. If w1 ∈ / Leaves(v) then, by the induction assumption, (w1 , w2 ) is
an edge of Graph(v) and hence of Graph(u). This may be not the case only if
u is an and node and w1 ∈ Leaves(u1 ) (u1 and u2 are as in the description
of processing of u). By definition of Graph(u), (w1 , u2 ) is an edge of Graph(u)
and, according to the previous paragraph, w1 does not have other outgoing
edges. Thus it remains to assume that w2 = u2 and the statement holds by
construction. 
Proof Sketch of Lemma 1 All the statements except 3 and 4 are proved by
induction like in Lemma 3. If u = mnode(u) is an input node then Graph(u) is
a labelled edge for which all the statements of this lemma are clearly true. So,
we assume below that u is either an or-node or an and-node.
Statement 1. Assume that Graph(u) does have a directed cycle C. Since
Graph(u) is loopless by construction, C contains at least one node w = u.
By construction, w is a node of some Graph(v) where v is as defined in the
proof of Lemma 3. Then C intersects with Leaves(v). Indeed, if we assume the
opposite then, applying the last statement of Lemma 4 inductively starting from
w, we conclude that all the edges of C belong to Graph(v) in contradiction
to its acyclicity by the induction assumption. Now, C does not intersect with
Leaves(u) because they have out-degree 0 by Lemma 4. Thus if C intersects
with Graph(v) then Leaves(v) cannot be a subset of Leaves(u). This is only
possible if u is an and-node and v = u1 (u1 and u2 are as in the description of
the processing of u). Let w ∈ Leaves(u1 ) be a node of C. By construction, u2 is
an out-neighbour of w and by Lemma 4, w does not have other out-neighbours
in Z. Thus u2 is the successor of w in C and hence C intersects with Graph(u2 )
while Leaves(u2 ) ⊆ Leaves(u) in contradiction to what we have just proved.
Thus C does not exists.
Statement 2. It is easy to verify by induction that Graph(u) has a path from u
to the rest of vertices, hence besides u, Graph(u) does not have any other roots.
Now u itself is a root by the acyclicity proved above. Since vertices of Leaves(u)
have out-degree 0 in Graph(u), clearly, they are all leaves. Suppose that some
w ∈ Graph(u) \ Leaves(u) is a leaf of u. By construction, w = u and hence
w is a node of some Graph(v) as above. Then w is a leaf of Graph(v) because
the latter is a subgraph of Graph(u) and hence, by the induction assumption,
374 I. Razgon

w ∈ Leaves(v). Hence, as in the previous paragraph, we need v such that

Leaves(v)  Leaves(u) and we conclude as above that v = u1 . But then u2
is an out-neighbour of w and hence w cannot be a leaf of u, a contradiction
showing that the leaves of Graph(u) are precisely Leaves(u).
Important Remark. In the rest of the section we use u and the root of
Graph(u) as well as Leaves(u) and the leaves of Graph(u) interchangeably
without explicit reference to statement 2 of this lemma.
Statements 3 and 4. Suppose that u is an or node and let P  be the suffix of P
starting at the second node v of P . By statement 1 of Lemma 4, v is some vi as
in the description of processing of u. Hence vertices of Leaves(v) ⊆ Leaves(u)
have out-degree 0 in Graph(u) (statement 2 of Lemma 4) and do not occur in
the middle of P  . Applying statement 3 of Lemma 4 inductively to P  starting
from v, we observe that P  is a path of Graph(v). The last node of P  is a leaf
of Graph(v) because it is a leaf of Graph(u). Thus statement 3 holds.
Suppose that u is an and-node. Then by statement 1 of Lemma 4, the second
node of P is u1 . Hence, one of Leaves(u1 ) must be an intermediate node of P .
Indeed, otherwise, by inductive application of statement 3 of Lemma 4, the suffix
of P starting at u1 is a path of Graph(u1 ). In particular, the last node w of P is
a node of Graph(u1 ). However, by Lemma 3, coord(w ) = coord(u) ⊂ coord(u1 ).
Hence, by Lemma 3, coord(w ) cannot belong to Graph(u1 ), a contradiction.
Let w ∈ Leaves(u1 ) be the first such node of P . By inductive application of the
last statement fo Lemma 4, the subpath P1 of P between u1 and w is a root-leaf
path of Graph(u1 ). By construction and the second statement of Lemma 4, u2
is the successor of w in P . Let P2 be the suffix of P starting at u2 . Arguing as
for the or case we conclude that P2 is a root-leaf path of u2 . Thus statement 4
holds.
Statement 5. We apply induction, taking into account that V ar(u) is the union
of all V ar(v) where v is an in-neighbour of and V ar(u) is the union of all V ar(v)
where v is as defined in the proof of Lemma 3. The details are omitted due to
space constraints.
Statement 6. Let P be a root-leaf path of Graph(u). Suppose u = mnode(u) is
an or-node. Then P = (u, vi )+P  , the notation as in statement 3. P  is read-once
by the induction assumption and the edge (u, vi ) is unlabelled by construction.
Hence P is read-once. If u is an and-node then P = (u, u1 ) + P1 + (w, u2 ) + P2 ,
all the notation as in statement 4. P1 and P2 are read-once by the induction
assumption, edges (u, u1 ) and (w, u2 ) are unlabelled. The variables of P1 and
of P2 are respective subsets of V ar(u1 ) and V ar(u2 ) equal to V ar(u1 ) and
V ar(u2 ), respectively, by statement 5 which, in turn, do not intersect due to the
decomposability property of and nodes of Z ∗ . Hence the variables of P1 and P2
do not intersect and P is read-once. 
Proof of Lemma 2. By induction as in Lemma 3. If u = mnode(u) is an input
node then F [Zu∗ ] = x where x is the literal labelling u and Graph(u) is a single
edge labelled with x, hence F [Graph(u)] = x.
 Quasipolynomial Simulation of DNNF by a NROBP 375

In the rest of the proof, we assume that the set of variables of all the con-
sidered functions is V ar, the set of variables of Z ∗ . Introduction of redundant
variables will simplify the reasoning because we can now make the induction
step without modifying the considered set of variables.
Assume that u is an or node. If S satisfies F [Zu∗ ] then there is an in-neighbour
v of u such that S satisfies F [Zv∗ ]. By construction there is an out-neighbour
v of u such that v = mnode(v). By the induction assumption, S satisfies
F [Graph(v)]. Let P  be a v − Leaves(v) path such that A(P  ) ⊆ S. Then,
by construction (u, v) + P  is a u − Leaves(u) path with the edge (u, v) unla-
belled. Hence A(P ) = A(P  ) ⊆ S and hence S satisfies F [Graph(u)]. Conversely,
if S satisfies F [Graph(u)] then there is a u − Leaves(u) path P with A(P ) ⊆ S.
By statement 3 of Lemma 1, P = (u, vi ) + P  , the notation as in the state-
ment. A(P  ) ⊆ A(P ) and hence S satisfies F [Graph(vi )] and, by the induction
assumption, S satisfies F [Zv∗i ] where vi = mnode(vi ). By definition of vi , vi is
an in-neighbour of u, hence S satisfies F [Zu∗ ].
Assume that u is an and node. Let u1 , u2 be as in the description of process-
ing of u with u1 = mnode(u1 ) and u2 = mnode(u2 ). Suppose that S satisfies
F [Zu∗ ]. Then S satisfies both F [Zu∗1 ] and F [Zu∗2 ]. Hence, by the induction assump-
tion S satisfies F [Graph(u1 )] and F [Graph(u2 )]. For each i ∈ {1, 2}, let Pi be
a ui − Leaves(ui ) path of Graph(ui ) with A(Pi ) ⊆ S. Let w be the last node
of P1 . Then P = (u, u1 ) + P1 + (w, u2 ) + P2 is a u = Leaves(u) path with the
edges (u, u1 ) and (w, u2 ) unlabelled. Hence A(P ) = A(P1 ) ∪ A(P2 ) ⊆ S and
thus S satisfies F [Graph(u)]. Conversely, suppose that S satisfies F [Graph(u)]
and let P be a u − Leaves(u) path with A(P ) ⊆ S. Then by statement 4 of
Lemma 1, P = (u, u1 ) + P1 + (w, u2 ) + P2 , the notation as in the statement.
Clearly, A(P1 ) ⊆ S and A(P2 ) ⊆ S, hence S satisfies both F [Zu∗1 ] and F [Zu∗2 ]
by the induction assumption and thus S satisfies F [Zu∗ ]. 

References
1. Beame, P., Li, J., Roy, S., Suciu, D.: Lower bounds for exact model counting and
applications in probabilistic databases. In: Proceedings of the Twenty-Ninth Con-
ference on Uncertainty in Artificial Intelligence, Bellevue, WA, USA, August 11–15,
2013 (2013)
2. Bova, S., Capelli, F., Mengel, S., Slivovsky, F.: Expander cnfs have exponential
DNNF size. CoRR, abs/1411.1995 (2014)
3. Darwiche, A.: Decomposable negation normal form. J. ACM 48(4), 608–647 (2001)
4. Darwiche, A., Marquis, P.: A knowledge compilation map. J. Artif. Intell. Res.
(JAIR) 17, 229–264 (2002)
5. Jukna, S.: Boolean Function Complexity: Advances and Frontiers. Springer-Verlag
(2012)
6. Krieger, M.P.: On the incompressibility of monotone DNFs. Theory Comput. Syst.
41(2), 211–231 (2007)
7. Oztok, U., Darwiche, A.: On compiling CNF into decision-DNNF. In: O’Sullivan,
B. (ed.) CP 2014. LNCS, vol. 8656, pp. 42–57. Springer, Heidelberg (2014)
8. Wegener, I.: Branching Programs and Binary Decision Diagrams. SIAM (2000)
 MiniSearch: A Solver-Independent Meta-Search
Language for MiniZinc

Andrea Rendl1(B) , Tias Guns2 , Peter J. Stuckey3 , and Guido Tack1

1
National ICT Australia (NICTA) and Faculty of IT, Monash University,
Melbourne, Australia
andrea.rendl@nicta.com.au, guido.tack@monash.edu
2
KU Leuven, Leuven, Belgium
tias.guns@cs.kuleuven.be
3
National ICT Australia (NICTA) and University of Melbourne, Victoria, Australia
pstuckey@unimelb.edu.au

Abstract. Much of the power of CP comes from the ability to create

complex hybrid search algorithms specific to an application. Unfortu-
nately there is no widely accepted standard for specifying search, and
each solver typically requires detailed knowledge in order to build com-
plex searches. This makes the barrier to entry for exploring different
search methods quite high. Furthermore, search is a core part of the
solver and usually highly optimised. Any imposition on the solver writer
to change this part of their system is significant.
In this paper we investigate how powerful we can make a uniform lan-
guage for meta-search without placing any burden on the solver writer.
The key to this is to only interact with the solver when a solution is found.
We present MINISEARCH, a meta-search language that can directly use
any FLATZINC solver. Optionally, it can interact with solvers through
an efficient C++ API. We illustrate the expressiveness of the language
and performance using different solvers on a number of examples.

1 Introduction

When using constraint programming (CP) technology, one often needs to exert
some control over the meta-search mechanism. Meta-search, such as Branch-
and-Bound search (BaB) or Large Neighbourhood Search (LNS), happens on
top of CP tree search, and aids finding good solutions, often by encoding meta-
information a modeller has about the problem.
Unfortunately, there is no widely accepted standard for controlling search or
meta-search. A wide range of high-level languages have been proposed that are
quite similar in how constraints are specified. However, they differ significantly
in the way search can be specified. This ranges from built-in minimisation and
maximisation only to fully programmable search. The main trade-off in devel-
oping search specification languages is the expressivity of the language versus
the required integration with the underlying solver. Fully programmable search,
including meta-search, is most expressive but requires deep knowledge of and

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 376–392, 2015.
DOI: 10.1007/978-3-319-23219-5 27
 MiniSearch: A Solver-Independent Meta-Search Language for MiniZinc 377

tight integration with a specific solver. Languages like OPL [24] and COMET [13]
provide convenient abstractions for programmable search for the solvers bundled
with these languages. On the other hand, some solver-independent languages
such as Esra [3], Essence [5] and Essence [6] do not support search specifica-
tions. Zinc [12] and MiniZinc [16] have support for specifying variable and value
ordering heuristics to CP solvers, but also no real control over the search.
Search combinators [19] was recently proposed as a generic meta-search lan-
guage for CP solvers. It interacts with the solver at every node in the search tree.
While very expressive, it requires significant engineering effort to implement for
existing solvers, since typically the solvers’ search engines are highly optimised
and tightly integrated with other components, such as the propagation engine
and state space maintenance.
In this paper, we introduce MiniSearch, a new combinator-like meta-search
language that has three design objectives: a minimal solver interface to facilitate
solver support, expressiveness, and, most importantly, solver-independence.
The objective to obtain a minimal solver interface stems from lessons learnt
in the design of search combinators that interact with the solver at every node.
In contrast, MiniSearch interacts with the underlying solving system only at
every solution, which is a minimal interface. At every solution, constraints can
be added or constraints in a well-defined scope can be removed, before asking
for the next solution. If the underlying solver does not support dynamic adding
and removing of constraints, MiniSearch can emulate this behaviour, for little
overhead.
Despite the lightweight solver interface, MiniSearch is surprisingly expres-
sive and supports many meta-search strategies such as BaB search, lexicographic
BaB, Large Neighbourhood Search variants, AND/OR search, diverse solution
search, and more. MiniSearch can also be used to create interactive optimi-
sation applications. Moreover, since MiniSearch builds upon MiniZinc, all
MiniZinc language features and built-ins can be used, for instance to formulate
custom neighbourhoods.
Solver-independence is the most important contribution of MiniSearch. All
solvers that can read and solve FlatZinc, which the majority of CP solvers
do [15,22], can be used with MiniSearch. Moreover, solvers that provide native
meta-search variants, such as branch-and-bound, can declare so and avoid exe-
cuting the MiniSearch decomposition instead. At the language level, this is
similar to the handling of global constraints in MiniZinc. Thus, solvers can
apply their strengths during meta-search, despite the minimal interface.

2 The MiniSearch language

MiniSearch is a high level meta-search language based on MiniZinc 2.0 [23].
MiniZinc is used for formulating the model and the constraints posted dur-
ing search, with language extensions for specifying the search. A MiniSearch
specification is provided, together with the search keyword, as an argument to
MiniZinc’s solve item. It is executed by the MiniSearch kernel (see Sec. 4).
378 A. Rendl et al.

With the built-in language extensions summarised in Tab. 1, users can define
functions such as the following branch-and-bound (BaB) minimisation:

Table 1. MiniSearch built-ins

MiniSearch built-ins Description

next() find the next solution
post(c) post the MiniZinc constraint c in the current scope
scope(s) open a local scope containing search s
s1 /\ s2 run s1 and iff successful, run s2
s1 \/ s2 run s1 and iff it fails, run s2
if s then s1 else s2 if s is successful, run s1 , otherwise s2
repeat(s) repeat search s until break is executed
repeat (i in 1..N )(s) repeat search s N times or until break is executed
break() break within a repeat
fail() return ‘failure’
skip() return ‘success’
time_limit(ms,s) run s until timelimit ms is reached
print(S) print MiniZinc output string S
print() print solution according to model output specification
c := v assign parameter c the value v
commit() commit to last found solution in function scope
sol(v) return solution value of variable v
hasSol() returns true if a solution has been found

1 include " minisearch . mzn " ;

2 var int : obj ; % other v a r i a b l e s and c o n s t r a i n t s not shown
3 solve search minimize_bab ( obj ) ;
4 output [ " Objective : " ++ show ( obj ) ];
5
6 function ann : minimize_bab ( var int : obj ) =
7 repeat ( if next () then commit () /\ print () /\ post ( obj < sol ( obj ) )
8 else break endif ) ;

The include item on line 1 includes the built-in MiniSearch function dec-
larations. This is necessary for any MiniZinc model that uses MiniSearch.
Line 3 contains the MiniZinc solve item followed by the new search keyword
and a user-defined MiniSearch function that takes a variable representing the
objective as argument. Line 4 is the MiniZinc output item, specifying how solu-
tions should be printed. Lines 7–8 contain the actual MiniSearch specification.
We will explain the different built-ins in more detail below, but the specification
can be read as follows: repeatedly try to find the next solution; and if that is suc-
cessful, commit to the solution, print it and add the constraint that the objective
must have a lower value than the current solution. If unsuccessful, break out of
the repeat.
All MiniSearch built-ins are typed as functions returning annotations.
Semantically, however, every MiniSearch built-in returns a value that repre-
sents either ‘success’ or ‘failure’, with respect to finding a solution. The handling
of these implicit return values is done by the MiniSearch interpreter (Sec. 4.1).
 MiniSearch: A Solver-Independent Meta-Search Language for MiniZinc 379

2.1 MiniSearch Built-Ins Involving the Solver

Communication with the solver is restricted to three forms: invoking the solver,
adding constraints/variables to the model, and scopes for temporary variables
and constraints.
Invoking the solver. The MiniSearch instruction for finding the next solution
is next(). It is successful if a solution has been found, and fails otherwise. The
variable/value labelling strategy (such as first-fail on the smallest domain value)
can be set in two ways: either by annotating the solve item (as in standard
MiniZinc), which sets the labelling globally, for every call to next(). Otherwise,
by annotating any MiniSearch function call, such as minimize_bab, with a
labelling strategy. Note, however, that as in MiniZinc, solvers may ignore these
annotations, for example if the labelling is not supported.
Solvers may declare support for native meta-search strategies, as with global
constraints, in which case these MiniSearch functions are treated as built-ins.
Adding constraints and variables. A constraint is added by calling the post()
built-in with a constraint as argument. Constraints can be formulated using the
same MiniZinc constructs as in the model, including global constraints, user-
defined functions and predicates. Variables can be dynamically added during
search too, using the MiniZinc let construct (see AND/OR search in Sec. 3.3).
Search Scopes. Search scopes define the lifespan of constraints and variables
in the model. MiniSearch has an implicit global search scope that contains
all variables and constraints of the model. A new search scope can be created
by using the scope(s) built-in that takes a MiniSearch specification s as an
argument. When entering a scope, search is assumed to start from the root again.
Whenever the execution leaves a scope, all constraints and variables that were
added in the scope are removed from the model and the solver. Execution in the
enclosing scope resumes from the point where it left off.

2.2 MiniSearch Control Built-Ins

All MiniSearch built-ins have an implicit return value that represents either
‘success’ (true) or ‘failure’ (false). Using this concept, we introduce MiniSearch
control built-ins. All built-ins execute their arguments in order.
And, Or, Repeat. The /\-built-in runs its arguments in order and stops to return
false as soon as one of its arguments fails. Similarly, the \/-built-in stops and
returns success as soon as one of its arguments succeeds. Existing control mech-
anisms of MiniZinc such as if then else endif expressions can be used as
well. The repeat(s) built-in takes a MiniSearch specification s and repeats
it until a break built-in is executed; returns false if a break happened, oth-
erwise returns what s returned. The delimited variant repeat(i in 1..N)(s)
will execute s for N iterations (or until break is executed).
Time-Limits. The built-in time_limit(ms,s) imposes a time limit ms (in mil-
liseconds) on any MiniSearch specification s. This way, s stops whenever the
380 A. Rendl et al.

time limit is reached, returning its current status. Time-limits are handled trans-
parently by the MiniSearch kernel as an exception.
Assigning values to constants. In standard MiniZinc constant parameters such
as int: N=10; cannot change their value. However, in MiniSearch we often
want to change constants across different iterations. For this purpose, we added
the assignment operator := which may only be used inside a MiniSearch spec-
ification. It overwrites that constant’s current value by the value supplied.

2.3 Solution Management

The strength of any meta-search language lies in using intermediate solutions to
guide the remaining search. For instance, branch-and-bound needs to access the
objective to post further constraints, and a Large Neighbourhood Search thaws
some of the variables in a solution to continue in that neighbourhood.
To facilitate working with solutions, the most recently found solution is
always accessible in MiniSearch using the sol built-in, where sol(x) returns
the value of x in the last solution. MiniSearch also provides a hasSol() built-in
to test whether a solution exists.
User-defined functions. When a MiniSearch strategy is defined as a MiniZinc
function, a local solution scope is created. This means that any solution found
by a call to next() inside the function is visible for the code in the function
body, but not for the caller of the function when the function returns. This
architecture allows for calls to next() to be encapsulated, i.e., a function can
make “tentative” calls to next in a nested search scope and only commit if these
succeed. Sect. 3.3 shows how AND/OR search can be implemented based on
this principle. In order to make the current solution accessible to the caller, the
function must call the commit built-in. A function returns ‘success’ if it called
commit() at least once, and ‘failure’ otherwise, and the last solution committed
by the function will then become the current solution of the caller.
Printing Solutions & Debugging. The print() function without any arguments
prints the last found solution in the format specified in the model’s output item.
Alternatively, print(s) provides more fine-grained control over the output. It
prints the string s, which can be constructed dynamically from values in the
solution using calls to sol. MiniSearch can be debugged using print() and
MiniZinc’s trace() function to display values of parameters, variables, and
arbitrary expressions during search. Furthermore, the MiniSearch interpreter
uses the C++ stack, so C++ debuggers can be used to follow the meta-search.

2.4 A Library of Search Strategies

Using the MiniSearch built-ins, we have defined and collected the most
common meta-search approaches in the standard library minisearch.mzn.1
1
The MiniSearch library is part of the open-source MiniSearch implementa-
tion [14].
 MiniSearch: A Solver-Independent Meta-Search Language for MiniZinc 381

These meta-search approaches can be used within any MiniZinc model that
includes the library. In the next section we present some of these meta-searches
in detail.

3 MiniSearch Examples
Despite MiniSearch’s limited communication with the solver, it provides
enough power to implement many useful complex searches that we illustrate
here.

3.1 Lexicographic BaB

In multi-objective optimisation, lexicographic optimisation can be used if the
objectives can be ranked according to their importance. The idea is to minimise
(or maximise) an array of objectives lexicographically. Lexicographic optimisa-
tion can be more efficient than the commonly used approach of obtaining a single
objective term by multiplying the components of the lexicographic objective with
different constants, as the latter approach leads to large objective values, and
potentially overflow.
Analogous to the implementation of branch-and-bound we post the global
constraint lex_less so that the next solution is lexicographically smaller than
the previous one. Below we show the respective MiniSearch specification.
1 function ann : minimize_lex ( array [ int ] of var int : objs ) =
2 next () /\ commit () /\ print () /\
3 repeat ( scope (
4 post ( lex_less ( objs , [ sol ( objs [ i ]) | i in index_set ( objs ) ]) ) /\
5 if next () then commit () /\ print () else break endif ) ) ;

In line 2 we search for an initial solution and, if successful, repeatedly open a new
scope (line 3). Then, we post the lexicographic (lex) constraint (line 4) and search
for another solution in line 5. This way, in each iteration of repeat, we add one
lex constraint, and all previously added lex constraints are removed due to the
scope. This is not required but beneficial, since posting several lex-constraints
can cause overhead if many intermediate solutions are found.

3.2 Large Neighbourhood Search (LNS)

Large area neighbourhood search (LNS) [20] is an essential method in the toolkit
of CP practitioners. It allows CP solvers to find very good solutions to very large
problems by iteratively searching large neighbourhoods (close) to optimality.

Randomised LNS explores a random neighbourhood of a given size, which can

be surprisingly effective in practice as long as the neighbourhood size is chosen
correctly. Cumulative scheduling is an example of a successful application [7].
The following MiniSearch specification of randomised LNS takes the objec-
tive variable, an array of decision variables that will be searched on, the number
382 A. Rendl et al.

of iterations, the destruction rate (the size of the neighbourhood) and a time limit
for exploring each neighbourhood. We have two scopes: in the global scope, we
post BaB style constraints (line 10); in the sub-scope (line 5), we search the
neighbourhoods. The predicate uniformNeighbourhood defines the neighbour-
hood: for each search variable we decide randomly whether to set it to its solution
value of the previous solution (line 15).
1 function ann : lns ( var int : obj , array [ int ] of var int : vars ,
2 int : iterations , float : destrRate , int : exploreTime ) =
3 repeat ( i in 1.. iterations ) (
4 print ( " Iteration " ++ show ( i ) ++ " \ n " ) /\
5 scope (
6 post ( u n i f o r m N e i g h b o u r h o o d ( vars , destrRate ) ) /\
7 time_limit ( exploreTime , minimize_bab ( obj ) ) /\
8 commit () /\ print ()
9 ) /\
10 post ( obj < sol ( obj ) )
11 );
12 predicate u n i f o r m N e i g h b o u r h o o d ( array [ int ] of var int : x , float : destrRate ) =
13 if hasSol () then
14 forall ( i in index_set ( x ) ) (
15 if uniform (0.0 ,1.0) > destrRate then x [ i ] = sol ( x [ i ]) else true endif )
16 else true endif ;

Adaptive LNS modifies the neighbourhood size over the course of the itera-
tions, depending on the success of previous iterations. Below is a simple variant,
where the neighbourhood size parameter nSize (line 3) is step-wise enlarged
each time no solution is found (line 9). The fail command fails the current
conjunction and will hence avoid that the post command on line 10 is executed.
1 function ann : adaptive_lns ( var int : obj , array [ int ] of var int : vars ,
2 int : iterations , int : initRate , int : exploreTime ) =
3 let { int : nSize = initRate , int : step = 1; } in
4 repeat ( i in 1.. iterations ) (
5 print ( " Iteration " ++ show ( i ) ++ " , rate = " ++ show ( nSize ) ++ " \ n " ) /\
6 scope ( ( post ( u n i f o r m N e i g h b o u r h o o d ( vars , nSize /100.0) ) /\
7 time_limit ( exploreTime , minimize_bab ( obj ) ) /\
8 commit () /\ print ()
9 ) \/ ( nSize := nSize + step /\ fail ) )
10 /\ post ( obj < sol ( obj ) ) );

Custom Neighbourhoods can sometimes be effective if they capture some

insight into the problem structure. For instance, in a Vehicle Routing Prob-
lem (VRP), we might want to keep certain vehicle tours or vehicle-customer
assignments. Below we show such a custom neighbourhood that is easily speci-
fied in MiniSearch. The predicate keepTour (line 1) posts the tour constraints
of a given vehicle number vNum. If a solution exists, the neighbourhood predi-
cate (line 4) determines the number of customers of each vehicle (line 7), and
then randomly chooses to keep the vehicle’s tour (line 8) where a high customer
usage results in a higher chance of keeping the vehicle. This predicate can be
used instead of the uniform neighbourhood in the LNS specifications above.
 MiniSearch: A Solver-Independent Meta-Search Language for MiniZinc 383

1 predicate keepTour ( int : vNum ) =

2 forall ( i in 1.. nbCustomers where sol ( vehicle [ i ]) == vNum )
3 ( successor [ i ] = sol ( successor [ i ]) ) ;
4 predicate v e h i c l e N e i g h b o u r h o o d () =
5 if hasSol () then
6 forall ( v in 1.. nbVehicles ) (
7 let { int : usage = sum ( c in 1.. nbCustomers ) ( sol ( vehicle [ c ]) == v ) }
in
8 if usage > uniform (0 , nbCustomers ) then
9 keepTour ( v ) % higher usage -> higher chance of keeping the
vehicle
10 else true endif )
11 else true endif ;

3.3 AND/OR Search

Search in CP instantiates variables according to a systematic variable labelling,
corresponding to an OR tree. AND/OR search decomposes problems into a mas-
ter and several conjunctive slave sub-problems. An AND/OR search tree consists
of an OR tree (master problem), an AND node with one branch for each sub-
problem, and OR trees underneath for the sub-problems. A prominent example
is stochastic two-stage optimisation [18], where the objective is to find optimal
first-stage variable assignments (master problem) such that all second-stage vari-
able assignments (sub-problems) are optimal for each scenario. AND/OR search
for stochastic optimisation is called policy based search [27]. AND/OR search is
also applied in other applications, such as graphical models [11].
Below is a MiniSearch example of AND/OR search for stochastic two-stage
optimisation. The variables and constraints of each scenario (sub-problem) are
added incrementally during search.
1 function ann : p o l i c y _ b a s e d _ s e a r c h _ m i n ( int : sc ) =
2 let {
3 array [1.. sc ] of int : sc_obj = [ 0 | i in 1.. sc ];
4 int : expectedCosts = infinity ;
5 } in (
6 repeat (
7 if next () then % solution for master
8 repeat ( s in 1.. sc ) (
9 scope ( % a local scope for each s u b p r o b l e m
10 let {
11 array [ int ] of var int : recourse ; % s u b p r o b l e m v a r i a b l e s
12 var 0.. maxCosts : scenarioCosts ;
13 } in (
14 post ( s e t F i r s t S t a g e V a r i a b l e s () /\ % assign master variables

15 secondStageCts (s , recourse ) ) /\ % s u b p r o b l e m
constraints
16 if minimize_bab ( scenarioCosts ) then
17 sc_obj [ s ] := sol ( scenarioCosts )
18 else print ( " No solution for scenario " ++ show ( s ) ++ " \ n " ) /\
19 break endif
20 ) )
21 ) /\ % end repeat
22 if expectedCosts > expectedValue ( sc_obj ) then
23 expectedCosts := expectedValue ( sc_obj ) /\
24 commit () % we found a better AND / OR s o l u t i o n
25 else skip () endif
26 else break endif % no master solution
27 ) );
384 A. Rendl et al.

28 % the f o l l o w i n g p r e d i c a t e s are d e f i n e d in the model a c c o r d i n g to the

problem class
29 predicate s e t F i r s t S t a g e V a r i a b l e s () ;
30 predicate secondStageCts ( int : scenario , array [ int ] of var : y ) ;
31 function int : expectedValue ( var int : sc_obj ) ;

Lines 3-4 initialise parameters that represent the costs for each scenario/sub-
problem and the combined, expected cost of the master and subproblems. Line 7
searches for a solution to the master problem (OR tree), and if this succeeds, we
continue with the AND search by finding the optimal solution for each subprob-
lem, based on the master solution (line 9): we create each subproblem in a local
scope, and add the respective variables (line 11- 12) and constraints (line 15).
Furthermore, we set the master variables to the values in the master solution
(line 14). Then we search for the minimal solution for the scenario (line 16), and,
if successful, store it in sc_obj (line 17). If we find a solution for each scenario,
then we compute the combined objective (expectedValue) and compare it to the
incumbent solution (line 22). If we found a better solution, we store its value
(line 23) and commit to it (line 24) and continue in the master scope (line 7).
Otherwise, if we find no solution for one of the scenarios (line 18), the mas-
ter solution is invalid. We therefore break (line 19) and continue in the master
problem scope, searching for the next master solution (line 7).

3.4 Diverse Solutions

Sometimes we don’t just require a satisfying or optimal solution, but a diverse

set of solutions [9]. The MiniSearch specification below implements the greedy
approximation method that iteratively constructs a set of K diverse solutions:
1 function ann : g r e e d y _ m a x D i v e r s e K s e t ( array [ int ] of var int : Vars , int : K ) =
2 let { array [ int , int ] of int : Store = array2d (1.. K , index_set ( Vars ) ,
3 [0 | x in 1.. K * length ( Vars ) ]) ,
4 int : L = 0 % current length
5 } in
6 next () /\ commit () /\ print () /\
7 repeat (
8 L := L +1 /\ repeat ( i in 1.. length ( Vars ) ) ( Store [L , i ] := sol ( Vars [ i ]) ) /\
9 if L <= K then
10 scope (
11 let { var int : obj ;} in
12 post ( obj = sum ( j in 1.. L , i in index_set ( Vars ) ) ( Store [j , i ] != Vars [ i ]) ) /\
13 maximize_bab ( obj ) /\ commit () /\ print () )
14 else print ( show ( Store ) ++ " \ n " ) /\ break endif
15 );

The first few lines initialise the Store, which will contain the K diverse solutions,
as well as the current length of the store up to which it already contains solutions.
On line 8 the length is increased and the previously found solution is saved. If the
length does not exceed K, we construct a new objective on line 12. This objective
expresses how different a solution is from the previously found solutions using
the Hamming distance. This objective is then maximised, resulting in the next
most diverse solution, and the process repeats.
 MiniSearch: A Solver-Independent Meta-Search Language for MiniZinc 385

3.5 Interactive Optimisation

Interactive optimisation lets users participate in the solving process by inspect-
ing solutions, adding constraints, and then re-solving. This has been shown to
improve the trust of end-users into decision support systems, and a number of
successful application have been implemented [2].
MiniSearch supports interactive optimisation by calling MiniZinc built-
ins that ask for user input. The MiniZinc library contains functions such as
read_int() which accept keyboard input. More advanced input facilities can
be added through user-defined MiniZinc built-ins that execute arbitrary C++
code, such as consulting a user through a graphical user interface or other means.
The following is an example of an interactive Vehicle Routing Problem solver
implemented using MiniSearch. We use a Large Neighbourhood search where
in every N -th iteration, the user can chose a vehicle route that should be kept.
This code can be used with the LNS implementations described earlier.
1 predicate i n t e r a c t i v e N e i g h b o u r h o o d ( int : iteration , int : N ) =
2 if iteration mod N = 0 then
3 let { string : msg = " Enter a vehicle tour to keep (0 for none ) :\ n " ;
4 int : n = read_int ( msg )
5 } in if n > 0 then keepTour ( n ) else true endif
6 else true endif ;

4 The MiniSearch Kernel

This section describes the architecture of the MiniSearch kernel, the engine
that executes MiniSearch specifications, interacting with both the MiniZinc
compiler and the backend solver.
First, let us briefly review how MiniZinc models are solved. The MiniZinc
compiler (usually invoked as mzn2fzn) takes text files containing the model and
instance data and compiles them into FlatZinc, a low-level language supported
by a wide range of solvers. The compiler generates FlatZinc that is specialised
for the capabilities of the particular target solver using a solver-specific library of
predicate declarations. Solvers read FlatZinc text files and produce text-based
output for the solutions. Version 2.0 of MiniZinc is based on the libminizinc
C++ library, which provides programmatic APIs, eliminating the need to com-
municate through text files. The library also defines an API for invoking solvers
directly in C++. MiniSearch is built on top of these APIs.

4.1 The MiniSearch Interpreter

The MiniSearch kernel implements an interpreter that processes MiniSearch
specifications and handles the communication between the MiniZinc compiler
and the solver. The interpreter assumes that the solver interface is incremen-
tal, i.e. variables and constraints can be added dynamically during search. We
provide an emulation layer (see Sec. 4.2) for solvers that do not support incre-
mental operations, including solvers that cannot be accessed through the C++
interface.
386 A. Rendl et al.

The MiniSearch interpreter is a simple stack-based interpreter that main-

tains the following state: A stack of solutions, one for each function scope; a
stack of time-outs and breaks; and a stack of search scopes, containing the solver
state of each scope. The interpreter starts by compiling the MiniZinc model
into FlatZinc, and then interprets each MiniSearch built-in as follows:

– next() invokes the solver for a new solution; if successful, it replaces the
top-most solution on solution stack. If a time-out has been set, the call to
the solver is only allowed to run up to the time-out. Returns true iff a new
solution was found.
– commit() replaces the parent solution (2nd on the stack) with the current
solution. This commits the current solution into the parent function scope.
Returns true. Aborts if not in a function call.
– function calls duplicate the top of the solution stack before executing the
function body. Return true if the function committed a new solution.
– time_limit(l,s) adds a new time-out now+l to the stack, executes s, and
then pops the time-out. During the execution of s, calls to next and repeat
check whether any time-outs t have expired (t > now), and if so they imme-
diately break. Returns whatever s returned.
– repeat(s) pushes a break scope with a Boolean flag on the stack, then
repeats the execution of s as long as the break flag is false and pops the
break scope when s is finished. The break construct sets the break flag in
the current break scope (similar to a time limit).
– post(c) compiles the MiniZinc expression c into FlatZinc. The compi-
lation is incremental, i.e., the result of compiling c is added to the existing
FlatZinc. The interpreter then adds the newly generated variables and
constraints to the current solver instance.
– scope(s) creates a new local scope. The current implementation copies the
flat model and creates a new solver instance based on this flat model.
– Other operations (/\,\/,print) are interpreted with respect to their seman-
tics.

4.2 Emulating Advanced Solver Behaviour

A key goal of this work is to make MiniSearch available for any solver that sup-
ports FlatZinc. Current FlatZinc solvers, however, neither support incremen-
tally adding constraints and variables, nor do they implement the libminizinc
C++ API. We therefore need to emulate the incremental API.
In order to emulate incrementality, we can implement the dynamic addition
of variables and constraints using a restart-based approach, re-running the solver
on the entire updated FlatZinc. To avoid re-visiting solutions, after each call to
next the emulation layer adds a no-good to the model that excludes the solution
that was just found. This emulation reduces the requirements on the solver to
simply being able to solve a given FlatZinc model.
In order to emulate the C++ API, we generate a textual representation of
the FlatZinc and call an external solver process. The emulator then converts
 MiniSearch: A Solver-Independent Meta-Search Language for MiniZinc 387

the textual output of the process back into libminizinc data structures. Using
this emulation of the incremental C++ solver API, any current FlatZinc solver
can be used with the MiniSearch kernel.

4.3 Built-in Primitives

Solvers can declare native support for a MiniSearch search specification. For
every MiniSearch call such as f(x,y,z), the kernel will check whether the dec-
laration of f has a function body. If it does not have a function body, the function
is considered to be a solver built-in, and it is executed by passing the call to a
generic solve function in the C++ solver API. This is similar to the handling of
global constraints in MiniZinc, where solvers can either use their own primitives
for a given global constraint, or use the respective MiniZinc decomposition of
the global constraint. This way solvers can easily apply their own primitives, but
can use alternatives if they do not support a certain feature.

5 Experiments

In our experimental evaluation, we analyse MiniSearch on practical exam-

ples to study the efficiency and compare the overhead of MiniSearch for dif-
ferent interfaces. Furthermore, we study the benefits of specialised, heuristic
MiniSearch approaches to standard branch-and-bound optimisation; the only
available option in solver-independent CP modelling languages.

5.1 Experimental Setup

The problems and instances are taken from the MiniZinc benchmarks reposi-
tory2 . The source code of MiniSearch and the respective solver interfaces will
be released in September 2015 [14].
Experiments are run on Ubuntu 14.04 machines with eight i7 cores and 16GB
of RAM. The MiniSearch kernel is based on the MiniZinc 2.0.1 toolchain. The
native Gecode interface and incremental C++ interface were implemented using
the latest Gecode source code (version 4.3.3+, 20 April 2015). The FZN solvers
used are: Gecode (20 April 2015), Choco 3.3.0, Or-tools source (17 February
2015), and Opturion CPX 1.0.2. All solvers use the same variable/value ordering
heuristic that is part of the MiniZinc model.

5.2 Overhead of Different MiniSearch Interfaces

First, we study the comparative overhead of MiniSearch with respect to

the different interface levels. We do this by analysing the performance of a
meta-search approach on the same solver, Gecode. We compare the following
four approaches: the solver’s native approach (without MiniSearch) through
2
https://github.com/MiniZinc/minizinc-benchmarks
388 A. Rendl et al.

Table 2. Rectangle Packing. Times (sec) averaged over 10 runs. Comparing FlatZinc-
solvers with MiniSearch BaB through FlatZinc (MS-F), MiniSearch BaB through
incremental API (MS-Inc), and native BaB with FlatZinc (Nat-F).

Rectangle Gecode or-tools choco Opturion CPX

Size (N ) MS-F MS-Inc Nat-F MS-F Nat-F MS-F Nat-F MS-F Nat-F
14 0.7 0.1 0.2 0.5 0.2 6.4 1.0 0.3 0.1
15 0.8 0.3 0.4 0.6 0.3 6.0 1.3 0.4 0.2
16 1.3 0.5 0.5 1.1 0.5 9.1 1.6 0.6 0.2
17 4.5 2.9 2.9 5.2 3.8 16.6 5.7 0.9 0.4
18 37.4 37.4 36.8 41.2 41.5 38.0 29.8 6.7 6.2
19 77.5 43.2 42.0 59.0 32.8 59.2 28.8 12.7 7.7
20 96.4 97.9 94.1 99.2 98.9 93.9 83.0 6.4 5.6
21 472.4 469.7 462.5 416.3 410.0 250.1 239.2 83.8 82.1

FlatZinc (Nat-F), the incremental API interface to MiniSearch (MS-Inc) and

the non-incremental text-based (MS-F) interface to MiniSearch. We use Rect-
angle Packing [21] as problem benchmark, and standard BaB as meta-search,
since it is available natively in FlatZinc for Gecode.
The results are summarised in Table 2, columns 2–4. It shows the runtimes
(in seconds) taken to find the optimal solution for the different approaches using
Gecode. The native approach through the FlatZinc interface is fastest, though
MiniSearch through the incremental API interface performs quite similarly.
Moreover, MiniSearch through the FZN interface is, as expected, slower, but
only by a small factor. These results are very promising, since they show that
the overhead of the MiniSearch interfaces is low.
In addition, we analyse the overhead of MiniSearch for other FlatZinc
solvers in Table 2, columns 5–10. We ran the Rectangle Packing problem for the
FlatZinc solvers or-tools, Choco, and Opturion CPX on a MiniZinc model
using both native branch-and-bound (Nat-F), and MiniSearch branch-and-
bound (MS-F), both through the FlatZinc interface. Overall the MiniSearch
FlatZinc interface, while slower, still gives decent performance. Choco seems
to incur most overhead of the repeated restarts.

5.3 Heuristic Search

MiniSearch strategies are not restricted to complete search. For example, dif-
ferent variants of Large Neighbourhood Search can be tried. To demonstrate the
benefits of having heuristic search at hand, we compare the Randomised LNS
approach from Section 3.2 with the standard out-of-the-box branch-and-bound
approach for a selection of FlatZinc solvers on the Capacitated Vehicle Routing
problem (CVRP).
We run each solver on (1) the standard MiniZinc model that uses branch-
and-bound, and (2) the MiniZinc model with the MiniSearch specification of
Randomised LNS. We use the Augerat et al [1] CVRP instance sets A,B and P,
where the 27 Set-A instances contain random customer locations and demands,
 MiniSearch: A Solver-Independent Meta-Search Language for MiniZinc 389

Table 3. Average relative improvement of MiniSearch LNS over standard BaB for
the achieved objective within a time limit for different solvers. The figures are averages
over the Augerat et al [1] CVRP instance sets A,B and P.

CVRP Instance-Set Gecode or-tools choco Opt. CPX

Set-A 11.60% 11.76% 11.17% 12.11%
Set-B 13.38% 11.82% 12.62% 14.92%
Set-P 9.78% 10.53% 7.98% 11.35%

the 22 Set-B instances have clustered locations and the 23 Set-P instances are
modified versions of instances from the literature. The LNS approach uses a
destroy rate of 0.3 (the neighbourhood size is 30%) and an exploration timeout
(for each neighbourhood) of 5 seconds. Both the LNS and BaB approach have
an overall 2 minute timeout, and we report the quality of the best solution found
on time-out.
We can see that LNS provides better solutions than BaB, and Table 3 reports
the average improvement of the objective of LNS over BaB on the three instance
sets. We observe an improvement between 8-15%, which is considerable within
the given time limit. Furthermore, we see that the improvements are similar for
all solvers across the instances sets, in particular for the Set-B instances. This
demonstrates the added value of MiniSearch search strategies to the existing
out-of-the-box optimisation approaches of standard modelling languages.

6 Related Work

The starting point for MiniSearch is search combinators [19], and MiniSearch
is indeed a combinator language. The MiniSearch language is more restricted
than search combinators. In particular, search combinators can interact with
the search at every node of the tree, essentially replacing the solver’s built-in
search, while MiniSearch only interacts with the solver’s search at the granu-
larity of solutions. Examples not possible in MiniSearch are the use of average
depth of failure or custom limits based on how many search nodes with a par-
ticular property have been encountered, and heuristics such as limited discrep-
ancy search [8] and iterative deepening. Other search combinator expressions are
however expressible with MiniSearch, as witnessed by the library of strategies
presented in this paper.
Neither search combinators nor MiniSearch support fully programmable
search. Especially not variable and value ordering heuristics, which interact with
the internal data structures of the solver. These kinds of search strategies are
important, but cannot be implemented without deep hooks into the solver.
Objective CP [26] can express complex meta-search at the model level, which
can then be executed by any of the Objective CP solvers. For example a CP-
style search can be implemented using a MIP solver. So it also provides solver-
independent search, but through a much more fine grained interface, supporting
interaction at each search node. The fine grained interface however means that
390 A. Rendl et al.

adding a new solver to Objective CP is complex, in particular supporting the

reverse mapping of information from solver objects to model objects.
Possibly the closest system to MiniSearch is OPL Script [25], a scripting
language for combining models. It essentially also communicates to solvers via
models and solutions. Distinct from MiniSearch is that it is object-oriented and
provides models as first-class objects, and hence supports the construction and
modification of multiple models, as opposed to our simpler mechanisms. It can
communicate more information than just variables and constraints, including
dual values, bases, and various statistics. However, the language is tightly linked
to the CP/MIP solver ILOG CPLEX CP Optimizer. MiniSearch on the other
hand can use any FlatZinc solver.
AMPL Script [4] is similar in many respects to OPL script, though a more
basic scripting language, not supporting function definitions for example. It does
not have a concept of scopes, but instead supports adding, dropping and restor-
ing variables and constraints from the model dynamically. It interacts with the
solver through requests to solve, which continue until some termination condi-
tion arises, like optimality proven or a timeout. Hence its natural interaction
level is more like minimize bab() than next().
Any meta-search that can be implemented in MiniSearch could also be
implemented directly in a solver such as Gecode, Choco, Comet, or CP Opti-
mizer, although the implementation burden for doing so could be significant.
Similarly, APIs for CP embedded in scripting languages such as Numberjack [10]
(through Python) could offer a MiniSearch-like interface. The benefits of
MiniSearch are that it is programming language- and solver-independent, that
it abstracts away solution and timeout handling. Furthermore, meta-search lan-
guages are easier to use when they extend the language for expressing constraints,
in this case MiniZinc, since this is often a vital component of the search.

7 Conclusion and Future Work

In this paper we present MiniSearch, a solver-independent lightweight lan-

guage to specify meta-search in MiniZinc models. The first contribution of
MiniSearch is its expressiveness: we showed how to formulate different com-
plex search approaches in MiniSearch, such as Large Neighbourhood Search or
AND/OR search. It reuses the MiniZinc modelling language for specifying con-
straints and variables to add during search. It even enables the implementation
of interactive optimisation approaches where the user guides the search process.
The second contribution is the minimal interface of MiniSearch. Since
the kernel interprets all statements, the communication between the solver and
MiniSearch is only necessary at every solution (instead of at every search node
as in search combinators). This makes it more realistic to implement for solver
developers. Furthermore, in our experiments we observed that both the incre-
mental and the text-based interface provide satisfactory performance.
The third and most important contribution is solver -independence. In con-
trast to existing search languages, MiniSearch can already use any solver that
 MiniSearch: A Solver-Independent Meta-Search Language for MiniZinc 391

can process FlatZinc, which is the majority of CP solvers. This allows users to
test different solvers with complex meta-searches without having to commit to
one single solver.
For future work, we plan to extend MiniSearch with parallelism, so that
independent searches can be executed in parallel and advanced parallel meta-
search approaches such as Embarrassingly Parallel Search [17] can be specified.
Furthermore, we plan to integrate portfolio search into MiniSearch, so that
different search functions can be executed by different solvers.

Acknowledgments. NICTA is funded by the Australian Government through the

Department of Communications and the Australian Research Council through the ICT
Centre of Excellence Program. T.G. is supported by a Postdoc grant from the Research
Foundation – Flanders.

References
1. Augerat, P., Belenguer, J., Benavent, E., Corberan, A., Naddef, D., Rinaldi, G.:
Computational results with a branch and cut code for the capacitated vehicle
routing problem. Technical Report 949-M. Universite Joseph Fourier, Grenoble
(1995)
2. Belin, B., Christie, M., Truchet, C.: Interactive design of sustainable cities with
a distributed local search solver. In: Simonis, H. (ed.) CPAIOR 2014. LNCS,
vol. 8451, pp. 104–119. Springer, Heidelberg (2014)
3. Flener, P., Pearson, J., Ågren, M.: Introducing esra, a relational language for
modelling combinatorial problems (Abstract). In: Rossi, F. (ed.) CP 2003. LNCS,
vol. 2833, pp. 971–971. Springer, Heidelberg (2003)
4. Fourer, R., Gay, D.M., Kernighan, B.W.: AMPL: A Modeling Language for Math-
ematical Programming. Cengage Learning (2002)
5. Frisch, A.M., Harvey, W., Jefferson, C., Hernández, B.M., Miguel, I.: Essence :
A constraint language for specifying combinatorial problems. Constraints 13(3),
268–306 (2008)
6. Gent, I.P., Miguel, I., Rendl, A.: Tailoring solver-independent constraint models:
a case study with Essence and Minion. In: Miguel, I., Ruml, W. (eds.) SARA
2007. LNCS (LNAI), vol. 4612, pp. 184–199. Springer, Heidelberg (2007)
7. Godard, D., Laborie, P., Nuijten, W.: Randomized large neighborhood search for
cumulative scheduling. In: Proceedings of the Fifteenth International Conference
on Automated Planning and Scheduling (ICAPS 2005), Monterey, California, USA,
June 5–10 2005, pp. 81–89 (2005)
8. Harvey, W.D., Ginsberg, M.L.: Limited discrepancy search. In: Proceedings of the
14th IJCAI, pp. 607–613 (1995)
9. Hebrard, E., Hnich, B., O’Sullivan, B., Walsh, T.: Finding diverse and similar
solutions in constraint programming. In: Veloso, M.M., Kambhampati, S. (eds.)
AAAI, pp. 372–377. AAAI Press / The MIT Press (2005)
10. Hebrard, E., O’Mahony, E., O’Sullivan, B.: Constraint programming and com-
binatorial optimisation in numberjack. In: Lodi, A., Milano, M., Toth, P. (eds.)
CPAIOR 2010. LNCS, vol. 6140, pp. 181–185. Springer, Heidelberg (2010)
11. Marinescu, R., Dechter, R.: AND/OR branch-and-bound search for combinatorial
optimization in graphical models. Artif. Intell. 173(16–17), 1457–1491 (2009)
392 A. Rendl et al.

12. Marriott, K., Nethercote, N., Rafeh, R., Stuckey, P.J., de la Banda, M.G., Wallace,
M.: The design of the Zinc modelling language. Constraints 13(3), 229–267 (2008)
13. Michel, L., Van Hentenryck, P.: The comet programming language and system. In:
van Beek, P. (ed.) CP 2005. LNCS, vol. 3709, p. 881. Springer, Heidelberg (2005)
14. MiniSearch release. http://www.minizinc.org/minisearch
15. MiniZinc challenge. http://www.minizinc.org/challenge.html
16. Nethercote, N., Stuckey, P.J., Becket, R., Brand, S., Duck, G.J., Tack, G.R.: MiniZ-
inc: towards a standard CP modelling language. In: Bessière, C. (ed.) CP 2007.
LNCS, vol. 4741, pp. 529–543. Springer, Heidelberg (2007)
17. Régin, J.-C., Rezgui, M., Malapert, A.: Embarrassingly parallel search. In: Schulte,
C. (ed.) CP 2013. LNCS, vol. 8124, pp. 596–610. Springer, Heidelberg (2013)
18. Ruszczyński, A., Shapiro, A.: Stochastic Programming. Handbooks in operations
research and management science. Elsevier (2003)
19. Schrijvers, T., Tack, G., Wuille, P., Samulowitz, H., Stuckey, P.J.: Search combi-
nators. Constraints 18(2), 269–305 (2013)
20. Shaw, P.: Using constraint programming and local search methods to solve vehicle
routing problems. In: Maher, M.J., Puget, J.-F. (eds.) CP 1998. LNCS, vol. 1520,
pp. 417–431. Springer, Heidelberg (1998)
21. Simonis, H., O’Sullivan, B.: Search strategies for rectangle packing. In: Stuckey,
P.J. (ed.) CP 2008. LNCS, vol. 5202, pp. 52–66. Springer, Heidelberg (2008)
22. Stuckey, P.J., Feydy, T., Schutt, A., Tack, G., Fischer, J.: The MiniZinc challenge
2008–2013. AI Magazine 35(2), 55–60 (2014)
23. Stuckey, P.J., Tack, G.: Minizinc with functions. In: Gomes, C., Sellmann, M. (eds.)
CPAIOR 2013. LNCS, vol. 7874, pp. 268–283. Springer, Heidelberg (2013)
24. Van Hentenryck, P.: The OPL Optimization Programming Language. MIT Press,
Cambridge (1999)
25. Van Hentenryck, P., Michel, L.: OPL script: composing and controlling models.
In: Apt, K.R., Kakas, A.C., Monfroy, E., Rossi, F. (eds.) Compulog Net WS 1999.
LNCS (LNAI), vol. 1865, pp. 75–90. Springer, Heidelberg (2000)
26. Van Hentenryck, P., Michel, L.: The objective-CP optimization system. In: Schulte,
C. (ed.) CP 2013. LNCS, vol. 8124, pp. 8–29. Springer, Heidelberg (2013)
27. Walsh, T.: Stochastic Constraint Programming. In: van Harmelen, F. (ed.) ECAI,
pp. 111–115. IOS Press (2002)
 Two Clause Learning Approaches
for Disjunctive Scheduling

Mohamed Siala1,2,3(B) , Christian Artigues2,4 , and Emmanuel Hebrard2,4

1
Insight Centre for Data Analytics, University College Cork, Cork, Ireland
mohamed.siala@insight-centre.org, siala@laas.fr
2
CNRS, LAAS, 7 Avenue du Colonel Roche, F-31400 Toulouse, France
3
Univ. de Toulouse, INSA, LAAS, F-31400 Toulouse, France
4
Univ. de Toulouse, LAAS, F-31400 Toulouse, France
{artigues,hebrard}@laas.fr

Abstract. We revisit the standard hybrid CP/SAT approach for solving

disjunctive scheduling problems. Previous methods entail the creation of
redundant clauses when lazily generating atoms standing for bounds mod-
ifications. We first describe an alternative method for handling lazily gen-
erated atoms without computational overhead. Next, we propose a novel
conflict analysis scheme tailored for disjunctive scheduling. Our exper-
iments on well known Job Shop Scheduling instances show compelling
evidence of the efficiency of the learning mechanism that we propose. In
particular this approach is very efficient for proving unfeasibility.

1 Introduction
Disjunctive scheduling refers to a large family of scheduling problems having in
common the Unary Resource Constraint. This constraint ensures that a set of
tasks run in sequence, that is, without any time overlap. The traditional con-
straint programming (CP) approaches for this problem rely on tailored propa-
gation algorithms (such as Edge-Finding [6,17,24]) and search strategies (such
as Texture [20]). The technique of Large Neighborhood Search [22] (LNS) was
also extensively used in this context [7,25].
A different type of approaches emerged recently, based on the so-called
Conflict-Driven Clause Learning (CDCL) algorithm for SAT [16]. This proce-
dure uses resolution to learn a new clause for each conflict during search. Recent
constraint programming approaches similarly trade off strong propagation-based
inference for a way to learn during search. For instance, a hybrid CP/SAT
method, Lazy Clause Generation (LCG) [8,9,18] was shown to be extremely effi-
cient on the more general Resource Constrained Project Scheduling Problem [21]
(RCPSP ). Even simple heuristic weight-based learning was shown to be very
efficient on disjunctive scheduling [10–13]. The so called light model combines
minimalist propagation with a slight variant of the weighted degree [5] heuristic.
In this paper, we propose a hybrid CP/SAT method based on this light
model. Similarly to LCG, our approach mimics CDCL. However, it differs from


c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 393–402, 2015.
DOI: 10.1007/978-3-319-23219-5 28
394 M. Siala et al.

LCG in two main respects: First, as the time horizon can be large, literals repre-
senting changes in the bounds of the tasks domains should be generated “lazily”
during conflict analysis as it was proposed in [8]. However, handling domain con-
sistency through clauses entails redundancies and hence suboptimal propagation.
We use a dedicated propagation algorithm, running in constant amortized time,
to perform this task. The second contribution is a novel conflict analysis scheme
tailored for disjunctive scheduling. This technique could be applied to any prob-
lem where search can be restricted to a predefined set of Boolean variables. This
is the case of disjunctive scheduling since once every pair of tasks sharing the
same resource is sequenced, we are guaranteed to find a complete solution in
polynomial time. Most methods therefore only branch on the relative order of
tasks sharing a common resource. We propose to use this property to design a
different conflict analysis method where we continue resolution until having a
nogood with only Boolean variables standing for task ordering. As a result, we
do not need to generate domain atoms.
We compare the two methods experimentally and show that the benefit of
not having to generate new atoms during search outweigh in many cases the
more expressive language of literals available in traditional hybrid solvers. The
novel approach is very efficient, especially for proving unfeasibility. We were
able to improve the lower bounds on several well known Job Shop Scheduling
Problem (JSSP ) benchmarks. However, a method implemented within IBM CP-
Optimizer studio recently found, in general, better bounds [25].
The rest of the paper is organized as follows: We give a short background
on hybrid CP/SAT solving in Section 2. Next, we briefly describe the baseline
CP model and strategies that we use in Section 3. We introduce in Section 4
our new lazy generation approach. In Section 5, we present our novel conflict
analysis scheme. Last, we give and discuss the experimental results in Section 6.

2 Hybrid CP/SAT Solving

In this section, we briefly describe the basic mechanisms and notations of hybrid
CP/SAT used in modern Lazy Clause Generation solvers [8,9,18].
Let X = [x1 , .., xn ] be a sequence of variables. A domain D maps every vari-
able x ∈ X to a finite set D(x)⊂ Z. D(x) is a range if max(D(x)) − min(D(x)) +
1 = |D(x)|. A constraint C is defined by a relation over the domains of a sequence
of variables. In this context, each constraint is associated to a propagator to
reduce the domains with respect to its relation. Propagators are allowed to use
only domain operations of a given type (often x = v, x = v, x ≤ v, and
x ≥ v). Every domain operation is associated to a literal p that can have an
atom (i.e. Boolean variable) generated “eagerly” before starting the search, or
“lazily” when learning a conflict involving this change. Let p be a literal corre-
sponding to a domain operation. level(p) is the number of nodes in the current
branch of the search tree when p becomes true (i.e., when this domain opera-
tion takes place). Moreover, rank(p) is the number of such domain operations
that took place at the same node of the search tree before p. Last, a literal p is
 Two Clause Learning Approaches for Disjunctive Scheduling 395

associated to a logical implication of the form ℘ ⇒ p where ℘ is a conjunction of

literals sufficient for a given propagator to deduce p. We call ℘ an explanation of
p. We assume that the function explain(p) returns the explanation ℘ for p. The
explanations can be computed at the moment of propagation (i.e., clausal and
forward explanations in [8]), or on demand during conflict analysis (backward ).
The notion of explanation is extended to a failure ⊥ in the natural way.
Whenever a failure occurs during search, a conflict analysis procedure is
used to compute a nogood, that is, an explanation of the failure which is not
yet logically implied by the model. Algorithm 1 depicts the conflict analysis
procedure based on the 1-UIP scheme [27]. The nogood Ψ is initialized in Line 1
to the explanation for the failure. It is returned once it contains a single literal
that has been propagated at the current level (Line 2). When this is not the
case, the last propagated literal is selected and its explanation isresolved with
the current nogood in Line 3. The final nogood Ψ can be seen as i∈[1,n] pi ⇒ p
The solver then backtracks to the level l = max(level(pi )).

Algorithm 1. 1-UIP-with-Propagators
1 Ψ ← explain(⊥) ;
2 while |{q ∈ Ψ | level(q) = current level}| > 1 do
p ← arg maxq ({rank(q) | level(q) = current level ∧ q ∈ Ψ }) ;
3 Ψ ← Ψ ∪ {q | q ∈ explain(p) ∧ level(q) > 0} \ {p} ;
return Ψ ;

Consider now the lazy generation of atoms (i.e. after computing Ψ ). When
x ≤ u has to be generated, the clauses ¬x ≤ a ∨ x ≤ u; ¬x ≤ u ∨ x ≤ b
where a and b are the nearest generated bounds to u with a < u < b must
be added to maintain the consistency of x ≤ u with respect to previously
generated atoms. In this case, the clause ¬x ≤ l ∨ x ≤ u becomes redundant.

3 A Simple CP Model for Job Shop Scheduling

A JSSP consists in sequencing the tasks from n jobs on m machines. The jobs
J = {Ji | 1 ≤ i ≤ n} define sequences of tasks to be scheduled in a given
order. Moreover, each of the m tasks of a job must be processed by a different
machine. We denote Tik the task of job Ji requiring machine Mk . Each task Tik
is associated to a processing duration pik . Let tik be the variable representing the
starting time of task Tik . The processing interval of a task Tik is [tik , tik + pik ).
For all k ∈ [1, m], the Unary Resource Constraint for machine Mk ensures that
for any i = j ∈ [1, n], there is no overlap between the processing interval of
Tik and Tjk . We use a simple decomposition with O(n2 ) Boolean variables δkij
(i < j ∈ [1, n]) per machine Mk using the following Disjunctive constraints:

0 ⇔ tik + pik ≤ tjk
δkij = (1)
1 ⇔ tjk + pjk ≤ tik
396 M. Siala et al.

A JSSP requires also a total order on the tasks of each job. We enforce this order
with simple precedence constraints. Last, the objective is to minimize the total
scheduling duration, i.e., the makespan. To this purpose, we have an objective
variable Cmax subject to precedence constraints with the last task of each job.
We use the same search strategy as that proposed in [10,13]. First, we com-
pute a greedy upper bound on the makespan. Then we use a dichotomic search to
reduce the gap between lower and upper bounds. Each step of this dichotomic
phase corresponds to the decision problem with the makespan variable Cmax
fixed to the mid-point value. Each step is given a fixed time cutoff, and exceed-
ing it is interpreted as the instance being unsatisfiable. Therefore, the gap might
not be closed at the end of this phase, so a classic branch and bound procedure
is used until either closing the gap or reaching a global cutoff.
We branch on the Boolean variables of the Disjunctive constraints follow-
ing [13] using the solution guided approach [3] for value ordering. We use the
weighed degree heuristic taskDom/tw in a similar way to [10,13] in addition to
VSIDS [16]. The former branches first on variables occurring most in constraints
triggering failures. The latter favors variables involved in conflict analysis.

4 Lazy Generation of Atoms

In this section we describe an efficient propagator to maintain the consistency
of lazily generated atoms. Recall that domain clauses are likely to be redun-
dant in this case (Section 2). Instead of generating clauses to encode the dif-
ferent relationships between the newly generated atoms, we propose to encode
such relations through a dedicated propagator in order to avoid this redundancy
and hence the associated overhead. We introduce the DomainFaithfulness
constraint. Its role is twofold: firstly, it simulates Unit-Propagation (UP) as if
the atoms were generated eagerly; secondly it performs a complete channeling
between the range variable and all its generated domain atoms.
We consider only the lazy generation of atoms of the type x ≤ u since all
propagators in our models perform only bound tightening operations. Neverthe-
less, the generalization with x = v is quite simple and straightforward. Let x
be a range variable, [v1 , . . . , vn ] be a sequence of integer values, and [b1 . . . bn ] be a
sequence of lazily generated Boolean variables s.t. bi is the atom x ≤ vi . We define
DomainFaithfulness(x, [b1 . . . bn ], [v1 , . . . , vn ]) as follows: ∀i, bi ↔ x ≤ vi .
We describe now the data structures and the procedures that allow to main-
tain Arc Consistency [4] (AC) with a constant amortized time complexity down
a branch of the search tree. We assume without loss of generality that n ≥ 1 and
that vi < vi+1 . The filtering is then organized in two parts:
1. Simulating UP as if the atoms were eagerly generated with all domain
clauses. That is, whenever an atom bi becomes assigned to 1 (respectively
0), the atom bi+1 (respectively bi−1 ) is assigned to 1 (respectively 0).
2. Performing a complete channeling between x and b1 , . . . , bn : We sketch the
process related to the upper bound of x. A similar process in applied with
the lower bound. There are two cases to distinguish:
 Two Clause Learning Approaches for Disjunctive Scheduling 397

(a) Changing the upper bound of x w.r.t. newly assigned atoms: When an
atom bi ↔ x ≤ vi  becomes assigned to 1, we check if vi can be the new
upper bound of x. Note that a failure can be triggered if vi < min(x).
(b) In the case of an upper bound propagation event, one has to assign
some atoms to be consistent with the new upper bound u of x.
Every atom bi such that vi ≥ u has to be assigned to 1. Let iub =
arg maxk (vk | D(bk ) = {1}). Observe first that the part simulating UP
(1) guarantees that all atoms bi where i > iub are already assigned
to 1. Consider now the set ϕ ={biub , biub −1 , biub −2 . . . , blastub } where
lastub = arg mink (vk | vk ≥ u)). It is now sufficient to assign every atom
in ϕ to 1 in order to complete the filtering.
Finally, since down a branch of the search tree the upper bound can only decrease,
we can compute the current value of iub , that is, arg maxk (vk | D(bk ) = {1}) by
exploring the sequence of successive upper bounds from where we previously
stopped. Therefore, each atom is visited at most once down a branch. This filter-
ing can thus be performed in constant amortized time, that is, in O(n) time down a
branch, however, we must store iub and ilb as “reversible” integers.

5 Learning Restricted to Task Ordering

Here we introduce a learning scheme as an alternative to lazy generation for
disjunctive scheduling problems. Recall that we branch only on Boolean variables
coming from the Disjunctive constraints. It follows that every bound literal
p s.t. level(p) > 0 has a non-empty explanation. We can exploit precisely this
property in order to avoid the generation of any bound atom.
The first step in this framework is to perform conflict analysis as usual by
computing the 1-UIP nogood Ψ . Next, we make sure that the latest literal in Ψ
is not a bound literal. Otherwise, we keep explaining the latest literal in Ψ until
having such UIP. We know that this procedure terminates because the worst
case would reach the last decision which corresponds to a UIP that is not a
bound literal. Let Ψ ∗ be the resulting nogood.
Consider now to be the set of bound literals in Ψ ∗ . Instead of generating
the corresponding atoms, we start a second phase of conflict analysis via the
procedure Substitute( , Ψ ) with ( , Ψ ∗ ) as input. Algorithm 2 details this pro-
cedure. In each step of the main loop, a bound literal p from is chosen (Line
1) and replaced in Ψ with its explanation (Line 2). is updated later at each
step in Line 3. The final nogood Ψ involves only Boolean variables. Note that
the backjump level remains the same as in Ψ ∗ since for every p there must exists
a literal in explain(p) with the same level of p.
The advantage of this approach is that since no atom need to be gener-
ated during search, we do not need to maintain the consistency between tasks’
domains and bounds atoms. In particular, it greatly reduces the impact that the
size of the scheduling horizon has on the space and time complexities. Note, how-
ever, that there may be a slight overhead in the conflict analysis step, compared
to the lazy generation mode, since there are more literals to explain. Moreover,
398 M. Siala et al.

Algorithm 2. Substitute( , Ψ )
visited ← ∅ ;
while |
| > 0 do
1 p ← choose p ∈
;
visited ← visited ∪ {p} ;
2 Ψ ← Ψ ∪ {q | q ∈ explain(p) ∧ level(q) > 0} \ visited ;
3
← {q | q ∈ Ψ ∧ q is a bound litteral};
return Ψ ;

since the language of literal is not as rich in this method, shorter proofs may be
possible with the standard approach.

6 Experimental Results

We first compare the two approaches described in this paper, that is, the imple-
mentation of lazy generation using DomainFaithfulness (lazy) as well as the
new learning scheme (disj) against the the CP model described in [11] on two
well known benchmarks for the JSSP: Lawrence [14] and Taillard [23]. Then, we
compare the lower bounds found by our approaches with the best known lower
bounds. All models are implemented within Mistral-2.0 and are tested on Intel
Xeon E5-2640 processors. The source code, the detailed results, and the experi-
mental protocol are available at http://siala.github.io/jssp/details.pdf .
We denote CP(task) the CP model using the taskDom/tw heuristic. For
the hybrid models, we use the notation H(θ, σ) where θ ∈ {vsids, task} is the
variable ordering and σ ∈ {lazy, disj} indicates the type of learning used.

6.1 Empirical Evaluation on the Job Shop Scheduling Problem

We ran every method 10 times using randomized geometric restarts [26]. Each
dichotomy step is limited to 300s and 4×106 propagation calls. The total runtime
for each configuration is limited to 3600s. The results are summarized in Table 1.
Table 1 is divided in two parts. The first part concerns instances that are
mostly proven optimal by our experiments. We report for these instances the
average percentage of calls where optimality was proven %O, the average CPU
time T , and the number of nodes explored by second (nds/s). The second part
concerns the rest of the instances (i.e. hard instances). We report for each data
set the speed of exploration (nds/s) along with the average percentage relative
deviation (PRD) of each model. The PRD of a model m for an instance C is
computed with the formula: 100∗ CmC−C best
best
, where Cm is the minimum makespan
found by model m for this instance ; and Cbest is the minimum makespan known
in the literature [1,2,25]. The bigger the PRD, the less efficient a model is.
Consider first small and medium sized instances, i.e., la-01-40, tai-01-10, and
tai-11-20. Table 1 shows clearly that the learning scheme that we introduced (i.e.
H(vsids/task, disj)) dominates the other models on these instances. For exam-
ple H(vsids, disj) proves 91.5% of Lawrence instances to optimality whereas
 Two Clause Learning Approaches for Disjunctive Scheduling 399

Table 1. Summary of the results

Instances CP(task) H(vsids, disj) H(vsids, lazy) H(task, disj) H(task, lazy)
Mostly proven optimal
%O T nds/s %O T nds/s %O T nds/s %O T nds/s %O T nds/s
la-01-40 87 522 8750 91.5 437 6814 88 632 2332 90.50 434 5218 88.75 509 2694
tai-01-10 89 768 5875 90 517 4975 88 1060 1033 90 634 3572 84 1227 1013
Hard instances
PRD nds/s PRD nds/s PRD nds/s PRD nds/s PRD nds/s
tai-11-20 1.8432 4908 1.1564 3583 1.3725 531 1.2741 2544 1.2824 489
tai-21-30 1.6131 3244 0.9150 2361 1.0841 438 0.9660 1694 0.8745 409
tai-31-40 5.4149 3501 4.0210 2623 3.7350 580 4.0536 1497 3.8844 510
tai-41-50 7.0439 2234 4.8362 1615 4.6800 436 4.9305 1003 5.0136 390
tai-51-60 3.0346 1688 3.2449 2726 3.7809 593 1.1156 1099 1.1675 575
tai-61-70 6.8598 1432 6.5890 2414 5.4264 578 3.9637 866 3.6617 533

CP(task) and H(vsids, lazy) achieve a ratio of 87% and 88%, respectively. More-
over, VSIDS generally performs better than weighted degree on these instances,
although this factor does not seem as important as the type of learning.
Consider now the rest of instances (tai-21-30 to tai-61-70). The impact of
clause learning is more visible on these instances. For example, with tai-31-
40, CP(task) has a PRD of 5.4149 while the worst hybrid model has a PRD
of 4.0536. The choice of the branching strategy seems more important on the
largest instances. For instance, on tai-51-60, the PRD of H(task, disj) is 1.1156
while H(vsids, disj) has a PRD of 3.2449.
Table 1 shows also that the CP model is often faster than any hybrid model
(w.r.t. nds/s). This is expected because of the overhead of propagating the learnt
clauses in the hybrid models. Lazy generation (lazy) slows the search down
considerably compared to the mechanism we introduced (disj).
Regarding the clauses size, we observed that the new method constantly
learns shorter clauses. For example when both methods use the heuristic VSIDS,
the average size on tai-11..20 is 31 for the disjunctive learning whereas the stan-
dard approach has an average size of 43. This may seem surprising, but several
bounds literals may share part of their explanations.
Overall, our hybrid models outperform the CP model in most cases. The
novel conflict analysis shows promising results especially for small and medium
sized instances. It should be noted that we did not find new upper bounds for
hard instances. However, our experiments show that our best model deviates
only of a few percents from the best known bounds in the literature. Note that
the state-of-the-art CP method [25] uses a timeout of 30000s (similarly to [19])
with a maximum of 600s per iteration in the bottom-up phase for improving the
lower bound. Moreover, they use the best known bounds as an extra information
along with a double threading phase. In our experiments, we use a single thread
and a fixed time limit of 3600s per instance.
400 M. Siala et al.

6.2 Lower Bound Computation

We ran another set of experiments with a slightly modified dichotomy phase.

Here we assume that the outcome of step i is satisfiable instead of unsatisfiable
when the cutoff is reached. Each dichotomy step is limited to 1400s and the
overall time limit is 3600s. We used the set of all 22 open Taillard instances before
the results of [25] and we found 7 new lower bounds (cf. Table 2). However, most

Table 2. New Lower Bounds for Taillard Instances

tai13 tai21 tai23 tai25 tai26 tai29 tai30

new old new old new old new old new old new old new old
1305 1282 1613 1573 1514 1474 1544 1518 1561 1558 1576 1525 1515 1485

of these bounds were already improved by Vilı́m et al. [25]. The main difference
is that their algorithm uses the best known bounds as an additional information
before starting search. Our models, however, are completely self-contained in the
sense where search is started from scratch (see Section 3).
We computed the average PRD w.r.t. the best known lower bound includ-
ing those reported in [25]. Excluding the set tai40-50, the best combination,
H(vsids, disj) using Luby restarts [15], finds lower bounds that are 2.78% lower,
in average, than the best known bound in the literature.
Finally, in order to find new lower bounds, we launched again the model
H(vsids, disj) with 2500s as a timeout for each dichotomy step and 7200s as a
global cutoff. We found a new lower bound of 1583 for tai-29 (previously 1573)
and one of 1528 for tai-30 (previously 1519).

7 Conclusion

We introduced two hybrid CP/SAT approaches for solving disjunctive schedul-

ing problems. The first one avoids a known redundancy issue regarding lazy
generation without computational overhead, whereas the second constitutes a
‘hand-crafted’ conflict analysis scheme for disjunctive scheduling.
These methods fit naturally the CDCL framework and, in fact, similar tech-
niques appear to have been implemented in existing hybrid solvers, such as
Chuffed, although they have never been published nor documented. Experi-
mental evaluation shows the promise of these techniques. In particular, the
novel learning scheme is extremely efficient for proving unfeasibility. Indeed, we
improved the best known lower bounds of two Taillard instances.

Acknowledgments. This work is funded by CNRS and Midi-Pyrénées region (grant

number 11050449). The authors would like to thank the Insight Centre for Data Ana-
lytics for kindly granting us access to its computational resources.
 Two Clause Learning Approaches for Disjunctive Scheduling 401

References
1. Best known lower/upper bounds for Taillard Job Shop instances. http://
optimizizer.com/TA.php (accessed April 15, 2015)
2. Taillard, É.: http://mistic.heig-vd.ch/taillard/problemes.dir/ordonnancement.dir/
jobshop.dir/best lb up.txt (accessed April 15, 2015)
3. Christopher, J.: Beck. Solution-guided multi-point constructive search for job shop
scheduling. Journal of Artificial Intelligence Research 29(1), 49–77 (2007)
4. Bessiere, C.: Constraint propagation. In: van Beek, P., Rossi, F., Walsh, T. (eds.)
Handbook of Constraint Programming, volume 2 of Foundations of Artificial Intel-
ligence, pp. 29–83. Elsevier (2006)
5. Boussemart, F., Hemery, F., Lecoutre, C., Sais, L.: Boosting systematic search
by weighting constraints. In: Proceedings of the 16th European Conference on
Artificial Intelligence, ECAI 2004, Valencia, Spain, pp. 146–150 (2004)
6. Carlier, J., Pinson, É.: An algorithm for solving the job-shop problem. Management
Science 35(2), 164–176 (1989)
7. Danna, E.: Structured vs. unstructured large neighborhood search: a case study
on job-shop scheduling problems with earliness and tardiness costs. In: Rossi, F.
(ed.) CP 2003. LNCS, vol. 2833, pp. 817–821. Springer, Heidelberg (2003)
8. Feydy, T., Schutt, A., Stuckey, P.J.: Semantic learning for lazy clause generation.
In: Proceedings of TRICS Workshop: Techniques foR Implementing Constraint
programming Systems, TRICS 2013, Uppsala, Sweden (2013)
9. Feydy, T., Stuckey, P.J.: Lazy clause generation reengineered. In: Gent, I.P. (ed.)
CP 2009. LNCS, vol. 5732, pp. 352–366. Springer, Heidelberg (2009)
10. Grimes, D., Hebrard, E.: Job shop scheduling with setup times and maximal time-
lags: a simple constraint programming approach. In: Lodi, A., Milano, M., Toth, P.
(eds.) CPAIOR 2010. LNCS, vol. 6140, pp. 147–161. Springer, Heidelberg (2010)
11. Grimes, D., Hebrard, E.: Models and strategies for variants of the job shop schedul-
ing problem. In: Lee, J. (ed.) CP 2011. LNCS, vol. 6876, pp. 356–372. Springer,
Heidelberg (2011)
12. Grimes, D., Hebrard, E.: Solving variants of the job shop scheduling prob-
lem through conflict-directed search. INFORMS Journal on Computing 27(2),
268–284 (2015)
13. Grimes, D., Hebrard, E., Malapert, A.: Closing the open shop: contradicting con-
ventional wisdom. In: Gent, I.P. (ed.) CP 2009. LNCS, vol. 5732, pp. 400–408.
Springer, Heidelberg (2009)
14. Lawrence, S.R.: Supplement to resource constrained project scheduling: an experi-
mental investigation of heuristic scheduling techniques. Technical report, Graduate
School of Industrial Administration, Carnegie Mellon University, Pittsburgh, PA
(1984)
15. Luby, M., Sinclair, A., Zuckerman, D.: Optimal speedup of Las Vegas algorithms.
Information Processing Letters 47(4), 173–180 (1993)
16. Matthew, W., Moskewicz, C.F., Madigan, Y.Z., Zhang, L., Malik, S.: Chaff: engi-
neering an efficient SAT solver. In: Proceedings of the 38th Annual Design Automa-
tion Conference, DAC 2001, Las Vegas, Nevada, USA, pp. 530–535 (2001)
17. Nuijten, W.: Time and resource constrained scheduling: a constraint satisfaction
approach. Ph.D thesis, Eindhoven University of Technology (1994)
18. Ohrimenko, O., Stuckey, P.J., Codish, M.: Propagation via Lazy Clause Genera-
tion. Constraints 14(3), 357–391 (2009)
402 M. Siala et al.

19. Pardalos, P.M., Shylo, O.V.: An algorithm for the job shop scheduling problem
based on global equilibrium search techniques. Computational Management Sci-
ence 3(4), 331–348 (2006)
20. Sadeh, N.M.: Lookahead techniques for micro-opportunistic job-shop scheduling.
Ph.D thesis, Carnegie Mellon University, Pittsburgh, PA, USA (1991)
21. Schutt, A., Feydy, T., Stuckey, P.J., Wallace, M.G.: Solving rcpsp/max by lazy
clause generation. Journal of Scheduling 16(3), 273–289 (2013)
22. Shaw, P.: Using constraint programming and local search methods to solve vehicle
routing problems. In: Maher, M.J., Puget, J.-F. (eds.) CP 1998. LNCS, vol. 1520,
pp. 417–431. Springer, Heidelberg (1998)
23. Taillard, É.: Benchmarks for basic scheduling problems. European Journal of Oper-
ational Research 64(2), 278–285 (1993). Project Management anf Scheduling
24. Vilı́m, P.: Edge finding filtering algorithm for discrete cumulative resources in
O(kn log (n)). In: Gent, I.P. (ed.) CP 2009. LNCS, vol. 5732, pp. 802–816. Springer,
Heidelberg (2009)
25. Vilı́m, P., Laborie, P., Shaw, P.: Failure-directed search for constraint-based
scheduling. In: Michel, L. (ed.) CPAIOR 2015. LNCS, vol. 9075, pp. 437–453.
Springer, Heidelberg (2015)
26. Walsh, T.: Search in a small world. In: Proceedings of the 16th International
Joint Conference on Artificial Intelligence, IJCAI 1999, Stockholm, Sweden,
pp. 1172–1177 (1999)
27. Zhang, L., Madigan, C.F., Moskewicz, M.H., Malik, S.: Efficient conflict driven
learning in a boolean satisfiability solver. In: Proceedings of the 2001 IEEE/ACM
International Conference on Computer-aided Design, ICCAD 2001, San Jose, Cal-
ifornia, pp. 279–285 (2001)
Bounding an Optimal Search Path with a Game
of Cop and Robber on Graphs

Frédéric Simard, Michael Morin(B) , Claude-Guy Quimper, François Laviolette,

and Josée Desharnais

Department of Computer Science and Software Engineering, Université Laval,

Québec, QC, Canada
{Frederic.Simard.10,Michael.Morin.3}@ulaval.ca,
{Claude-Guy.Quimper,Francois.Laviolette,Josee.Desharnais}@ift.ulaval.ca

Abstract. In search theory, the goal of the Optimal Search Path (OSP)
problem is to find a finite length path maximizing the probability that
a searcher detects a lost wanderer on a graph. We propose to bound
the probability of finding the wanderer in the remaining search time
by relaxing the problem into a stochastic game of cop and robber from
graph theory. We discuss the validity of this bound and demonstrate its
effectiveness on a constraint programming model of the problem. Exper-
imental results show how our novel bound compares favorably to the
DMEAN bound from the literature, a state-of-the-art bound based on a
relaxation of the OSP into a longest path problem.

Keywords: Optimal search path · Cop and robber · Constraint relax-

ation · Pursuit games

1 Introduction
The Optimal Search Path (OSP) problem [1,2] is a classical problem from search
theory [3] of which the SAROPS [4] and the SARPlan [5] are two examples of
successful systems. The goal of an OSP is to find the best finite length path
a searcher has to take in order to maximize the probability of finding a lost
wanderer moving randomly on a graph. It is a common assumption, in search
theory, to consider unconstrained searcher’s motion. The assumption holds, for
instance, in contexts where the searcher is fast in comparison to the target of
interest. There are, however, many cases, such as in land search and rescue [3],
where the searcher is as slow as the moving wanderer. Washburn [6] also notes
how bounded speeds (for searcher and wanderer) is used in modeling searches
for hostile submarines or friendly vessels.
Most of the early work on the OSP problem consisted in computing bounds
for a depth-first branch-and-bound algorithm initially developed by Stewart [2].1
At the opposite of stochastic constraint programming, the OSP requires to com-
pute a single scenario given by the decision variables and the probability that
1
A review of the bounds developed prior to 1993 is found in the work of Washburn [6].

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 403–418, 2015.
DOI: 10.1007/978-3-319-23219-5 29
404 F. Simard et al.

this scenario works is automatically computed through constraint propagation.

Recent works on the OSP problem include generalization of the problem to con-
sider, for instance, non-uniform travel time [7], searching from a distance [8],
and path dependent detections [9]. Along with a novel OSP generalization, Lau
et al. [7] propose the DMEAN bound. The bound is proved to effectively prune
the branch-and-bound search tree.
In this paper, we present bounding techniques for a constraint programming
(CP) model of the OSP problem introduced in [10]. We discuss, in Section 2, the
DMEAN bound which has never been implemented in CP. Then, as our main
contribution, we develop a bound on the OSP objective function (see Section 3).
This bound, we call the Copwin bound, is obtained by relaxing the OSP into
a game of cop and robber [11–14] which pertains to a well-studied game from
graph theory (see Section 3). In Section 4, we show the benefits of using our
novel Copwin bound by providing experimental results obtained on an existing
CP model of the problem [10]. We conclude in Section 5.

2 The Optimal Search Path Problem

The OSP models a search operation where one searcher searches for a lost, and
possibly moving, search object. The problem is known to be N P-hard [1]. It
is used, in search theory, as a model for path-constrained search operations [3]
and can be formulated as follows. A searcher and a wanderer are moving on
the vertices of a graph G = (V (G), E(G)). Their respective initial positions
correspond to some vertex of V (G) or of some subset of it. Each of their moves,
that we can assume simultaneous, is performed along one of the edges e ∈ E(G)
of the graph and consumes one unit of time. There is a maximal number T of
steps taken for the search operation. The wanderer W is invisible to the searcher
S, but each time they share a vertex v ∈ V (G), the latter has a given probability
of detecting and removing him2 from the graph:

pod(v) := probability of detecting W when both W and S are on vertex v.

Let πt = v1 , v2 , . . . , vt be a path of length t ≤ T representing the t first moves

of the searcher (vi ∈ V (G), i = 1, . . . , t). We define πt[..k] := v1 , v2 , . . . , vk for
k ≤ t. We define the following probabilities for a searcher S and a wanderer W:

pocπt(v) := probability3 of W being on vertex v at time t and not having

previously been detected, given that S’s first t steps were along πt ;
cosπt := Cumulative probability Of Success of detecting W up to time t
t
when S follows πt , i.e., it is given as i=1 pocπt[..t−1] (vi )pod(vi ) ,
where pocπt[..0] is the initial distribution of the wanderer;

2
Following the convention of cop and robber games on graphs, we suppose that the
searcher (or cop) is a woman and that the wanderer (or robber) is a man.
3
poct (v) also stands for the probability of containment [10].
 Bounding an Optimal Search Path 405

cos∗πt := maximal cumulative probability of success up to T if S’s t first

steps are along πt , that is, max{cosπT | πT has prefix πt } .
The goal of the searcher is to find the path πT with maximal cumulative
probability of success cosπT . This probability is actually cos∗π0 , where π0 is the
empty path.
The wanderer is passive in that he does not react to the searcher’s moves. It
is the usual assumption to model the wanderer’s motion by using a Markovian
transition matrix M with source vertices on rows and destination vertices on
columns. Thus, M (r, r ) is the probability of a wanderer’s move from vertex r to
vertex r within one time step. Each time the searcher searches a vertex r along
path πt and the search operation does not stop, then either the wanderer is not
there, or he is undetected; the latter happens with probability 1 − pod(r ). That
is, as long as the search goes on, the wanderer is not found and the knowledge
about his location evolves according to:
pocπt(r ) = pocπt[..t−1](vt )(1 − pod(vt ))M (vt , r ) (1)

+ pocπt[..t−1](r)M (r, r ).
r∈V (G)\{vt }

The next section summarizes the CP model of the OSP used in [10].

2.1 Modeling the OSP in Constraint Programming

A CP model for the OSP problem follows directly from the problem’s definition.
We define, for each step t (1 ≤ t ≤ T ), a decision variable PATHt that represents
the searcher’s position. The PATHt domain is an enumerated domain such that
dom(PATHt ) = V (G). The searcher’s path is constrained by a possible limited
choice for its initial position PATH1 and by the fact that
(PATHt−1 , PATHt ) ∈ E(G) (2)
for any steps t ∈ {2, . . . , T }. The encoding in the CP-framework is done via a
sequence of Table constraints [15].
We define two sets of implicit variables with bounded domains for probabil-
ities: POCt (v) with dom(POCt (v)) = [0, 1], and POSt (v) with dom(POSt (v)) =
[0, 1] for 1 ≤ t ≤ T and for v ∈ V (G). If the searcher is located in a vertex v,
then she obtains a success probability as modeled by the following constraint set
defined over all time steps t ∈ {1, . . . , T } and vertices v ∈ V (G), where pod(v)
is the known probability of detection in vertex v:
PATHt = v ⇒ POSt (v) = POCt (v)pod(v). (3)
If the searcher is not located in vertex v at step t, then the probability of detecting
the object in v at step t is null. This is modeled by the following constraint set
defined over all time steps t ∈ {1, . . . , T } and vertices v ∈ V (G):
PATHt = v ⇒ POSt (v) = 0. (4)
406 F. Simard et al.

The value of the POC1 (r) variables is fixed given the initial distribution on
the wanderer’s location, a value known for all r ∈ V (G). For all subsequent
time steps t ∈ {2, . . . , T } and vertices v ∈ V (G), the evolution of the searcher’s
knowledge on the wanderer’s location is modeled by the following constraint set:

POCt (v) = M (r, r ) (POCt−1 (r) − POSt−1 (r)) . (5)
r∈V (G)

A bounded domain variable Z with dom(Z) = [0, 1] is added to represent the

objective value to maximize which corresponds to the cumulative overall prob-
ability of success as defined above by cosπT where πT is an assignment to the
PATHt variables. Following [10], we observe that the searcher is in one vertex at
a time leading to
T

Z= max POSt (v) (6)
v∈V (G)
t=1

and improving filtering over the usual double sum definition of the objective.
The objective is finally
max Z, (7)
PATH1 ,...,PATHT

subject to Constraints (2) to (6) which leads to cos∗π0 value.

2.2 The DMEAN Bound from the Literature

Martins [16] proposes to bound the objective function using the MEAN bound.
Lau et al. [7] improve on this bound by presenting the discounted MEAN bound
(DMEAN) which tightens the MEAN at a reasonable computational cost. In
the literature, these bounds are usually applied in an OSP-specific depth-first
branch-and-bound algorithm. In this section, we show how to use the DMEAN
bound in our OSP CP model. We chose to implement DMEAN among other
bounds from the literature since both DMEAN and our novel bound are based
on graph theory problems. While DMEAN aims at solving a longest path prob-
lem, the Copwin bound, as will be explained, is based on pursuit games. Just as
DMEAN, MEAN is based on the maximization of the expected number of detec-
tions in the remaining time. The core idea of the MEAN bound is to construct
a directed acyclic graph (DAG) at time t ∈ {1, . . . , T } on vertices consisting of
pairs (r, k) where r ∈ V (G) and t < k ≤ T . A directed edge is added from vertex
(r, k) to vertex (r , k + 1) if and only if (r, r ) ∈ E(G). The maximal expected
number of detections ZDAG from step t to T is obtained by following the longest
path on the DAG, which can be computed in time linear on |V (G)| + |E(G)|.
Given the beginning πt of a path, it is sufficient to know pocπt (v) for all v ∈ V (G)
to be able to compute the expected detection at step k with t ≤ k ≤ T in vertex
r ∈ V (G). An incident edge into a node (r , k + 1) is weighted as follows:

RMEAN (r , k + 1) = pocπt (v)M k+1−t (v, r ) pod(r ) (8)
v∈V (G)
 Bounding an Optimal Search Path 407

where M k+1−t is the transition matrix M to the power of k + 1 − t.

The DMEAN bound is based on the additional observation that a wanderer
moving from vertex r to r has survived a detection on r to possibly being
captured on vertex r . On each outgoing edge from a node (r, k) into a node
(r , k + 1), if k = t we let the weight become RDMEAN (r , k + 1) = RMEAN (r , k +
1). Otherwise, if k > t, we let:

RDMEAN (r , k + 1) = RMEAN (r , k + 1) − RMEAN (r, k)M (r, r )pod(r ). (9)

That is, the wanderer has a probability RMEAN (r , k + 1) of reaching vertex
r on time step k + 1 and being detected there, from which we substract his
probability RMEAN (r, k)M (r, r )pod(r ). This last term is the probability the
wanderer had of being on vertex r at the preceding time step k, transiting on r
and being captured there. The RDMEAN (r , k + 1) value thus corresponds to the
probability that the wanderer reaches r, survives an additional search, reaches
r and gets caught. Given the beginning of a path πt , an admissible4 bound for
the OSP is obtained by first solving a longest path problem of which ZDAG is
the optimal value and then by summing ZDAG and cosπt [7].
Remark 1. Given a graph with n vertices and m edges, DMEAN asks for the
construction of a DAG with nT new vertices and mT new edges. Then, if the
current time is k, the longest path can be solved in O ((T − k + 1)(n + m)) steps.
It is convenient, to apply the DMEAN bound in our CP model of the OSP,
to order the path variables in a static order of the searcher’s positions in time.
This is the natural order to solve an OSP since the PATHt variables are the only
decision variables and since the implicit variables representing the probabilities,
including the objective variable Z that represents cos∗π0 , are entirely determined
by the searcher’s path. When opening a node for a searcher’s position at a time
t, we are able to compute, using the chosen bound, a tighter upper bound on the
domain of the objective variable. The DMEAN bound is proved admissible [7],
that is, it never underestimates the real objective value. Whenever the bound
computed on the opened node of the searcher’s position at a time t is lower than
the best known lower bound on the objective value (e.g., the objective value of
the current incumbent solution) that searcher’s move is proven unpromising.

3 Bounding the Optimal Search Path Using Search

Games on Graphs
The cop and robber game on a graph consists in finding a winning strategy for
a cop to catch a robber, considering perfect information for both players (i.e.,
each player knows her/his position and her/his adversary’s position). We focus
on a recent variant where the robber is random (or drunk) [17–19]. We show
how a valid upper bound on the objective function of the OSP can be derived
4
A bound is admissible if and only if it never underestimate (resp. overestimate) the
value of the objective to maximize (resp. minimize).
408 F. Simard et al.

by considering, after t steps, the best possible scenario for the searcher that
corresponds to allowing her the ability of seeing the wanderer for the remainder
of the game.

3.1 A Game of Cop and Drunk Robber

In this game, a cop and a robber move in turn on a graph G, the cop moving
first. In contrast with the OSP, the cop has probability 1 of catching the robber
when sharing the same vertex, and she sees the robber, who, as does the wan-
derer in the previous game, walks randomly according to a stochastic transition
matrix M . The value M (r, r ) is the probability that the robber moves to r pro-
vided he is in r at the beginning of the turn. It is positive only if (r, r ) ∈ E(G).
The game ends whenever the cop and the robber share the same vertex. The
following definition generalizes Nowakowski and Winkler’s [11] relational charac-
terization for the classic cop and robber game to the stochastic one. We want to
define the relation wnM (r, c) as standing for the probability that the cop catches
the robber within n moves given their positions (r, c) and the robber’s random
movement model M . Its definition is based upon the observation that in order to
maximize her capture probability, the cop only needs to average her probability
of capture at the next turn on the robber’s possible transitions.
Definition 1. Given r, c the respective positions of the robber and the cop in G,
M the robber’s random walk matrix, we define:
w0M (r, c) := 1 if r = c; otherwise, it is 0;
⎧
⎨1  if c ∈ N [r], n ≥ 1;
wnM (r, c) := M (r, r )wn−1
M
(r , c ) if c ∈
/ N [r], n ≥ 1. (10)
⎩ max
c ∈N [c]
r  ∈N [r]

where N [z] := {z} ∪ {v ∈ V (G) | (v, z) ∈ E(G)} is the closed neighbourhood of

z.
The following proposition gives sense to the previous definition.

Proposition 1. If the cop plays first on G and M governs the robber’s random
walk, then wnM (r, c) is the probability a cop starting on vertex c captures the
drunk robber on his start vertex r in n steps or less.

Proof. By induction. The base case, with n = 0, is clear. Suppose the proposition
holds for n − 1 ≥ 0 with n > 0 and let us now prove it for n. If c ∈ N [r], then
wnM (r, c) = 1 and the result follows because the cop indeed catches the robber. If
c∈/ N [r], then let the cop move to some vertex c . The position of the robber at
the end of the round is r with probability M (r, r ). The probability that the cop
catches the robber
 depends on thisMlast one’s next move and on wn−1 following
the expression r ∈N [r] M (r, r )wn−1 (r , c ). Hence, the best possible move for

the cop is argmaxc ∈N [c] r ∈N [r] M (r, r )wn−1
M
(r , c ). The wanted probability is
thus wnM (r, c).

 Bounding an Optimal Search Path 409

Proposition 1 could provide a trivial upper bound on the probability of find-

ing the wanderer, but a tighter one is presented in the following section. The
bound we will derive is inspired from Proposition 1 but is closer to the OSPs
objective function.

3.2 Markov Decision Process (MDP) and the OSP Bound

In Section 3.1, we presented a game where the robber is visible, a setting that
we will use as a bound on the OSP’s objective function. We formulate this
game as a MDP from which we derive the Copwin bound and the proof that
it is admissible. As a recall, a MDP [20–22] is a stochastic decision process
generalizing Markov chains. In a MDP, an agent evolves in a state space S.
Every time the agent finds itself in a state s ∈ S, it takes an action a in a
set A of possible actions. The system then randomly transits to another valid
state s according to a given transition probability P [s | s, a] (where | denotes
a conditional). In order to guide the agent’s choice of actions, a reward function
R : S × A × S → R is defined that assigns to every triple (s, a, s ) a real number.
The goal of the agent is then to maximize its expected reward within a time
bound T . We first formulate the game of cop and robber defined above as an
MDP and then deduce a valid OSP bound. The following definition presents the
MDP; its solution encodes the optimal strategy for the cop.
Definition 2 (MDP Cop and Drunk Defending Robber). Let G be a
graph, M a stochastic transition matrix on V (G) and T the maximal number of
time steps. We define a MDP M = (S, A, P, R) as follows. A state is a triplet
(r, c, t) consisting in both positions of the robber and the cop r, c ∈ V (G) in
addition to the current time t ∈ {1, . . . , T + 1}.

S := (V (G) ∪ {jail}) × V (G) × {1, 2, . . . , T + 1}.

The set of actions is V (G), the vertices on which the cop moves.

A := V (G).

From a pair of positions of the cop and the robber at a certain time, once the
cop chooses an action another state is chosen randomly with probability P .
 
P (r , c , t ) | (r, c, t), a := 0 whenever a = c or c ∈ N [c] or t = t + 1

otherwise, P is defined as : ⎧
⎪
⎪ 1 if r = r = jail;
  ⎪
⎨pod(r) if r = c , r = jail;
P (r , c , t + 1) | (r, c, t), c := (11)
⎪
⎪

(1 − pod(r))M (r, r ) if r = c , r = jail;
⎪
⎩
M (r, r ) if r ∈ {c, c , jail}.

1 if r = jail = r, t ≤ T ;
R((r , c , t ) | (r, c, t), a) := (12)
0 otherwise.
410 F. Simard et al.

The game is initialized as follows: the cop chooses her initial position c on a
subset X of the graph vertices, and then the initial position r of the robber is
picked at random according to the probability distribution pocπ0, which results
in an initial state (r, c, 1) for the MDP. A turn starts with a cop move. If the
cop transits to the robber state (r = c ), then there is probability pod(r) that
she catches the robber, which results in the robber going to jail (r = jail) and
staying there for the rest of the game. The cop then receives a reward of 1.
If the catch fails (r = jail, with probability 1 − pod(r)) or if the cop did not
transit to the robber state (r = c ), the robber is still free to roam, following
M . Note that the state transition probabilities (11) are non-null only when valid
moves are considered (time goes up by one and a = c ∈ N [c]). Note also that,
when the robber is in jail, no more reward is given to the cop. In the MDP M,
the cop’s goal is to find a strategy, also called policy, to maximize her expected
reward, that is, the probability to capture the robber before a total of T steps
is reached. A strategy in M consists in first choosing an initial position (for
the cop) and then in following a function f : S → A that, given the current
state, tells the cop which action to take, that is, which state to transit to. Note
that, since the current position of the robber is included in the current state
of the MDP, the cop therefore has full information on the system when she is
elaborating her strategy.
Because of M’s Markov property, whenever a strategy f is fixed, one can com-
pute the value uf (r, c, t) of each state (r, c, t) of the MDP, that is, the expected
reward the cop can obtain from that state when following the strategy f . The
optimal strategy, noted u∗ , is therefore the one that gives the highest value on all
states (r, c, t). The cop’s optimal strategy consists in moving to the robber’s posi-
tion if possible, and then trying to capture him. If the robber is not positioned
on one of the cop’s neighbours, the cop moves to the position that maximizes her
probability of capture in the remaining time allowed. Similarly to Proposition 1,
the value of this optimal strategy is:
⎧ 
⎪
⎪
⎪ max M (r, r )u∗ (r , c , t + 1) if r ∈
/ N [c], t < T ;
⎨c ∈N [c] r ∈N [r]


u∗ (r, c, t) = (13)
⎪
⎪ 1 − (1 − pod(r))T +1−t if r ∈ N [c], t ≤ T ;
⎪
⎩0 if r ∈
/ N [c], t = T.

If r ∈ N [c], the cop, who moves first, must choose a next state that will result in
the best probability of eventual capture, given the robber’s present position, and
knowing that the robber’s next move is governed by M . If r ∈ N [c], the cop tries
to catch the robber with probability of success pod(r); if she fails, the robber
will transit to one of his neighbours, and hence the cop can keep on trying to
catch the robber until success or until the maximal time has been reached. It is
important to note here that the robber is completly visible and the game is not
played simultaneously, hence why the cop can follow the robber. Equation (13)
is analogous to (10) with time steps reversed and pod(r) = 1 for all vertices.
The formula follows from the fact that at the beginning of time step t, the cop
has T + 1 − t remaining attempts.
 Bounding an Optimal Search Path 411

Since the optimal probability of capture in the OSP problem is always lower
than the optimal probability of capture in the cop and robber game, we have
the following proposition:
Proposition 2. The probability cos∗π0 of finding the wanderer is always at most
that of catching the robber:

cos∗π0 ≤ max pocπ0(r)u∗ (r, c, 1),
c∈X
r∈V (G)

where X is the subset of V (G) of possible initial positions of the cop.

Proof. Clearly, cos∗π0 is bounded by the optimal probability of capture of the cop
and robber game. In the MDP, the optimal probability of capture is obtained
if the cop’s first choice maximizes his probability of capture considering that at
that moment the robber is not yet positioned on the graph but will be according
to the probability distribution pocπ0 .


Unfortunately, Proposition 2’s bound is of no use in a branch-and-bound attempt

for solving the OSP problem, because a bound for each πt and each t = 0, . . . , T
is needed. The next proposition generalizes it appropriately.

Proposition 3. Let πt = v1 , v2 , . . . , vt . Then


cos∗πt ≤ cosπt + max pocπt(r )u∗ (r , c , t + 1), (14)
c ∈N [vt ]
r  ∈V (G)

where N [v0 ] is interpreted as the possible initial positions of the cop.

Proof. As in the preceding proof, cos∗πt is bounded by the optimal reward

obtained when first playing the OSP game along πt and then (if the wanderer is
not yet detected) switching to the cop and robber game: this is done by making
the wanderer (robber) visible to the searcher (cop). When starting this second
phase (at the t + 1 step), the cop must choose his next position in order to max-
imize the probability of capture; at this moment, the robber is not yet visible

but his next position
 is governed by poc πt . If the cop chooses c , his probability
∗  
of capture will be r ∈V (G) pocπt (r )u (r , c , t + 1) and the result follows. 

Remark 2. An important aspect of Copwin is its ability to be mostly precom-

puted. For any vertices r, c ∈ V (G) and time t ∈ {1, . . . , T + 1}, the values
u∗ (r, c, t) are computed recursively (starting with t = T ) and then stored.
Then, when the bound is called the next value of Equation (14) depends on
the neighbours of the searcher’s position and the vertices of the graph, requiring
O ((Δ + 1)n) extra operations on an n vertex graph of maximal degree Δ. Since
DMEAN’s complexity is O ((T − k + 1)(n + m)), Copwin is faster on most time
steps and on many instances.
412 F. Simard et al.

We note that Pralat and Kehagias [18] also formulated the game of cop and
visible drunk robber as an MDP, but instead of using this framework to compute
the probability of capture of the robber it was formulated to obtain the expected
number of time steps before capture.
Applying the Copwin bound in CP requires filtering the upper bound of the
objective function variable Z. The solver can be configured to branch in a static
order of the PATHt variable, which is, as discussed in Section 2.2, the natural
order of the decision variables to solve an OSP. We proved that our bound is
admissible in that it never underestimates the real objective value. In the next
section, we use the Copwin bound jointly with a CP model of the OSP and
compare its performance to that of the DMEAN bound from literature.

4 Experiments

All experiments were run on the supercomputer Colosse from Université Laval
using the Choco 2.1.5 solver [23] along with the Java Universal Network/Graph
(JUNG) 2.0.1 framework [24]. The total allowed CPU time is 5 minutes (or 5
million backtracks) per run. We used a total of six graphs representing different
search environments. In the grid G+ , the nodes are connected to their North,
South, East, and West neighbour. The grid G∗ has the same connections than
G+ but also adds connectivity to the North-East, North-West, South-East, and
South-West neighbours. The last grid is a typical hexagonal grid we call GH
with the searcher moving on its edges. Grid-like accessibility constraints are
common in search theory problems in that they model aerial or marine search
operations where the searcher has few accessibility constraints but her own phys-
ical constraints. This justifies the presence of grids in our OSP benchmarks. We
also added a fourth grid with supplementary constraints on accessibility. This
last grid GV is a hexagonal grid with a supplementary node in the center of
each hexagon. We then had removed edges randomly to simulate supplementary
accessibility constraints. The last two graphs, GL and GT , are real underground
environments. The GL graph is the graph of the Université Laval’s tunnels. The
GT graph is the graph of the London Underground subway (also known as the
Tube). Both neighbours of an intermediate station, i.e., a node of degree 2, were
connected together while the node was removed. This practice results in a more
interesting graph environment for searches.
For each instance, we selected a p ∈ {0.3, 0.6, 0.9} and put pod(r) = p
for all vertex r. For the wanderer’s Markovian transition matrix, we pick
ρ ∈ {0.3, 0.6, 0.9} and considered that at any position v, the wanderer has prob-
ρ
ability ρ of staying in v and probability 1 − deg(v) to move to one of his deg(v)
neighbours. We uniformly distributed the remaining probability mass 1−ρ on the
neighbouring vertices. Finally, we chose T ∈ {5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25}.
This leads to a total of 594 instances.
We conducted the experiment in three parts. In the first two parts, we com-
pare our novel bound to the DMEAN bound we reviewed in Section 2.2. First,
we implemented the two bounds as separate constraints on the upper bound of
 Bounding an Optimal Search Path 413

the objective value. Both bounds are evaluated on the same problem instances.
This enables us to assess the bounds performance with respect to the objective
value of the best incumbent solution found by the solver and to the time required
to attain it.
In the second part, we implemented both bounds as a single constraint on
the upper bound of the objective value. Both bounds were computed simultane-
ously by the constraint. The minimal value of both bounds was used to update
the upper bound of the domain of the Z variable representing the objective if
required. That is, whenever that minimal value is lower than the upper bound
on the domain of the Z variable. If the domain of Z is emptied, then one of the
bounds produced a value that is lower than the best known lower bound on the
objective thus leading to a backtrack. For each bounding technique (DMEAN
and Copwin), we count how many times, during the search, the bounding of
the objective value causes a backtrack. We gave a single point to a bound that
leads to a backtrack that way. We note that on occasions both bounds could
receive a point. In all cases, the solver branches on the PATHt decision variables
in their natural order, i.e., in ascending order of the time steps. We thus were
able to record the exact time step at which each backtrack occurred. Such an
experiment enables one to empirically assess which bound is more efficient at
pruning the search tree no matter without considering any time constraints.
We show, in the third part, how the bounds perform when paired with the
total detection (TD) heuristic used as a value-selection heuristic [10]. The TD
heuristic is on many account similar to the bound derived in this paper which
gives its theoretical justification. However, rather than using the cop and robber
games theory to bound the objective value, TD is a heuristic that assigns values
to the variables PATHt . Even in this context, the Copwin bound performs and
stays competitive.

4.1 Results and Discussion

Figure 1 shows the best incumbent objective value (COS) for all OSP problem
instances. Hence, COS is the estimation of cos∗π0 at the end of the solving process.
Each dot represents the solution to a single OSP instance. The figure compares
the objective value obtained when the solver uses the DMEAN bound against
the objective value obtained when the solver uses the Copwin bound for all
instances. The closer a dot is to the diagonal line, the closer it is to being the
same value for both bounds which is viewed as both bounds having had the same
effectiveness on this instance. It does appear the Copwin bound tends to lead
the solver to better objective values in the total allowed time. Thus, Copwin is
not only faster than DMEAN, it is sharp enough to help the solver to exploit
more promising branches.
Figure 2 shows the number of times each bound empties the domain of the
objective value variable in the search tree for each type of graph. We chose
to plot the score dependant on the time step (or level in the search tree) to
better analyse where branches were being pruned. It appears that on average
414 F. Simard et al.

Fig. 1. Two by two comparison of DMEAN best incumbent COS value against Copwin
best incumbent COS value for all 594 OSP instances; the incumbent solution of each
instance is represented by a dot on the graph.

Fig. 2. Total number of cuts in the solver’s search tree for both bounds against
searcher’s path length with T = 17; results for each instance are aggregated on motion
models ρ and on detection models pod(r). refers to the use of Copwin whereas
refers to DMEAN.
 Bounding an Optimal Search Path 415

Copwin empties the objective value variable’s domain more often than DMEAN
independently of the graph, even though on some instances such as GL both lead
to a great number of prunings. This is a strong indication that Copwin performs
well on all accounts on most instances: for a fixed resolution time, it leads to
better objective values while pruning more branches than the DMEAN bound.
We present on Figure 3, the objective value of the best so far solution against
computational time for both bounding strategies. We present these comparisons
on three representative instances: one in which Copwin is better than DMEAN,

0.100 0.5

0.6
0.4
0.075
COS value

COS value
COS value

0.3
0.4
0.050
0.2

0.2
0.025
0.1

0.0 0.000 0.0

0 100 200 300 0 100 200 300 0 100 200 300
Time (ms) Time (ms) Time (ms)

(a) GT , 0.9, 0.9 (b) G+ , 0.9, 0.6 (c) GL , 0.3, 0.9

Fig. 3. The COS value achieved and the time it was found for each bound. Red lines
are associated with Copwin and blue ones with DMEAN. Each title refers to the graph
type, the instance pod(r), the instance ρ and maximal time steps of 17.

Fig. 4. Comparison of the best incumbent COS value obtained with the TD heuristic
for the Copwin bound against the best COS value obtained with the TD heuristic for
the DMEAN bound for all OSP problem instances; the incumbent solution of each
instance is represented by a dot on the graph.
416 F. Simard et al.

Table 1. Results on OSP problem instances; bold values are better.

pod(r) ρ Objective (COS) Time† (s) Backtracks

DMEAN Copwin DMEAN Copwin DMEAN Copwin
GT with T = 17
0.3 0.3 0.111 0.113 298 197 3578 2791
0.6 0.139 0.129 200 144 2045 2157
0.9 0.362 0.354 275 208 2973 3304
0.9 0.3 0.252 0.258 131 287 1400 4613
0.6 0.318 0.325 259 236 3082 3544
0.9 0.736 0.736 292 185 3543 2706
G+ with T = 17
0.3 0.3 0.011 0.014 96 37 283 118
0.6 0.007 0.036 88 281 243 1307
0.9 0.001 0.022 90 111 276 497
0.9 0.3 0.030 0.077 95 296 290 1394
0.6 0.020 0.103 95 272 283 1229
0.9 0.001 0.047 26 262 19 1174
GV with T = 17
0.3 0.3 0.035 0.079 154 296 2510 7240
0.6 0.038 0.105 146 206 2510 4922
0.9 0.036 0.229 158 196 2510 4898
0.9 0.3 0.099 0.215 159 240 2510 5539
0.6 0.112 0.304 155 264 2510 6040
0.9 0.109 0.531 161 219 2510 4996
GL with T = 17
0.3 0.3 0.344 0.318 273 145 9224 7749
0.6 0.417 0.359 287 139 9666 7476
0.9 0.516 0.480 297 122 9298 6462
0.9 0.3 0.638 0.651 278 119 11304 6249
0.6 0.713 0.713 289 146 11603 7890
0.9 0.811 0.833 249 162 9489 8727
G∗ with T = 17
0.3 0.3 0.069 0.072 270 286 1264 2264
0.6 0.107 0.117 273 183 1261 1482
0.9 0.235 0.324 276 281 1254 2461
0.9 0.3 0.192 0.205 274 252 1264 2077
0.6 0.304 0.333 264 219 1261 1842
0.9 0.671 0.711 259 231 1253 1925
GH with T = 17
0.3 0.3 0.023 0.087 258 241 522 618
0.6 0.015 0.122 255 277 519 742
0.9 0.001 0.318 250 233 514 623
0.9 0.3 0.064 0.227 274 258 522 686
0.6 0.043 0.342 260 286 520 719
0.9 0.003 0.816 248 280 516 680
† The time to last incumbent solution.

one in which it is worse and another where both bounds lead to a similar perfor-
mance. These graphs confirm the intuition that, most of the time, Copwin finds
good solutions more quickly than DMEAN.
 Bounding an Optimal Search Path 417

Figure 4 summarizes further experiments where Copwin and DMEAN were

paired with the TD heuristic. Even though both bounds lead the solver to
greater objective values, we observe that Copwin’s objective values are on aver-
age slightly better. Thus, the behavior observed on Figure 1 is preserved when
a good heuristic is added.
As an addition to the graphs of Figures 1 to 4, we include Table 1 which
contains the results of the experiments of the Copwin versus the DMEAN bound
for all graphs with T = 17.

5 Conclusion
We tackled the OSP problem from search theory using constraint programming.
As a first contribution, we provided the first implementation of the DMEAN
bound from the search theory literature in CP. As our main contribution, we
developed the Copwin bound, a novel and competitive bound based on MDPs.
This bound is derived from a simple and elegant relaxation of a search problem
into a graph theory pursuit game. Involving a polynomial computational cost,
the Copwin bound leads to an improved solver performance on the vast majority
of OSP problem instances in our benchmark. Altough we used the bound on a
CP model of the OSP, it remains a general technique applicable to other OSP
algorithms.

Acknowledgments. This research is funded by the Natural Sciences and Engineering

Research Council of Canada (NSERC) and the Fonds de recherche du Québec - Nature
et technologies (FRQ-NT).
Computations were made on the supercomputer Colosse from Université Laval,
managed by Calcul Québec and Compute Canada. The operation of this supercomputer
is funded by the Canada Foundation for Innovation (CFI), NanoQubec, RMGA and
the Fonds de recherche du Québec - Nature et technologies (FRQ-NT).

References
1. Trummel, K., Weisinger, J.: The complexity of the optimal searcher path problem.
Operations Research 34(2), 324–327 (1986)
2. Stewart, T.: Search for a moving target when the searcher motion is restricted.
Computers and Operations Research 6(3), 129–140 (1979)
3. Stone, L.: Theory of Optimal Search. Academic Press, New York (2004)
4. Netsch, R.: The USCG search and rescue optimal planning system (SAROPS)
via the commercial/joint mapping tool kit (c/jmtk). In: Proceedings of the 24th
Annual ESRI User Conference, vol. 9, August 2004
5. Abi-Zeid, I., Frost, J.: A decision support system for canadian search and rescue
operations. European Journal of Operational Research 162(3), 636–653 (2005)
6. Washburn, A.R.: Branch and bound methods for a search problem. Naval Research
Logistics 45(3), 243–257 (1998)
7. Lau, H., Huang, S., Dissanayake, G.: Discounted MEAN bound for the optimal
searcher path problem with non-uniform travel times. European Journal of Oper-
ational Research 190(2), 383–397 (2008)
418 F. Simard et al.

8. Morin, M., Lamontagne, L., Abi-Zeid, I., Lang, P., Maupin, P.: The optimal
searcher path problem with a visibility criterion in discrete time and space.
In: Proceedings of the 12th International Conference on Information Fusion,
pp. 2217–2224 (2009)
9. Sato, H., Royset, J.O.: Path optimization for the resource-constrained searcher.
Naval Research Logistics, 422–440 (2010)
10. Morin, M., Papillon, A.-P., Abi-Zeid, I., Laviolette, F., Quimper, C.-G.: Constraint
programming for path planning with uncertainty. In: Milano, M. (ed.) CP 2012.
LNCS, vol. 7514, pp. 988–1003. Springer, Heidelberg (2012)
11. Nowakowski, R., Winkler, P.: Vertex-to-vertex pursuit in a graph. Discrete Math-
ematics 43(2), 235–239 (1983)
12. Quilliot, A.: Problème de jeux, de point fixe, de connectivité et de représentation
sur des graphes, des ensembles ordonnés et des hypergraphes. Ph.D thesis, Univer-
sité de Paris VI (1983)
13. Bonato, A., Nowakowski, R.: The game of cops and robbers on graphs. American
Mathematical Soc. (2011)
14. Fomin, F.V., Thilikos, D.M.: An annotated bibliography on guaranteed graph
searching. Theoretical Computer Science 399(3), 236–245 (2008)
15. Beldiceanu, N., Demassey, S.: Global constraint catalog (2014). http://sofdem.
github.io/gccat/ (accessed 2015–04)
16. Martins, G.H.: A new branch-and-bound procedure for computing optimal search
paths. Master’s thesis, Naval Postgraduate School (1993)
17. Kehagias, A., Mitsche, D., Pralat, P.: Cops and invisible robbers: The cost of
drunkenness. Theoretical Computer Science 481, 100–120 (2013)
18. Kehagias, A., Pralat, P.: Some remarks on cops and drunk robbers. Theoretical
Computer Science 463, 133–147 (2012)
19. Komarov, N., Winkler, P.: Capturing the Drunk Robber on a Graph. arXiv preprint
arXiv:1305.4559 (2013)
20. Bäuerle, N., Rieder, U.: Markov Decision Processes with Applications to Finance.
Springer (2011)
21. Puterman, M.L.: Markov decision processes: discrete stochastic dynamic program-
ming, vol. 414. John Wiley & Sons (2009)
22. Barto, A.G., Sutton, R.S.: Reinforcement learning: An introduction. MIT press
(1998)
23. Laburthe, F., Jussien, N.: Choco solver documentation. École de Mines de Nantes
(2012)
24. O’Madadhain, J., Fisher, D., Nelson, T., White, S., Boey, Y.B.: The JUNG (Java
universal network/graph) framework (2010)
 Restricted Path Consistency Revisited

Kostas Stergiou(B)

Department of Informatics and Telecommunications Engineering,

University of Western Macedonia, Kozani, Greece
kstergiou@uowm.gr

Abstract. Restricted path consistency (RPC) is a strong local consis-

tency for binary constraints that was proposed 20 years ago and was
identified as a promising alternative to arc consistency (AC) in an early
experimental study of local consistencies for binary constraints. How-
ever, and in contrast to other strong local consistencies such as SAC
and maxRPC, it has been neglected since then. In this paper we revisit
RPC. First, we propose RPC3, a new lightweight RPC algorithm that is
very easy to implement and can be efficiently applied throughout search.
Then we perform a wide experimental study of RPC3 and a light version
that achieves an approximation of RPC, comparing them to state-of-the-
art AC and maxRPC algorithms. Experimental results clearly show that
restricted RPC is by far more efficient than both AC and maxRPC when
applied throughout search. These results strongly suggest that it is time
to reconsider the established perception that MAC is the best general
purpose method for solving binary CSPs.

1 Introduction
Restricted path consistency (RPC) is a local consistency for binary constraints
that is stronger than arc consistency (AC). RPC was introduced by Berlandier
[4] and was further studied by Debruyne and Bessiere [7,8]. An RPC algorithm
removes all arc inconsistent values from a domain D(x), and in addition, for any
pair of values (a, b), with a ∈ D(x) and b ∈ D(y) s.t. b is the only support for a
in a D(y), it checks if (a, b) is path consistent. If it is not then a is removed from
D(x). In this way some of the benefits of path consistency are retained while
avoiding its high cost.
Although RPC was identified as a promising alternative to AC as far back
as 2001 [8], it has been neglected by the CP community since then. In contrast,
stronger local consistencies such as max restricted path consistency (maxRPC)
[7] and singleton arc consistency (SAC) [8] have received considerable attention
in the past decade or so [1–3,5,9,11,13,14]. However, despite the algorithmic
developments on maxRPC and SAC, none of the two outperforms AC when
maintained during search, except for specific classes of problems. Therefore,
MAC remains the predominant generic algorithm for solving binary CSPs.
In this paper we revisit RPC and make two contributions compared to pre-
vious works that bring the state-of-the-art regarding RPC up to date. The first
is algorithmic and the second experimental.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 419–428, 2015.
DOI: 10.1007/978-3-319-23219-5 30
420 K. Stergiou

The two algorithms that have been proposed for RPC, called RPC1 [4] and
RPC2 [7], are based on the AC algorithms AC4 and AC6 respectively. As a result
they suffer from the same drawbacks as their AC counterparts. Namely, they use
heavy data structures that are too expensive to maintain during search. In recent
years it has been shown that in the case of AC lighter algorithms which sacrifice
optimality display a better performance when used inside MAC compared to
optimal but heavier algorithms such as AC4, AC6, AC7, and AC2001/3.1. Hence,
the development of the residue-based version of AC3 known as AC3r [10,12]. A
similar observation has been made with respect to maxRPC [1]. Also, it has
been noted that cheap approximations of local consistencies such as maxRPC
and SAC are more cost-effective than the full versions. In the case of maxRPC,
the residue-based algorithm lmaxRPC3r , which achieves an approximation of
maxRPC, is the best choice when applying maxRPC [1].
Following these trends, we propose RPC3, an RPC algorithm that makes use
of residues in the spirit of ACr and lmaxRPCr and is very easy to implement. As
we will explain, for each constraint (x, y) and each value a ∈ D(x), RPC3 stores
two residues that correspond to the two most recently discovered supports for a
in D(y). This enables the algorithm to avoid many redundant constraint checks.
We also consider a restricted version of the algorithm (simply called rRPC3)
that achieves a local consistency property weaker than RPC, but still stronger
than AC, and is considerably faster in practice.
Our second and most important contribution concerns experiments. Given
that the few works on RPC date from the 90s, the experimental evaluations of the
proposed algorithms were carried out on limited sets of, mainly random, prob-
lems. Equally importantly, there was no evaluation of the algorithms when used
during search to maintain RPC. We carry out a wide evaluation on benchmark
problems from numerous classes that have been used in CSP solver competi-
tions. Surprisingly, results demonstrate that an algorithm that applies rRPC3
throughout search is not only competitive with MAC, but it clearly outperforms
it on the overwhelming majority of tested instances, especially on structured
problems. Also, it clearly outperforms lmaxRPC3r . This is because RPC, and
especially its restricted version, achieves a very good balance between the prun-
ing power of maxRPC and the low cost of AC.
Our experimental results provide strong evidence of a local consistency that
is clearly preferable to AC when maintained during search. Hence, perhaps it is
time to reconsider the common perception that MAC is the best general purpose
solver for binary problems.

2 Background
A Constraint Satisfaction Problem (CSP) is defined as a triplet (X , D, C) where:
X = {x1 , . . . , xn } is a set of n variables, D = {D(x1 ), . . . , D(xn )} is a set of
domains, one for each variable, with maximum cardinality d, and C = {c1 , . . . , ce }
is a set of e constraints. In this paper we are concerned with binary CSPs. A
binary constraint cij involves variables xi and xj .
 Restricted Path Consistency Revisited 421

At any time during the solving process if a value ai has not been removed
from the domain D(xi ), we say that the value is valid. A value ai ∈ D(xi ) is arc
consistent (AC) iff for every constraint cij there exists a value aj ∈ D(xj ) s.t.
the pair of values (ai , aj ) satisfies cij . In this case aj is called an support of ai .
A variable is AC iff all its values are AC. A problem is AC iff there is no empty
domain in D and all the variables in X are AC.
A pair of values (ai , aj ), with ai ∈ D(xi ) and aj ∈ D(xj ), is path consistent
PC iff for any third variable xk constrained with xi and xj there exists a value
ak ∈ D(xk ) s.t. ak is a support of both ai and aj . In this case aj is a PC-support
of ai in D(xj ) and ak is a PC-witness for the pair (ai , aj ) in D(xk ).
A value ai ∈ D(xi ) is restricted path consistent (RPC) iff it is AC and for each
constraint cij s.t. ai has a single support aj ∈ D(xj ), the pair of values (ai , aj )
is path consistent (PC) [4]. A value ai ∈ D(xi ) is max restricted path consistent
(maxRPC) iff it is AC and for each constraint cij there exists a support aj for
ai in D(xj ) s.t. the pair of values (ai , aj ) is path consistent (PC) [7]. A variable
is RPC (resp. maxRPC) iff all its values are RPC (resp. maxRPC). A problem
is RPC (resp. maxRPC) iff there is no empty domain and all variables are RPC
(resp. maxRPC).

3 The RPC3 Algorithm

The RPC3 algorithm is based on the idea of seeking two supports for a value,
which was first introduced in RPC2 [7]. But in contrast to RPC2 which is based
on AC6, it follows an AC3-like structure, resulting in lighter use of data struc-
tures, albeit with a loss of optimality. As explained below, we can easily obtain
a restricted but more efficient version of the algorithm that only approximates
the RPC property. Crucially, the lack of heavy data structures allows for the
use of the new algorithms during search without having to perform expensive
restorations of data structures after failures.
In the spirit of ACr , RPC3 utilizes two data structures, R1 and R2 , which
hold residual data used to avoid redundant operations. Specifically, for each
constraint cij and each value ai ∈ D(xi ), R1xi ,ai ,xj and R2xi ,ai ,xj hold the two
most recently discovered supports of ai in D(xj ). Initially, all residues are set to
a special value NIL, considered to precede all values in any domain.
The pseudocode of RPC3 is given in Algorithm 1 and Function 2. Being
coarse-grained like AC3, Algorithm 1 uses a propagation list Q, typically imple-
mented as a fifo queue. We use a constraint-oriented description, meaning that Q
handles pairs of variables involved in constraints. A variable-based one requires
minor modifications.
Once a pair of variables (xi , xj ) is removed from Q, the algorithm iterates
over D(xi ) and for each value ai first checks the residues R1xi ,ai ,xj and R2xi ,ai ,xj
(line 5). If both are valid then ai has at least two supports in D(xj ). Hence,
the algorithm moves to process the next value in D(xi ). Otherwise, function
findTwoSupports is called. This function will try to find two supports for ai in
D(xj ). In case it finds none then ai is not AC and will thus be deleted (line 13).
422 K. Stergiou

Algorithm 1. RPC3:boolean
1: while Q = ∅ do
2: Q ← Q−{(xi , xj )};
3: Deletion ← FALSE;
4: for each ai ∈ D(xi ) do
5: if both R1xi ,ai ,xj and R2xi ,ai ,xj are valid then
6: continue;
7: else
8: if only one of R1xi ,ai ,xj and R2xi ,ai ,xj is valid then
9: R ← the valid residue;
10: else
11: R ← NIL;
12: if findTwoSupports(xi , ai , xj , R) = FALSE then
13: remove ai from D(xi );
14: Deletion ← TRUE;
15: if D(xi ) = ∅ then
16: return FALSE;
17: if Deletion = TRUE then
18: for each (xk , xi ) ∈ C s.t. (xk , xi ) ∈ / Q do
19: Q ← Q ∪ {(xk , xi )};
20: for each (xl , xk ) ∈ C s.t. xl = xi and (xl , xi ) ∈ C and (xl , xk ) ∈
/ Q do
21: Q ← Q ∪ {(xl , xk )};
22: return TRUE;

In case it finds only one then it will check if ai is RPC. If it is not then it will
be deleted. Function findTwoSupports takes as arguments the variables xi and
xj , the value ai , and a parameter R, which is set to the single valid residue of
ai in D(xj ) (line 9) or to NIL if none of the two residues is valid.
Function findTwoSupports iterates over the values in D(xj ) (line 3). For each
value aj ∈ D(xj ) it checks if the pair (ai , aj ) satisfies constraint cij (this is what
function isConsistent does). If both residues of ai in D(xj ) are not valid then
after a support is found, the algorithm continues to search for another one.
Otherwise, as soon as a support is found that is different than R, the function
returns having located two supports (lines 9-11).
If only one support aj is located for ai then the algorithm checks if the
pair (ai , aj ) is path consistent. During this process it exploits the residues to
save redundant work, if possible. Specifically, for any third variable xk that is
constrained with both xi and xj , we first check if one of the two residues of ai is
valid and if aj is consistent with that residue (line 16). If this is the case then we
know that there is a PC-witness for the pair (ai , aj ) in D(xk ) without having to
iterate over D(xk ). If it is not the case then the check is repeated for the residues
of aj in D(xk ). If we fail to verify the existense of a PC-witness in this way then
we iterate over D(xk ) checking if any value ak is consistent with both ai and aj .
If a PC-witness is found, we proceed with the next variable that is constrained
with both xi and xj . Otherwise, the function returns false, signaling that ai is
not RPC.
 Restricted Path Consistency Revisited 423

Function 2. findTwoSupports(xi , ai , xj , R):Boolean

1: if R = NIL then oneSupport ← FALSE;
2: else oneSupport ← TRUE;
3: for each aj ∈ D(xj ) do
4: if isConsistent(ai , aj ) then
5: if oneSupport = FALSE then
6: oneSupport ← TRUE;
7: R1xi ,ai ,xj ← aj ;
8: else
9: if aj = R then
10: R2xi ,ai ,xj ← aj ;
11: return TRUE;
12: if oneSupport = FALSE then
13: return FALSE
14: else
15: for each xk ∈ X, xk = xi and xk = xj , s.t. (xk , xi ) ∈ C and (xk , xj ) ∈ C do
16: if there is a valid residue R∗xi ,ai ,xk and isConsistent(Rx∗ i ,ai ,xk , aj ) or if there
is a valid residue R∗xj ,aj ,xk and isConsistent(Rx∗ j ,aj ,xk , ai )
17: then continue;
18: PCwitness ← FALSE;
19: for each ak ∈ D(xk ) do
20: if isConsistent(ai , ak ) and isConsistent(aj , ak ) then
21: PCwitness ← TRUE;
22: break;
23: if PCwitness = FALSE then
24: return FALSE;
25: return TRUE;

Moving back to Algorithm 1, if at least one value is deleted from a domain

D(xi ), some pairs of variables must be enqueued so that the deletions are prop-
agated. Lines 18-19 enqueue all pairs of the form (xk , xi ). This ensures that if a
value in a domain D(xk ) has lost its last support in D(xi ), it will be processed
by the algorithm when the pair (xk , xi ) is dequeued, and it will be removed. In
addition, it ensures that if a value in D(xk ) has been left with only one support
in D(xi ), that is not a PC-support, it will be processed and deleted once (xk , xi )
is dequeued. This means that if we only enqueue pairs of the form (xk , xi ), we
can achieve stronger pruning than AC. However, this is not enough to achieve
RPC. We call the version of RPC3 that only enqueues such pairs restricted RPC3
(rRPC3).
To achieve RPC, for each pair (xk , xi ) that is enqueued, we also enqueue all
pairs of the form (xl , xk ) s.t. xl is constrained with xi . This is because after the
deletions from D(xi ) the last PC-witness in D(xi ) for some pair of values for
variables xk and xl may have been deleted. This may cause further deletions
from D(xl ).
424 K. Stergiou

The worst-case time complexity of RPC3, and rRPC3, is O(ned3 )1 . The space
complexity is determined by the space required to store the residues, which is
O(ed). The time complexities of algorithms RPC1 and RPC2 are O(ned3 ) and
O(ned2 ) respectively, while their space complexities, for stand-alone use, are
O(ed2 ) and O(end). RPC3 has a higher time complexity than RPC2, and a lower
space complexity than both RPC1 and RPC2. But most importantly, RPC3
does not require the typically quite expensive restoration of data structures
after failures when used inside search. In addition, this means that its space
complexity remains O(ed) when used inside search, while the space complexities
of RPC1 and RPC2 will be even higher than O(ed2 ) and O(end).

4 Experiments

We have experimented with 17 classes of binary CSPs taken from C.Lecoutre’s

XCSP repository: rlfap, graph coloring, qcp, qwh, bqwh, driver, job shop,
haystacks, hanoi, pigeons, black hole, ehi, queens, geometric, composed, forced
random, model B random. A total of 1142 instances were tested. Details about
these classes of problems can be found in C.Lecoutre’s homepage. All algorithms
used the dom/wdeg heuristic for variable ordering [6] and lexicographic value
ordering. The experiments were performed on a FUJITSU Server PRIMERGY
RX200 S7 R2 with Intel(R) Xeon(R) CPU E5-2667 clocked at 2.90GHz, with 48
GB of ECC RAM and 16MB cache.
We have compared search algorithms that apply rRPC3 (resp. RPC3) during
search to a baseline algorithm that applies AC (i.e. MAC) and also to an algo-
rithm that applies lmaxRPC. AC and lmaxRPC were enforced using the ACr
and lmaxRPC3 algorithms respectively. For simplicity, the four search algorithms
will be denoted by AC, rRPC, RPC, and maxRPC hereafter. Note that a MAC
algorithm with ACr and dom/wdeg for variable ordering is considered as the
best general purpose solver for binary CSPs.
A timetout of 3600 seconds was imposed on all four algorithms for all the
tested instances. Importantly, rRPC only timed out on instances where
AC and maxRPC also timed out. On the other hand, there were several
cases where rRPC finished search within the time limit but one (or both) of AC
and maxRPC timed out. There were a few instances where RPC timed out while
rRPC did not, but the opposite never occured.
Table 1 summarizes the results of the experimental evaluation for specific
classes of problems. For each class we give the following information:

– The mean node visits and run times from non-trivial instances that were
solved by all algorithms within the time limit. We consider as trivial any
instance that was solved by all algorithms in less than a second.
– Node visits and run time from the single instance where AC displayed its
best performance compared to rRPC.

1
The proof is quite simple but it is omitted for space reasons.
 Restricted Path Consistency Revisited 425

– Node visits and run time from the single instance where maxRPC displayed
its best performance compared to rRPC.
– Node visits and run time from a representative instance where rRPC dis-
played good performance compared to AC, excluding instances where AC
timed out.
– The number of instances where AC, RPC, maxRPC timed out while rRPC
did not. This information is given only for classes where at least one such
instance occured.
– The number of instances where AC, rRPC, RPC, or maxRPC was the win-
ning algorithm, excluding trivial instances.

Comparing AC to rRPC we can note the following. rRPC is more efficient in

terms of mean run time performance on all classes of structured problems with
the exception of queens. The difference in favor of rRPC can be quite stunning,
as in the case of qwh and qcp. The numbers of node visits in these classes suggest
that rRPC is able to achieve considerable extra pruning, and this is then reflected
on cpu times.
Importantly, in all instances of 16 classes (i.e. all classes apart from queens)
AC was at most 1.7 times faster than rRPC. In contrast, there were 7 instances
from rlfap and 12 from graph coloring where AC timed out while rRPC finished
within the time limit. The mean cpu time of rRPC on these rlfap instances
was 110 seconds while on the 12 graph coloring instances the cpu time of rRPC
ranged from 1.8 to 1798 seconds. In addition, there were numerous instances
where rRPC was orders of magnitude faster than AC. This is quite common in
qcp and qwh, as the mean cpu times demonstrate, but such instances also occur
in graph coloring, bqwh, haystacks and ehi.
Regarding random problems, rRPC achieves similar performance to AC on
geometric (which is a class with some structure) and is slower on forced ran-
dom and model B random. However, the differences in these two classes are not
significant. The only class where there are significant differences in favour of
AC is queens. Specifically, AC can be up to 5 times faster than rRPC on some
instances, and orders of magnitude faster than both RPC and maxRPC. This
is because all algorithms spend a lot of time on propagation but evidently the
strong local consistencies achieve little extra pruning. Importantly, the low cost
of rRPC makes its performance reasonable compared to the huge run times of
RPC and maxRPC.
The comparison between RPC and AC follows the same trend as that of
rRPC and AC, but importantly the differences in favour of RPC are not as large
on structured problems where AC is inefficient, while AC is clearly faster on
random problems, and by far superior on dense structured problems like queens
and pigeons.
Comparing the mean performance of rRPC to RPC and maxRPC we can
note that rRPC is more efficient on all classes. There are some instances where
RPC or/and maxRPC outperform rRPC due to their stronger pruning, but the
differences in favour of RPC and maxRPC are rarely significant. In contrast,
rRPC can often be orders of magnitude faster. An interesting observation that
426 K. Stergiou

Table 1. Node visits (n), run times in secs (t), number of timeouts (#TO) (if appli-
cable), and number of wins (winner) in summary. The number in brackets after the
name of each class gives the number of instances tested.

class AC rRPC RPC maxRPC

(n) (t) (n) (t) (n) (t) (n) (t)
rlfap (24)
mean 29045 64.1 11234 32.3 10878 39.0 8757 134.0
best AC 12688 9.3 11813 14.5 10048 18.2 5548 33.2
best maxRPC 8405 10.1 3846 4.4 3218 6.86 1668 4.8
good rRPC 19590 28.8 5903 8.2 5197 10.4 8808 23.7
#TO 7 3 6
winner 1 18 1 0
qcp (60)
mean 307416 345.4 37725 44.8 43068 167.1 49005 101.3
best AC 36498 63.5 36286 73.7 57405 354.3 63634 173.5
best maxRPC 20988 16.8 7743 7.6 4787 11.9 1723 1.8
good rRPC 1058477 761 65475 53.5 67935 162.9 54622 63.1
winner 2 8 0 4
qwh (40)
mean 205232 1348.2 20663 46.2 28694 177.9 24205 64.7
best AC 6987 4.5 3734 3.0 5387 11.9 3174 3.2
best maxRPC 231087 461.4 30926 72.3 30434 187.6 13497 35.8
good rRPC 445771 859.6 35923 79.5 56965 375.4 37582 103.9
winner 0 9 0 6
bqwh (200)
mean 28573 7.6 7041 2.4 5466 2.9 6136 2.7
best AC 5085 1.2 4573 1.4 4232 2.0 3375 1.3
best maxRPC 324349 85.3 122845 46.1 56020 33.8 64596 29.3
good rRPC 83996 22.5 7922 2.6 10858 5.6 9325 4.2
winner 2 36 13 20
graph coloring (177)
mean 322220 88.6 261882 60.4 192538 73.7 263227 138.79
best AC 1589650 442.5 1589650 743.8 1266416 930.8 1589650 1010.0
best maxRPC 1977536 647.9 1977536 762.4 1265930 613.7 1977536 759.8
good rRPC 31507 189.1 3911 15.6 2851 18.2 10477 62.7
#TO 12 1 5
winner 8 35 17 0
geometric (100)
mean 111611 58.4 54721 58.6 52416 97.3 38227 190.9
best AC 331764 203.1 169871 218.1 160428 358.1 113878 696.2
best maxRPC 67526 28.1 31230 28.1 30229 53.5 20071 73.6
good rRPC 254304 123.3 119248 117.4 117665 203.0 84410 363.3
winner 12 11 1 0
forced random (20)
mean 348473 143.5 197994 177.2 191114 309.8 154903 455.4
best AC 1207920 491.5 677896 596.7 654862 1040.4 538317 1551.4
best maxRPC 26729 7.6 12986 9.0 12722 13.5 9372 22.4
good rRPC 455489 201.6 270345 258.5 262733 462.0 207267 651.2
winner 20 0 0 0
model B random (40)
mean 741124 194.3 383927 224.5 361926 346.5 28871 1044.9
best AC 2212444 669.8 1197369 805.7 1136257 1283.8 - TO
best maxRPC 345980 81.2 181127 94.4 171527 142.7 130440 405.8
good rRPC 127567 32.4 43769 24.4 41328 39.1 51455 171.5
#TO 0 0 14
winner 39 1 0 0
queens (14)
mean 2092 14.2 797 59.2 2476 1032.2 953 2025.2
best AC 150 9.2 149 43.0 149 499.1 149 3189.2
best maxRPC 7886 10.8 2719 11.5 9425 910.5 3341 367.9
good rRPC 7886 10.8 2719 11.5 9425 910.5 3341 367.9
#TO 0 0 1
winner 4 0 0 0
 Restricted Path Consistency Revisited 427

requires further investigation is that in some cases the node visits of rPRC are
fewer than RPC and and/or maxRPC despite the weaker pruning. This usually
occurs on soluble instances and suggests that the interaction with the dom/wdeg
heuristic can guide search to solutions faster.
Finally, the classes not shown in Table 1 mostly include instances that are
either very easy or very hard (i.e. all algorithms time out). Specifically, instances
in composed and hanoi are all trivial, and the ones in black hole and job shop are
either trivial or very hard. Instances in ehi typically take a few seconds for AC
and under a second for the other three algorithms. Instances in haystacks are
very hard except for a few where AC is clearly outperformed by the other three
algorithms. For example, in haystacks-04 AC takes 8 seconds and the other three
take 0.2 seconds. Instances in pigeons are either trivial or very hard except for
a few instances where rRPC is the best algorithm followed by AC. For example
on pigeons-12 AC and rRPC take 709 and 550 seconds respectively, while RPC
and maxRPC time out. Finally, driver includes only 7 instances. Among them,
3 are trivial, rRPC is the best algorithm on 3, and AC on 1.

5 Conclusion

RPC was recognized as a promising alternative to AC but has been neglected for
the past 15 years or so. In this paper we have revisited RPC by proposing RPC3,
a new algorithm that utilizes ideas, such as residues, that have become standard
in recent years when implementing AC or maxRPC algorithms. Using RPC3 and
a restricted variant we performed the first wide experimental study of RPC when
used inside search. Perhaps surprisingly, results clearly demostrate that rRPC3
is by far more efficient than state-of-the-art AC and maxRPC algorithms when
applied during search. This challenges the common perception that MAC is the
best general purpose solver for binary CSPs.

References
1. Balafoutis, T., Paparrizou, A., Stergiou, K., Walsh, T.: New algorithms for max
restricted path consistency. Constraints 16(4), 372–406 (2011)
2. Balafrej, A., Bessiere, C., Bouyakh, E., Trombettoni, G.: Adaptive singleton-based
consistencies. In: Proceedings of the Twenty-Eighth AAAI Conference on Artificial
Intelligence, pp. 2601–2607 (2014)
3. Barták, R., Erben, R.: A new algorithm for singleton arc consistency. In:
Proceedings of the Seventeenth International Florida Artificial Intelligence,
pp. 257–262 (2004)
4. Berlandier, P.: Improving domain filtering using restricted path consistency. In:
Proceedings of IEEE CAIA 1995, pp. 32–37 (1995)
5. Bessiere, C., Cardon, S., Debruyne, R., Lecoutre, C.: Efficient Algorithms for Sin-
gleton Arc Consistency. Constraints 16, 25–53 (2011)
6. Boussemart, F., Hemery, F., Lecoutre, C., Sais, L.: Boosting systematic search by
weighting constraints. In: Proceedings of ECAI 2004, Valencia, Spain (2004)
428 K. Stergiou

7. Debruyne, R., Bessière, C.: From restricted path consistency to max-restricted

path consistency. In: Smolka, G. (ed.) CP 1997. LNCS, vol. 1330, pp. 312–326.
Springer, Heidelberg (1997)
8. Debruyne, R., Bessière, C.: Domain Filtering Consistencies. JAIR 14, 205–230
(2001)
9. Grandoni, F., Italiano, G.F.: Improved algorithms for max-restricted path con-
sistency. In: Rossi, F. (ed.) CP 2003. LNCS, vol. 2833, pp. 858–862. Springer,
Heidelberg (2003)
10. Lecoutre, C., Hemery, F.: A study of residual supports in arc cosistency. In:
Proceedings of IJCAI 2007, pp. 125–130 (2007)
11. Lecoutre, C., Prosser, P.: Maintaining singleton arc consistency. In: 3rd Interna-
tional Workshop on Constraint Propagation and Implementation (CPAI 2006),
pp. 47–61 (2006)
12. Likitvivatanavong, C., Zhang, Y., Bowen, J., Shannon, S., Freuder, E.: Arc consis-
tency during search. In: Proceedings of IJCAI 2007, pp. 137–142 (2007)
13. Prosser, P., Stergiou, K., Walsh, T.: Singleton consistencies. In: Dechter, R. (ed.)
CP 2000. LNCS, vol. 1894, pp. 353–368. Springer, Heidelberg (2000)
14. Vion, J., Debruyne, R.: Light algorithms for maintaining max-RPC during search.
In: Proceedings of SARA 2009 (2009)
 Machine Learning of Bayesian Networks
Using Constraint Programming

Peter van Beek(B) and Hella-Franziska Hoffmann

Cheriton School of Computer Science, University of Waterloo,

Waterloo, ON N2L 3G1, Canada
vanbeek@cs.uwaterloo.ca

Abstract. Bayesian networks are a widely used graphical model with

diverse applications in knowledge discovery, classification, prediction,
and control. Learning a Bayesian network from discrete data can be
cast as a combinatorial optimization problem and there has been much
previous work on applying optimization techniques including proposals
based on ILP, A* search, depth-first branch-and-bound (BnB) search,
and breadth-first BnB search. In this paper, we present a constraint-based
depth-first BnB approach for solving the Bayesian network learning prob-
lem. We propose an improved constraint model that includes powerful
dominance constraints, symmetry-breaking constraints, cost-based prun-
ing rules, and an acyclicity constraint for effectively pruning the search
for a minimum cost solution to the model. We experimentally evaluated
our approach on a representative suite of benchmark data. Our empir-
ical results compare favorably to the best previous approaches, both in
terms of number of instances solved within specified resource bounds and
in terms of solution time.

1 Introduction
Bayesian networks are a popular probabilistic graphical model with diverse appli-
cations including knowledge discovery, classification, prediction, and control (see,
e.g., [1]). A Bayesian network (BN) can either be constructed by a domain expert
or learned automatically from data. Our interest here is in the learning of a BN
from discrete data, a major challenge in machine learning. Learning a BN from
discrete data can be cast as a combinatorial optimization problem—the well-
known score-and-search approach—where a scoring function is used to evaluate
the quality of a proposed BN and the space of feasible solutions is systemati-
cally searched for a best-scoring BN. Unfortunately, learning a BN from data is
NP-hard, even if the number of parents per vertex in the DAG is limited to two
[2]. As well, the problem is unlikely to be efficiently approximatable with a good
quality guarantee, thus motivating the use of global (exact) search algorithms
over local (heuristic) search algorithms [3].
Global search algorithms for learning a BN from data have been studied
extensively over the past several decades and there have been proposals based on
dynamic programming [4–6], integer linear programming (ILP) [7,8], A* search

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 429–445, 2015.
DOI: 10.1007/978-3-319-23219-5 31
430 P. van Beek and H.-F. Hoffmann

[9–11], depth-first branch-and-bound (BnB) search [12,13], and breadth-first

BnB search [10,11,14,15]. In this paper, we present a constraint-based depth-first
BnB approach for solving the Bayesian network learning problem. We propose
an improved constraint model that includes powerful dominance constraints,
symmetry-breaking constraints, cost-based pruning rules, and an acyclicity con-
straint for effectively pruning the search for a minimum cost solution to the
model. We experimentally evaluated our approach on a representative suite of
benchmark data. Our empirical results compare favorably to the best previous
approaches, both in terms of number of instances solved within specified resource
bounds and in terms of solution time.

2 Background

In this section, we briefly review the necessary background in Bayesian networks

before defining the Bayesian network structure learning problem (for more back-
ground on these topics see, for example, [16,17]).
A Bayesian network (BN) is a probabilistic graphical model that consists of a
labeled directed acyclic graph (DAG) in which the vertices V = {v1 , . . . , vn } cor-
respond to random variables, the edges represent direct influence of one random
variable on another, and each vertex vi is labeled with a conditional probability
distribution P (vi | parents(vi )) that specifies the dependence of the variable vi
on its set of parents parents(vi ) in the DAG. A BN can alternatively be viewed as
a factorized representation of the joint probability distribution over the random
variables and as an encoding of conditional independence assumptions.
The predominant method for BN structure learning from data is the score-
and-search method1 . Let G be a DAG over random variables V , and let I =
{I1 , . . . , IN } be a set of multivariate discrete data, where each instance Ii is
an n-tuple that is a complete instantiation of the variables in V . A scoring
function σ(G | I) assigns a real value measuring the quality of G given the data
I. Without loss of generality, we assume that a lower score represents a better
quality network structure.

Definition 1. Given a discrete data set I = {I1 , . . . , IN } over random variables

V and a scoring function σ, the Bayesian network structure learning problem
is to find a directed acyclic graph G over V that minimizes the score σ(G | I).

Scoring functions commonly balance goodness of fit to the data with a

penalty term for model complexity to avoid overfitting. Common scoring func-
tions include BIC/MDL [18,19] and BDeu [20,21]. An important property of
these (and all commonly used) scoring functions is decomposability, where the

1
An alternative method, called constraint-based structure learning in the literature, is
based on statistical hypothesis tests for conditional independence. We do not discuss
it further here except to note that the method is known to scale to large instances
but to have the drawback that it is sensitive to a single failure in a hypothesis test
(see [17, p. 785]).
 Machine Learning of Bayesian Networks Using Constraint Programming 431

A: {D}, 9.6 {C}, 9.9 {E}, 10.0 {}, 15.4

B: {C, D}, 12.1 {C}, 12.2 {E}, 12.3 {}, 14.1
E
C: {E}, 3.6 {D}, 5.2 {A, B}, 10.9 {A}, 11.4 {}, 17.0
D: {E}, 3.6 {C}, 5.2 {A, B}, 10.9 {A}, 11.4 {}, 17.0 C D
E: {D}, 3.7 {A}, 4.2 {A, B}, 11.2 {C}, 11.6 {}, 17.0
B

(a) (b)

Fig. 1. (a) Random variables and possible parent sets for Example 1; (b) minimum
cost DAG structure with cost 38.9.

score of the entire network σ(G | I) can be rewritten as the sum of local scores
 n
i=1 σ(vi , parents(vi ) | I) that only depend on vi and the parent set of vi in
G. Henceforth, we assume that the scoring function is decomposable and that,
following previous work, the local score σ(vi , p | I) for each possible parent set
p ⊆ 2V −{vi } and each random variable vi has been computed in a preprocessing
step prior to the search for the best network structure. Pruning techniques can
be used to reduce the number of possible parent sets that need to be considered,
but in the worst-case the number of possible parent sets for each variable vi is
2n−1 , where n is the number of vertices in the DAG.

Example 1. Let A, B, C, D, and E be random variables with the possible parent

sets and associated scores shown in Figure 1(a). For example, if the parent set
{C, D} for random variable B is chosen there would be a directed edge from C
to B and a directed edge from D to B and those would be the only incoming
edges to B. The local score for this parent set is 12.1. If the parent set {} for
random variable A is chosen, there would be no incoming edges to A; i.e., A
would be a source vertex. Figure 1(b) shows the minimum cost DAG with cost
15.4 + 4.2 + 3.6 + 3.6 + 12.1 = 38.9.

3 Constraint Programming Approach

In this section, we present a constraint model and a depth-first branch-and-
bound solver for the Bayesian network structure learning problem. Table 1 sum-
marizes the notation.
Our constraint model consists of vertex variables, ordering variables, depth
variables, and constraints over those variables. The ordering and depth variables,
although redundant, improve the search for a solution.
Vertex (possible parent set) variables. There is a vertex variable vi , i ∈ V ,
for each random variable in V and the domain of vi , dom(vi ), consists of the
possible parent sets for vi . The assignment vi = p denotes that vertex vi has
parents p in the DAG; i.e., the vertex variables represent the DAG over the set
of random variables V . Associated with each domain value is a cost and the goal
432 P. van Beek and H.-F. Hoffmann

Table 1. Notation for specifying constraint model.

V set of random variables

n number of random variables in the data set
cost(v) cost (score) of variable v
dom(v) domain of v
parents(v) set of parents of v in the DAG
min(dom(v)) the minimum value in the domain of v
v1 , . . . , vn vertex (possible parent set) variables
o1 , . . . , on ordering (permutation) variables
d1 , . . . , dn depth variables
depth(p | o1 , . . . , oi−1 ) depth of p ∈ dom(vj ), where vj occurs at position i in the ordering

is to minimize the total cost, cost(v1 ) + · · · + cost(vn ), subject to the constraint

that the graph is acyclic. A global constraint is introduced to enforce that the
vertex variables form a DAG,

acyclic(v1 , . . . , vn ), (1)

where the constraint is satisfied if and only if the graph designated by the parent
sets is acyclic. The DAG is not necessarily connected. A satisfiability checker for
the acyclic constraint is given in Section 3.7, which in turn can be used to
propagate the constraint.
Ordering (permutation) variables. There is an ordering variable oi for each
random variable and dom(oi ) = V , the set of random variables. The assignment
oi = j denotes that vertex vj is in position i in the total ordering of the variables.
The ordering variables represent a permutation of the random variables. A global
constraint is introduced to enforce that the order variables form a permutation
of the vertex variables,
alldifferent(o1 , . . . , on ). (2)
The alldifferent constraint is propagated by, whenever a variable becomes instan-
tiated, simply removing that value from the domains of the other variables.
Depth variables. There is a depth variable di for each random variable and
dom(di ) = {0, ..., n − 1}. The depth variables and the ordering variables are in
one-to-one correspondence. The assignment di = k denotes that the depth of the
vertex variable vj that occurs at position i in the ordering is k, where the depth
is the length of the longest path from a source vertex to vertex vj in the DAG.

Example 2. A constraint model for Example 1 would have variables vA , . . ., vE ,

o1 , . . ., o5 , and d1 , . . ., d5 , with domains dom(vA ) = {{C}, {D}, {E}, {}}, . . .,
dom(vE ) = {{D}, {A}, {A, B}, {C}, {}}, dom(oi ) = {A, . . . , E}, and dom(di ) =
{0, . . . , 4}.

To more formally state additional constraints, we introduce the following

notation for the depth of a domain value p for a vertex variable vj .
 Machine Learning of Bayesian Networks Using Constraint Programming 433

Definition 2. The depth of a domain value p for a vertex variable vj that

occurs at position i in the ordering, denoted depth(p | o1 , . . . , oi−1 ), is defined
as: 0 if p = {}; one plus the maximum depth of the elements of p if each element
of p occurs in a parent set of a vertex variable earlier in the ordering; and ∞
otherwise.

Example 3. In Example 2, suppose that o1 = A, vA = {}, d1 = 0, o2 = E,

vE = {A}, and d2 = 1. The value of depth(p | o1 , o2 ) for variable C is 0 if p = {},
1 if p = {A}, 2 if p = {E}, and ∞ if p = {D} or p = {A, B}.

Constraints 3 & 4 establish the correspondence between the three types of

variables,

∀j • ∀p • vj = p ⇐⇒ ∃!i • oi = j ∧ di = depth(p | o1 , . . . , oi−1 ), (3)

∀i • ∀j • oi = j ⇐⇒ ∃!p • vj = p ∧ di = depth(p | o1 , . . . , oi−1 ), (4)

where the indices i and j range over 1 ≤ i, j ≤ n and the value p ranges over
dom(vj ). The constraints are propagated as follows. A value p ∈ dom(vj ) can
be pruned iff ∀i • j ∈ dom(oi ) ⇒ depth(p | o1 , . . . , oi−1 ) ∈ dom(di ). A value
j ∈ dom(oi ) can be pruned iff ∀p ∈ dom(vj ) • depth(p | o1 , . . . , oi−1 ) ∈ dom(di ).
Only bounds are maintained on the depth variables. Hence, the notation
depth(p | o1 , . . . , oi−1 ) ∈ dom(di ) is to be interpreted as depth(p | o1 , . . . , oi−1 ) <
min(dom(di )) ∨ depth(p | o1 , . . . , oi−1 ) > max(dom(di )). When propagating
Constraints 3 & 4, we must determine depth(p | o1 , . . . , oi−1 ). In general, this
is a difficult problem. We restrict ourselves to two easy special cases: (i) all of
o1 , . . . , oi−1 have been instantiated, or (ii) some of o1 , . . . , oi−1 have been instan-
tiated and all of the p ∈ dom(vj ) are subsets of these ordering variables. We leave
to future work further ways of safely approximating the depth to allow further
propagation.

Example 4. In Example 2, suppose that o1 = A, vA = {}, d1 = 0, o2 = E,

vE = {A}, d2 = 1, and that, as a result of some propagation, min(di ) = 1,
i = 3, 4, 5. The value {} can be pruned from each of the domains of vB , vC , and
vD .

One can see that the vertex variables together with the acyclic constraint
are sufficient alone to model the Bayesian network structure learning problem.
Such a search space over DAGs forms the basis of Barlett and Cussens’ integer
programming approach [8]. One can also see that the ordering (permutation)
variables together with the alldifferent constraint are sufficient alone, as the min-
imum cost DAG for a given ordering is easily determinable. Larranaga et al. [22]
were the first to propose the search space of all permutations and Teyssier and
Koller [23] successfully applied it within a local search algorithm. The permu-
tation search space also forms the basis of the approaches based on dynamic
programming [4–6] and on the approaches based on searching for the shortest
path in an ordering graph using A* search [9–11], depth-first branch-and-bound
DFBnB search [13], and BFBnB search [10,11,15].
434 P. van Beek and H.-F. Hoffmann

The unique aspects of our model lie in combining the DAG and permutation
search spaces and introducing the depth variables. As is shown in the next sec-
tions, the combination of variables allows us to identify and post many additional
constraints that lead to a considerable reduction in the search space.

3.1 Symmetry-Breaking Constraints (I)

Many permutations and prefixes of permutations, as represented by the ordering
variables, are symmetric in that they lead to the same minimum cost DAG,
or are dominated in that they lead to a DAG of equal or higher cost. The
intent behind introducing the auxiliary depth variables is to rule out all but the
lexicographically least of these permutations. A lexicographic ordering is defined
over the depth variables—and over the ordering variables, in the case of a tie
on the values of the depth variables. The following constraints are introduced to
enforce the lexicographic order.

d1 = 0 (5)
di = k ⇐⇒ (di+1 = k ∨ di+1 = k + 1), i = 1, . . . , n − 1 (6)
di = di+1 =⇒ oi < oi+1 , i = 1, . . . , n − 1 (7)

The constraints are readily propagated. Constraint 6 is a dominance con-

straint.
Example 5. In Example 2, consider the ordering prefix (o1 , . . . , o4 ) =
(E, C, A, D) with associated vertex variables (vE , vC , vA , vD ) = ({}, {E}, {C},
{E}) and depths (d1 , . . . , d4 ) = (0, 1, 2, 1). The cost of this ordering prefix is
33.8. The ordering prefix violates Constraint 6. However, the ordering prefix
(o1 , . . . , o4 ) = (E, C, D, A) with depths (d1 , . . . , d4 ) = (0, 1, 1, 2) and vertex vari-
ables (vE , . . . , vD ) = ({}, {E}, {E}, {D}) satisfies the constraint and has a lower
cost of 33.5.
Constraint 7 is a symmetry-breaking constraint.
Example 6. In Example 2, consider the ordering (o1 , . . . , o5 ) = (A, E, D, C, B)
with (d1 , . . . , d5 ) = (0, 1, 2, 2, 3) and (vA , . . . , vB ) = ({}, . . . , {C, D}). The order-
ing violates Constraint 7. However, the symmetric ordering (o1 , . . . , o5 ) =
(A, E, C, D, B), which represents the same DAG, satisfies the constraint and
has equal cost.

Theorem 1. Any ordering prefix o1 , . . . , oi can be safely pruned if the associated

depth variables d1 , . . . , di do not satisfy Constraints 5–7.

3.2 Symmetry-Breaking Constraints (II)

In the previous section, we identified symmetries and dominance among the

ordering variables. In this section, we identify symmetries among the vertex
 Machine Learning of Bayesian Networks Using Constraint Programming 435

variables. Let [x/y]dom(v) be the domain that results from replacing all occur-
rences of y in the parent sets by x. Two vertex variables v1 and v2 are symmetric
if [v1 /v2 ]dom(v1 ) = dom(v2 ); i.e., the domains are equal once the given substi-
tution is applied. The symmetry is broken by enforcing that v1 must precede v2
in the ordering,
∀i • ∀j • oi = 1 ∧ oj = 2 =⇒ i < j. (8)

Example 7. In Example 2, consider vertex variables vC and vD . The variables

are symmetric as, [vC /vD ]dom(vC ) = {{E}, {C}, {A, B}, {A}, {}} = dom(vD ).

3.3 Symmetry-Breaking Constraints (III)

A BN can be viewed as an encoding of conditional independence assumptions.

Two BN structures (DAGS) are said to be I-equivalent if they encode the same
set of conditional independence assumptions (see, e.g., [16,17]). The efficiency
of the search for a minimal cost BN can be greatly improved by recognizing
I-equivalent partial (non-)solutions. Chickering [24,25] provides a local transfor-
mational characterization of equivalent BN structures based on covered edges
that forms the theoretical basis of these symmetry-breaking constraints.

Definition 3 (Chickering [24]). An edge x → y in a Bayesian network is a

covered edge if parents(y) = parents(x) ∪ {x}.

Theorem 2 (Chickering [24]). Let G be any DAG containing the edge x → y,

and let G  be the directed graph identical to G except that the edge between x
and y in G  is oriented as y → x. Then G  is a DAG that is equivalent to G if
and only if x → y is a covered edge in G.

Example 8. In Figure 1(b) the edge A → E is a covered edge and the Bayesian
network with the edge reversed to be E → A is an I-equivalent Bayesian network.

In what follows, we identify three cases that consist of sequences of one or

more covered edge reversals and break symmetries by identifying a lexicographic
ordering. Experimental evidence suggests that these three cases capture much of
the symmetry due to I-equivalence. Symmetries are only broken if the costs of the
two I-equivalent DAGs would be equal; otherwise there is a negative interaction
with pruning based on the cost function (discussed in Section 3.8).
Case 1. Consider vertex variables vi and vj . If there exists domain values
p ∈ dom(vi ) and p ∪ {vi } ∈ dom(vj ), this pair of assignments includes a covered
edge vi → vj ; i.e., vi and vj would have identical parents except that vi would
also be a parent of vj . Thus, there exists an I-equivalent DAG with the edge
reversed. We keep only the lexicographically least: the pair of assignments would
be permitted iff i < j.
Case 2. Consider vertex variables vi , vj , and vk . If there exists domain values
p ∈ dom(vi ), p ∪ {vi } ∈ dom(vj ), p ∪ {vj } ∈ dom(vk ), where i < j and k < j,
there is a covered edge vi → vj and, if this covered edge is reversed, the covered
edge vj → vk is introduced, which in turn can be reversed. Thus, there exists an
436 P. van Beek and H.-F. Hoffmann

I-equivalent DAG with the edges {vi → vj , vj → vk } and an I-equivalent DAG

with the edges {vk → vj , vj → vi }. We keep only the lexicographically least: the
triple of assignments would be permitted iff i < k.
Case 3. Consider vertex variables vi , vj , vk , and vl . If there exists domain
values p ∈ dom(vi ), p ∪ {vi } ∈ dom(vj ), p ∪ {vi , vj } ∈ dom(vk ), p ∪ {vj , vk } ∈
dom(vl ), where i < j, l < j, j < k, there exists an I-equivalent DAG with the
edges {vi → vj , vi → vk , vj → vk , vj → vl , vk → vl } and an I-equivalent DAG
with the edges {vl → vj , vl → vk , vj → vk , vj → vi , vk → vi }. We keep only the
lexicographically least: the triple of assignments would be permitted iff i < l.
In our empirical evaluation, these symmetry-breaking rules were used only as
a satisfiability check at each node in the search tree, as we found that propagating
the I-equivalence symmetry-breaking rules did not further improve the runtime.

3.4 Dominance Constraints (I)

Given an ordering prefix o1 , . . . , oi−1 , a domain value p for a vertex variable vj
is consistent with the ordering if each element of p occurs in a parent set of a
vertex variable in the ordering. The domain value p assigned to vertex variable
vj that occurs at position i in an ordering should be the lowest cost p consistent
with the ordering; assigning a domain value with a higher cost can be seen to be
dominated as the values can be substituted with no effect on the other variables.
Theorem 3. Given an ordering prefix o1 , . . . , oi−1 , a vertex variable vj , and
domain elements p, p ∈ dom(vj ), p = p , if p is consistent with the ordering and
cost(p) ≤ cost(p ), p can be safely pruned from the domain of vj .
Example 9. In Example 2, consider the prefix ordering (o1 , o2 ) = (C, D). The
values {C}, {E}, and {} can be pruned from each of the domains of vA and vB ,
and the values {A}, {A, B}, {C}, and {} can be pruned from the domain of vE .

3.5 Dominance Constraints (II)

Teyssier and Koller [23] present a pruning rule that is now routinely used in
score-and-search approaches as a preprocessing step before search begins.
Theorem 4 (Teyssier and Koller [23]). Given a vertex variable vj , and
domain elements p, p ∈ dom(vj ), if p ⊂ p and cost(p) ≤ cost(p ), p can be
safely pruned from the domain of vj .
Example 10. In Example 2, the value {A, B} can be pruned from the domain of
vE .
We generalize the pruning rule so that it is now applicable during the search.
Suppose that some of the vertex variables have been assigned values. These
assignments induce ordering constraints on the variables.
Example 11. In Example 2, suppose vA = {D} and vC = {A, B}. These assign-
ments induce the ordering constraints D < A, A < C, and B < C.
 Machine Learning of Bayesian Networks Using Constraint Programming 437

Definition 4. Given a set of ordering constraints induced by assignments to

the vertex variables, let ip(p) denote the induced parent set where p has been
augmented with any and all variables that come before in the ordering; i.e., if
y ∈ p and x < y then x is added to p.

The generalized pruning rule is as follows.

Theorem 5. Given a vertex variable vj , and domain elements p, p ∈

dom(vj ), p = p , if p ⊆ ip(p ) and cost(p) ≤ cost(p ), p can be safely pruned
from the domain of vj .

Example 12. Continuing with Example 11, consider vE with p = {D} and p =
{A}. The induced parent set ip(p ) is given by {A, D} and cost(p) ≤ cost(p ).
Thus, p can be pruned. Similarly, p = {C} can be pruned.

3.6 Dominance Constraints (III)

Consider an ordering prefix o1 , . . . , oi with associated vertex and depth variables

and let π be a permutation over {1, . . . , i}. The cost of completing the partial
solutions represented by the prefix ordering o1 , . . . , oi and the permuted prefix
ordering oπ(1) , . . . , oπ(i) are identical. This insight is used by Silander and Myl-
lymäki [5] in their dynamic programming approach and by Fan et al. [9–11] in
their best-first approaches based on searching for the shortest path in the order-
ing graph. However, all of these approaches are extremely memory intensive.
Here, we use this insight to prune the search space.

Theorem 6. Let cost(o1 , . . . , oi ) be the cost of a partial solution represented by

the given ordering prefix. Any ordering prefix o1 , . . . , oi can be safely pruned if
there exists a permutation π such that cost(oπ(1) , . . . , oπ(i) ) < cost(o1 , . . . , oi ).

Example 13. In Example 2, consider the ordering prefix O = (o1 , o2 ) = (E, A)

with associated vertex variables (vE , vA ) = ({}, {E}) and cost of 27.0. The order-
ing prefix O can be safely pruned as there exists a permutation (oπ(1) , oπ(2) ) =
(A, E) with associated vertex variables (vA , vE ) = ({}, {A}) that has a lower
cost of 19.6.

Clearly, in its full generality, the condition of Theorem 6 is too expensive to

determine exactly as it amounts to solving the original problem. However, we
identify three strategies that are easy to determine and collectively were found
to be very effective at pruning in our experiments while maintaining optimality.
Strategy 1. We consider permutations that differ from the original permu-
tation only in the last l or fewer places (l = 4 in our experiments).
Strategy 2. We consider permutations that differ from the original permu-
tation only in the swapping of the last variable oi with a variable earlier in the
ordering.
Strategy 3. We consider whether a permutation oπ(1) , . . . , oπ(i) of lower cost
was explored earlier in the search. To be able to efficiently determine this, we
438 P. van Beek and H.-F. Hoffmann

use memoization for ordering prefixes and only continue with a prefix if it has a
better cost than one already explored (see, e.g., [26,27]). Our implementation of
memoization uses hashing with quadratic probing and the replacement policy is
to keep the most recent if the table becomes too full. It is well known that there
is a strong relationship between backtracking search with memoization/caching
and dynamic programming using a bottom-up approach, but memoization allows
trading space for time and top-down backtracking search allows pruning the
search space.

3.7 Acyclic Constraint

In this section we describe a propagator for the acyclicity constraint that achieves
generalized arc consistency in polynomial time. We first present and analyze an
algorithm that checks satisfiability for given possible parent sets. We then explain
how this algorithm can be used to achieve generalized arc consistency.
Algorithm 1 can check whether a collection of possible parent sets allows a
feasible parent set assignment, i.e. an assignment that represents a DAG. Its
correctness is based on the following well-known property of directed acyclic
graphs that is also used in the ILP approaches [7,8].

Theorem 7. Let G be a directed graph over vertices V and let parents(v) be the
parents of vertex v in the graph. G is acyclic if and only if for every non-empty
subset S ⊂ V there is at least one vertex v ∈ S with parents(v) ∩ S = {}.

The algorithm works as follows. First, it searches possible sources for the
DAG, i.e. vertices for which {} is a possible parent set. These vertices are stored
in W 0 . Note that if a directed graph does not have a source, it must contain a
cycle by Theorem 7. Thus, if W 0 remains empty, there is no parent set assignment
satisfying the acyclicity constraint. In the next iteration, the algorithm searches
for vertices that have a possible parent set consisting of possible sources only.
These vertices form set W 1 . Again, if there are no such vertices, then no vertex
in V \ W 0 has a possible parent set completely outside V \ W 0 , which violates
the acyclicity characterization of Theorem 7. Thus, there is no consistent parent
set assignment. We continue this process until all vertices
k are included in one of
the W k sets or until we find a contradicting set V \ ( i=0 W i ) for some k.

Theorem 8. We can test satisfiability of the acyclic constraint in time O(n2 d),
where n is the number of vertices and d is an upper bound on the number of
possible parent sets per vertex.

Example 14. Let vA , vB , vC , and vD be vertex variables with the possible parent
sets,
dom(vA ) = {{B}, {D}}, dom(vC ) = {{B}, {D}},
dom(vB ) = {{A}, {C}}, dom(vD ) = {{A}, {C}}.
Algorithm 1 returns false as W 0 is found to be empty.
 Machine Learning of Bayesian Networks Using Constraint Programming 439

Algorithm 1. Checking satisfiability of acyclic constraint

Input: V = {v1 , . . . , vn }, set dom(vi ) of possible parent sets for each vertex vi
in V .
Output: True if there is a feasible parent set assignment and false otherwise.
Additionally, the variables Si represent a feasible assignment if one
exists.
k ← 0;
Si ← nil for all vi ∈ V ;
k−1 j
while j=0 W = V do
W k ← {};
k−1 j
for all vertices vi not in j=0 W do
k−1 j
if vi has a possible parent set p ∈ dom(vi ) with p ⊆ j=0 W then
Si ← p;
W k ← W k ∪ {vi };
end if
end
if W k = {} then return false;
;
k ← k + 1;
end while
return true;

The algorithm for checking satisfiability can be used to achieve general-

ized arc consistency by iteratively testing, for each vertex vi , whether each
p ∈ dom(vi ) has a support. We simply substitute the set of possible parent
sets dom(vi ) for vi by the set {p}. A satisfiability check on the resulting instance
successfully tests whether there is a consistent parent set assignment containing
vi = p. If we find a parent set p that cannot appear in any feasible solution, we
remove p from the corresponding domain. Note that we only prune a value p
from a domain dom(vi ) if vi = p cannot be part of any feasible solution. This
means that vi = p can also not be part of the support of any other variable value
q ∈ dom(vj ). Therefore, the removal of p cannot cause another supported value
to become unsupported. Hence, we do not have to rerun any of the tests; we can
simply run the test once for every value in every domain. These considerations
show that we can enforce arc consistency for the acyclicity constraint in O(n3 d2 )
steps.
In our empirical evaluation, we found that achieving generalized arc consis-
tency did not pay off in terms of reduced runtime. Hence, in the current set of
experiments Algorithm 1 was used only as a satisfiability check at each node
in the search tree. Instead, a limited form of constraint propagation was per-
formed based on necessary edges between vertex variables. An edge vi → vj
is necessary if vi occurs in every currently valid parent set for variable vj ; i.e.,
∀p ∈ dom(vj ) • vi ∈ p. If a directed edge vi → vj is a necessary edge, the
directed edge vj → vi cannot be an edge in a valid DAG, as a cycle would
440 P. van Beek and H.-F. Hoffmann

be introduced. Thus, any parent set that contains vj can be removed from the
domain of vi . Removing domain elements may introduce additional necessary
edges and pruning can be based on chaining necessary edges.

Example 15. Let vA , vB , vC , and vD be vertex variables with the possible parent
sets,
dom(vA ) = {{}, {B}, {C}} dom(vC ) = {{B}}
dom(vB ) = {{A}, {A, C}} dom(vD ) = {{A}, {A, C}}
Since the edge B → C is necessary, the value {A, C} can be pruned from the
domain of vB . This causes the edge A → B to become necessary, and the values
{B} and {C} can be pruned from the domain of vA .

We conclude with the following observation. Let iv be the index of the set W i
in which we include vertex v in the satisfiability algorithm. Then, iv is a lower
bound on the number that vertex v can have in any topological numbering. This
lower bound can be used in propagating Constraint 4.

3.8 Solving the Constraint Model

A constraint-based depth-first branch-and-bound search is used to solve the con-

straint model; i.e., the nodes in the search tree are expanded in a depth-first
manner and a node is expanded only if the propagation of the constraints suc-
ceeds and a lower bound estimate on completing the partial solution does not
exceed the current upper bound.
The branching is over the ordering (permutation) variables and uses the
static order o1 , . . . , on . Once an ordering variable is instantiated as oi = j, the
associated vertex variable vj and depth variable di are also instantiated.
The lower bound is based on the lower bound proposed by Fan and Yuan
[11]. In brief, prior to search, the strongly connected components (SCCs) of the
graph based on the top few lowest cost elements in the domains of the vertex
variables are found and pattern databases are constructed based on the SCCs.
The pattern databases allow a fast and often accurate lower bound estimate
during the search (see [11] for details).
The initial upper bound, found before search begins, is based on the local
search algorithm proposed by Teyssier and Koller [23]. The algorithm uses
restarts and first-improvement moves, the search space consists of all permu-
tations, and the neighborhood function consists of swapping the order of two
variables. Of course, as better solutions are found during the search the upper
bound is updated.
As a final detail, additional pruning on the vertex variables can be performed
based on the (well-known) approximation of bounds consistency on a cost func-
tion that is a knapsack constraint: z = cost(v1 ) + · · · + cost(vn ). Let the bounds
on cost(vi ) be [li , ui ], and let lb and ub be the current lower bound and upper
bound on the cost, respectively. At any point in the search we have the  constraint
lb ≤ z < ub and a value p ∈ dom(vi ) can be pruned if cost(p) + j=i uj < lb
 Machine Learning of Bayesian Networks Using Constraint Programming 441

 
or if cost(p) + j=i lj ≥ ub. Note that the expression cost(p) + j=i lj can
be replaced with any lower bound on the cost of a solution that includes p and
respects the current domains and instantiations, as long as the lower bound never
over estimates. Fortunately, we have a fast and effective method of querying such
lower bounds and we use it when performing this pruning.

4 Experimental Evaluation

In this section, we compare a bespoke C++ implementation of our constraint-

based approach, called CPBayes 2 , to the current state-of-the-art on benchmark
instances and show that our approach compares favorably both in terms of num-
ber of instances solved within specified resource bounds and in terms of solution
time.
The set of benchmark instances are derived from data sets obtained from the
UCI Machine Learning Repository 3 and data generated from networks obtained
from the Bayesian Network Repository 4 . Following previous work, the local
score for each possible parent set and each random variable was computed in a
preprocessing step (either by us or by others) prior to the search for the best
network structure and we do not report the preprocessing time. Note that the
computations of the possible parent sets for each variable are independent and
can be determined in parallel. The BIC/MDL [18,19] and BDeu [20,21] scoring
methods were used.
Table 2 shows the results of comparing CPBayes (v1.0) against Barlett and
Cussens’ [8] GOBNILP system (v1.4.1) based on integer linear programming, and
Fan, Malone, and Yuan’s [10,11,15] system (v2015) based on A* search. These
two systems represent the current state-of-the-art for global (exact) approaches.
Breadth-first BnB search [10,11,15] is also competitive but its effectiveness is
known to be very similar to that of A*. Although for space reasons we do not
report detailed results, we note that on these benchmarks CPBayes far outpaces
the previous best depth-first branch-and-bound search approach [13]. GOBNILP
(v1.4.1) 5 and A* (v2015) 6 are both primarily written in C/C++. A* (v2015)
is the code developed by Fan et al. [10,15], but in the experiments we included
our implementation of the improved lower bounds recently proposed by Fan and
Yuan [11]. Thus, both CPBayes (v1.0) and A* (v2015) use exactly the same
lower bounding technique (see Section 3.8). The experiments were performed on
a cluster, where each node of the cluster is equipped with four AMD Opteron
CPUs at 2.4 GHz and 32.0 GB memory. Resource limits of 24 hours of CPU time
and 16 GB of memory were imposed both for the preprocessing step common
to all methods of obtaining the local scores and again to determine the minimal
cost BN using a method. The systems were run with their default values.
2
CPBayes code available at: cs.uwaterloo.ca/∼vanbeek/research
3
archive.ics.uci.edu/ml/
4
www.bnlearn.com/bnrepository/
5
www.cs.york.ac.uk/aig/sw/gobnilp/
6
bitbucket.org/bmmalone/
442 P. van Beek and H.-F. Hoffmann

Table 2. For each benchmark, time (seconds) to determine minimal cost BN using
various systems (see text), where n is the number of random variables in the data set,
N is the number of instances in the data set, and d is the total number of possible
parents sets for the random variables. Resource limits of 24 hours of CPU time and 16
GB of memory were imposed: OM = out of memory; OT = out of time. A blank entry
indicates that the preprocessing step of obtaining the local scores for each random
variable could not be completed within the resource limits.

BDeu BIC
GOBN. A* CPBayes GOBN. A* CPBayes
Benchmark n N d v1.4.1 v2015 v1.0 d v1.4.1 v2015 v1.0
shuttle 10 58,000 812 58.5 0.0 0.0 264 2.8 0.1 0.0
adult 15 32,561 768 1.4 0.1 0.0 547 0.7 0.1 0.0
letter 17 20,000 18,841 5,060.8 1.3 1.4 4,443 72.5 0.6 0.2
voting 17 435 1,940 16.8 0.3 0.1 1,848 11.6 0.4 0.1
zoo 17 101 2,855 177.7 0.5 0.2 554 0.9 0.4 0.1
tumour 18 339 274 1.5 0.9 0.2 219 0.4 0.9 0.2
lympho 19 148 345 1.7 2.1 0.5 143 0.5 1.0 0.2
vehicle 19 846 3,121 90.4 2.4 0.7 763 4.4 2.1 0.5
hepatitis 20 155 501 2.1 4.9 1.1 266 1.7 4.8 1.0
segment 20 2,310 6,491 2,486.5 3.3 1.3 1,053 13.2 2.4 0.5
mushroom 23 8,124 438,185 OT 255.5 561.8 13,025 82,736.2 34.4 7.7
autos 26 159 25,238 OT 918.3 464.2 2,391 108.0 316.3 50.8
insurance 27 1,000 792 2.8 583.9 107.0 506 2.1 824.3 103.7
horse colic 28 300 490 2.7 15.0 3.4 490 3.2 6.8 1.2
steel 28 1,941 113,118 OT 902.9 21,547.0 93,026 OT 550.8 4,447.6
flag 29 194 1,324 28.0 49.4 39.9 741 7.7 12.1 2.6
wdbc 31 569 13,473 2,055.6 OM 11,031.6 14,613 1,773.7 1,330.8 1,460.5
water 32 1,000 159 0.3 1.6 0.6
mildew 35 1,000 166 0.3 7.6 1.5 126 0.2 3.6 0.6
soybean 36 266 5,926 789.5 1,114.1 147.8
alarm 37 1,000 672 1.8 43.2 8.4
bands 39 277 892 15.2 4.5 2.0
spectf 45 267 610 8.4 401.7 11.2
sponge 45 76 618 4.1 793.5 13.2
barley 48 1,000 244 0.4 1.5 3.4
hailfinder 56 100 167 0.1 9.9 1.5
hailfinder 56 500 418 0.5 OM 9.3
lung cancer 57 32 292 2.0 OM 10.5
carpo 60 100 423 1.6 OM 253.6
carpo 60 500 847 6.9 OM OT

5 Discussion and Future Work

The Bayesian Network Repository classifies networks as small (< 20 random
variables), medium (20–60 random variables), large (60–100 random variables),
very large (100–1000 random variables), and massive (> 1000 random variables).
The benchmarks shown in Table 2 fall into the small and medium classes. We
are not aware of any reports of results for exact solvers for instances beyond the
medium class (Barlett and Cussens [8] report results for GOBNILP for n > 60,
 Machine Learning of Bayesian Networks Using Constraint Programming 443

but they are solving a different problem, severely restricting the cardinality of
the parent sets to ≤ 2).
Benchmarks from the small class are easy for the CPBayes and A* methods,
but can be somewhat challenging for GOBNILP depending on the value of the
parameter d, the total number of parent sets for the random variables. Along
with the integer linear programming (ILP) solver GOBNILP, CPBayes scales
fairly robustly to medium instances using a reasonable restriction on memory
usage (both use only a few GB of memory, far under the 16 GB limit used
in the experiments; in fairness, the scalability of the A* approach on a very
large memory machine is still somewhat of an open question). CPBayes also has
several other advantages, which it shares with the ILP approach, over A*, DP,
and BFBnB approaches. Firstly, the constraint model is a purely declarative
representation and the same model can be given to an exact solver or a solver
based on local search, such as large neighborhood search. Secondly, the constraint
model can be augmented with side structural constraints that can be important
in real-world modeling (see [28]). Finally, the solver is an anytime algorithm
since, as time progresses, the solver progressively finds better solutions.
Let us now turn to a comparison between GOBNILP and CPBayes. CPBayes
scales better than GOBNILP along the dimension d which measures the size of
the possible parent sets. A partial reason is that GOBNILP uses a constraint
model that includes a (0,1)-variable for each possible parent set. GOBNILP
scales better than CPBayes along the dimension n which measures the number
of random variables. CPBayes has difficulty at the topmost range of n proving
optimality. There is some evidence that n = 60 is near the top of the range for
GOBNILP as well. Results reported by Barlett and Cussens [8] for the carpo
benchmark using larger values of N and the BDeu scoring method—the scoring
method which usually leads to harder optimization instances than BIC/MDL—
showed that instances could only be solved by severely restricting the cardinality
of the parent sets. A clear difficulty in scaling up all of these score-and-search
methods is in obtaining the local scores within reasonable resource limits.
In future work on Bayesian network structure learning, we intend to focus on
improving the robustness and scalability of our CPBayes approach. A direction
that appears especially promising is to improve the branch-and-bound search by
exploiting decomposition and lower bound caching during the search [29,30]. As
well, our approach, as with all current exact approaches, assumes complete data.
An important next step is to extend our approach to handle missing values and
latent variables (cf. [31]).

Acknowledgments. This research was partially funded through an NSERC Discov-

ery Grant. We thank Claude-Guy Quimper, Alejandro López-Ortiz, Mats Carlsson,
and Christian Schulte for helpful discussions, and Brandon Malone and James Cussens
for providing test instances and their code.
444 P. van Beek and H.-F. Hoffmann

References
1. Witten, I.H., Frank, E., Hall, M.A.: Data Mining, 3rd edn. Morgan Kaufmann
(2011)
2. Chickering, D., Meek, C., Heckerman, D.: Large-sample learning of Bayesian net-
works is NP-hard. In: Proc. of UAI, pp. 124–133 (2003)
3. Koivisto, M.: Parent assignment is hard for the MDL, AIC, and NML costs. In:
Lugosi, G., Simon, H.U. (eds.) COLT 2006. LNCS (LNAI), vol. 4005, pp. 289–303.
Springer, Heidelberg (2006)
4. Koivisto, M., Sood, K.: Exact Bayesian structure discovery in Bayesian networks.
J. Mach. Learn. Res. 5, 549–573 (2004)
5. Silander, T., Myllymäki, P.: A simple approach for finding the globally optimal
Bayesian network structure. In: Proc. of UAI, pp. 445–452 (2006)
6. Malone, B., Yuan, C., Hansen, E.A.: Memory-efficient dynamic programming for
learning optimal Bayesian networks. In: Proc. of AAAI, pp. 1057–1062 (2011)
7. Jaakkola, T., Sontag, D., Globerson, A., Meila, M.: Learning Bayesian network
structure using LP relaxations. In: Proc. of AISTATS, pp. 358–365 (2010)
8. Barlett, M., Cussens, J.: Advances in Bayesian network learning using integer pro-
gramming. In: Proc. of UAI, pp. 182–191 (2013)
9. Yuan, C., Malone, B.: Learning optimal Bayesian networks: A shortest path per-
spective. J. of Artificial Intelligence Research 48, 23–65 (2013)
10. Fan, X., Malone, B., Yuan, C.: Finding optimal Bayesian network structures with
constraints learned from data. In: Proc. of UAI, pp. 200–209 (2014)
11. Fan, X., Yuan, C.: An improved lower bound for Bayesian network structure learn-
ing. In: Proc. of AAAI (2015)
12. Tian, J.: A branch-and-bound algorithm for MDL learning Bayesian networks. In:
Proc. of UAI, pp. 580–588 (2000)
13. Malone, B., Yuan, C.: A depth-first branch and bound algorithm for learning opti-
mal bayesian networks. In: Croitoru, M., Rudolph, S., Woltran, S., Gonzales, C.
(eds.) GKR 2013. LNCS, vol. 8323, pp. 111–122. Springer, Heidelberg (2014)
14. de Campos, C.P., Ji, Q.: Efficient structure learning of Bayesian networks using
constraints. Journal of Machine Learning Research 12, 663–689 (2011)
15. Fan, X., Yuan, C., Malone, B.: Tightening bounds for Bayesian network structure
learning. In: Proc. of AAAI, pp. 2439–2445 (2014)
16. Darwiche, A.: Modeling and Reasoning with Bayesian Networks. Cambridge Uni-
versity Press (2009)
17. Koller, D., Friedman, N.: Probabilistic Graphical Models: Principles and Tech-
niques. The MIT Press (2009)
18. Schwarz, G.: Estimating the dimension of a model. Ann. Stat. 6, 461–464 (1978)
19. Lam, W., Bacchus, F.: Using new data to refine a Bayesian network. In: Proc. of
UAI, pp. 383–390 (1994)
20. Buntine, W.L.: Theory refinement of Bayesian networks. In: Proc. of UAI,
pp. 52–60 (1991)
21. Heckerman, D., Geiger, D., Chickering, D.M.: Learning Bayesian networks: The
combination of knowledge and statistical data. Machine Learning 20, 197–243
(1995)
22. Larranaga, P., Kuijpers, C., Murga, R., Yurramendi, Y.: Learning Bayesian net-
work structures by searching for the best ordering with genetic algorithms. IEEE
Trans. Syst., Man, Cybern. 26, 487–493 (1996)
 Machine Learning of Bayesian Networks Using Constraint Programming 445

23. Teyssier, M., Koller, D.: Ordering-based search: a simple and effective algorithm
for learning Bayesian networks. In: Proc. of UAI, pp. 548–549 (2005)
24. Chickering, D.M.: A transformational characterization of equivalent Bayesian net-
work structures. In: Proc. of UAI, pp. 87–98 (1995)
25. Chickering, D.M.: Learning equivalence classes of Bayesian network structures.
Journal of Machine Learning Research 2, 445–498 (2002)
26. Michie, D.: “memo” functions and machine learning. Nature 218, 19–22 (1968)
27. Smith, B.M.: Caching search states in permutation problems. In: van Beek, P. (ed.)
CP 2005. LNCS, vol. 3709, pp. 637–651. Springer, Heidelberg (2005)
28. Cussens, J.: Integer programming for Bayesian network structure learning. Quality
Technology & Quantitative Management 1, 99–110 (2014)
29. Kitching, M., Bacchus, F.: Symmetric component caching. In: Proc. of IJCAI,
pp. 118–124 (2007)
30. Kitching, M., Bacchus, F.: Exploiting decomposition in constraint optimization
problems. In: Stuckey, P.J. (ed.) CP 2008. LNCS, vol. 5202, pp. 478–492. Springer,
Heidelberg (2008)
31. Friedman, N.: Learning belief networks in the presence of missing values and hidden
variables. In: Proc. of ICML, pp. 125–133 (1997)
 Hybridization of Interval CP and Evolutionary
Algorithms for Optimizing Difficult Problems

Charlie Vanaret1(B) , Jean-Baptiste Gotteland2 ,

Nicolas Durand2 , and Jean-Marc Alliot1
1
Institut de Recherche en Informatique de Toulouse, 2 Rue Charles Camichel,
31000 Toulouse, France
charlie.vanaret@enseeiht.fr, jean-marc.alliot@irit.fr
2
Ecole Nationale de l’Aviation Civile, 7 Avenue Edouard Belin,
31055 Toulouse Cedex 04, France
{gottelan,durand}@recherche.enac.fr

Abstract. The only rigorous approaches for achieving a numerical proof

of optimality in global optimization are interval-based methods that
interleave branching of the search-space and pruning of the subdomains
that cannot contain an optimal solution. State-of-the-art solvers gener-
ally integrate local optimization algorithms to compute a good upper
bound of the global minimum over each subspace. In this document,
we propose a cooperative framework in which interval methods cooper-
ate with evolutionary algorithms. The latter are stochastic algorithms
in which a population of candidate solutions iteratively evolves in the
search-space to reach satisfactory solutions.
Within our cooperative solver Charibde, the evolutionary algorithm
and the interval-based algorithm run in parallel and exchange bounds,
solutions and search-space in an advanced manner via message passing.
A comparison of Charibde with state-of-the-art interval-based solvers
(GlobSol, IBBA, Ibex) and NLP solvers (Couenne, BARON) on a bench-
mark of difficult COCONUT problems shows that Charibde is highly
competitive against non-rigorous solvers and converges faster than rig-
orous solvers by an order of magnitude.

1 Motivation
We consider n-dimensional continuous constrained optimization problems over
a hyperrectangular domain D = D1 × . . . × Dn :

(P) min f (x)

x∈D⊂Rn
s.t. gi (x) ≤ 0, i ∈ {1, . . . , m} (1)
hj (x) = 0, j ∈ {1, . . . , p}

When f , gi and hj are non-convex, the problem may have multiple local
minima. Such difficult problems are generally solved using generic exhaustive
branch and bound (BB) methods. The objective function and the constraints are

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 446–462, 2015.
DOI: 10.1007/978-3-319-23219-5 32
 Hybridization of Interval CP and Evolutionary Algorithms 447

bounded on disjoint subspaces using enclosure methods. By By keeping track of

the best known upper bound f˜ of the global minimum f ∗ , subspaces that cannot
contain a global minimizer are discarded (pruned).
Several authors proposed hybrid approaches in which a BB algorithm cooper-
ates with another technique to enhance the pruning of the search-space. Hybrid
algorithms may be classified into two categories [24]: integrative approaches,
in which one of the two methods replaces a particular operator of the other
method, and cooperative methods, in which the methods are independent and
are run sequentially or in parallel. Previous works include

– integrative approaches: [34] integrates a stochastic genetic algorithm (GA)

within an interval BB. The GA provides the direction along which a box
is partitioned, and an individual is generated within each subbox. At each
generation, the best evaluation updates the best known upper bound of the
global minimum. In [9], the crossover operator is replaced by a BB that
determines the best individual among the offspring.
– cooperative approaches: [28] sequentially combines an interval BB and a
GA. The interval BB generates a list L of remaining small boxes. The GA’s
population is initialized by generating a single individual within each box
of L. [12] (BB and memetic algorithm) and [7] (beam search and memetic
algorithm) describe similar parallel strategies: the BB identifies promising
regions that are then explored by the metaheuristic. [1] hybridizes a GA and
an interval BB. The two independent algorithms exchange upper bounds and
solutions through shared memory. New optimal results are presented for the
rotated Michalewicz (n = 12) and Griewank functions (n = 6).

In this communication, we build upon the cooperative scheme of [1]. The

efficiency and reliability of their solver remain very limited; it is not competi-
tive against state-of-the-art solvers. Their interval techniques are naive and may
lose solutions, while the GA may send evaluations subject to roundoff errors.
We propose to hybridize a stochastic differential evolution algorithm (close to a
GA), described in Section 2, and a deterministic interval branch and contract
algorithm, described in Section 3. Our hybrid solver Charibde is presented in
Section 4. Experimental results (Section 5) show that Charibde is highly com-
petitive against state-of-the-art solvers.

2 Differential Evolution
Differential evolution (DE) [29] is among the simplest and most efficient meta-
heuristics for continuous problems. It combines the coordinates of existing
individuals (candidate solutions) with a given probability to generate new
individuals. Initially devised for continuous unconstrained problems, DE was
extended to mixed problems and constrained problems [23].
Let NP denote the size of the population, W > 0 the amplitude factor and
CR ∈ [0, 1] the crossover rate. At each generation (iteration), NP new individ-
uals are generated: for each individual x = (x1 , . . . , xn ), three other individuals
448 C. Vanaret et al.

u = (u1 , . . . , un ) (called base individual ), v = (v1 , . . . , vn ) and w = (w1 , . . . , wn ),

all different and different from x, are randomly picked in the population. The
coordinates yi of the new individual y = (y1 , . . . , yn ) are computed according to

ui + W × (vi − wi ) if i = R or ri < CR
yi = (2)
xi otherwise
where ri is picked in [0, 1] with uniform probability. The index R, picked in
{1, . . . , n} with uniform probability for each x, ensures that at least a coordinate
of y differs from that of x. y replaces x in the population if it is “better” than x
(e.g. in unconstrained optimization, y is better than x if it improves the objective
function).
Figure 1 depicts a two-dimensional crossover between individuals x, u (base
individual), v and w. The contour lines of the objective function are shown in
grey. The difference v − w, scaled by W , yields the direction (an approximation
of the direction opposite the gradient) along which u is translated to yield y.

Fig. 1. Crossover of the differential evolution

3 Reliable Computations

Reliable (or rigorous) methods provide bounds on the global minimum, even
in the presence of roundoff errors. The only reliable approaches for achieving a
numerical proof of optimality in global optimization are interval-based methods
that interleave branching of the search-space and pruning of the subdomains
that cannot contain an optimal solution.
Section 3.1 introduces interval arithmetic, an extension of real arithmetic.
Reliable global optimization is detailed in Section 3.2, and interval contractors
are mentioned in Section 3.3.
 Hybridization of Interval CP and Evolutionary Algorithms 449

3.1 Interval Arithmetic

An interval X with floating-point bounds defines the set {x ∈ R | X ≤ x ≤ X}.

IR denotes the set of all intervals. The width of X is w(X) = X − X. m(X) =
X+X
2 is the midpoint of X. A box X is a Cartesian product of intervals. The
width of a box is the maximum width of its components. The convex hull (X, Y )
of X and Y is the smallest interval enclosing X and Y .
Interval arithmetic [19] extends real arithmetic to intervals. Interval arith-
metic implemented on a machine must be rounded outward (the left bound is
rounded toward −∞, the right bound toward +∞) to guarantee conservative
properties. The interval counterparts of binary operations and elementary func-
tions produce the smallest interval containing the image. Thanks to the conser-
vative properties of interval arithmetic, we define interval extensions (Def 1) of
functions that may be expressed as a finite composition of elementary functions.

Definition 1 (Interval extension). Let f : Rn → R. F : IRn → IR is an

interval extension (or inclusion function) of f iff

∀X ∈ IRn , f (X) := {f (x) | x ∈ X} ⊂ F (X)

n n (3)
∀X ∈ IR , ∀Y ∈ IR , X ⊂ Y ⇒ F (X) ⊂ F (Y )

Interval extensions with various sharpnesses may be defined (Example 1).

The natural interval extension FN replaces the variables with their domains and
the elementary functions with their interval counterparts. The Taylor interval
extension FT is based on the Taylor expansion at point c ∈ X.

Example 1 (Interval extensions). Let f (x) = x2 −x, X = [−2, 0.5] and c = −1 ∈

X. The exact range is f (X) = [−0.25, 6] Then

– FN (X) = X 2 − X = [−2, 0.5]2 − [−2, 0.5] = [0, 4] − [−2, 0.5] = [−0.5, 6];
– FT (X, c) = 2 + (2X − 1)(X + 1) = 2 + [−5, 0][−1, 1.5] = [−5.5, 7].

Example 1 shows that interval arithmetic often overestimates the range of a

real-valued function. This is due to the dependency problem, an inherent behavior
of interval arithmetic. Dependency decorrelates multiple occurrences of the same
variable in an analytical expression (Example 2).

Example 2 (Dependency). Let X = [−5, 5]. Then

X − X = [−10, 10] = {x1 − x2 | x1 ∈ X, x2 ∈ X}

(4)
⊃ {x − x | x ∈ X} = {0}

Interval extensions (FN , FT ) have different convergence orders, that is the

overestimation decreases at different speeds with the width of the interval.
450 C. Vanaret et al.

3.2 Global Optimization

Interval arithmetic computes a rigorous enclosure of the range of a function

over a box. The first branch and bound algorithms for continuous optimization
based on interval arithmetic were devised in the 1970s [20][27], then refined
during the following years [14]: the search-space is partitioned into subboxes.
The objective function and the constraints are evaluated on each subbox using
interval arithmetic. The subspaces that cannot contain a global minimizer are
discarded and are not further explored. The algorithm terminates when f˜−f ∗ <
ε.
To overcome the pessimistic enclosures of interval arithmetic, interval branch
and bound algorithms have recently been endowed with filtering algorithms
(Section 3.3) that narrow the bounds of the boxes without loss of solutions.
Stemming from the Interval Analysis and Interval Constraint Programming com-
munities, filtering (or contraction) algorithms discard values from the domains by
enforcing local (each constraint individually) or global (all constraints simultane-
ously) consistencies. The resulting methods, called interval branch and contract
(IBC) algorithms, interleave steps of contraction and steps of bisection.

3.3 Interval Contractors

State-of-the-art contractors (contraction algorithms) include HC4 [6], Box [32],

Mohc [2], 3B [17], CID [31] and X-Newton [3]. Only HC4 and X-Newton are
used in this communication.
HC4Revise is a two-phase algorithm that exploits the syntax tree of a con-
straint to contract each occurrence of the variables. The first phase (evaluation)
evaluates each node (elementary function) using interval arithmetic. The second
phase (propagation) uses projection functions to inverse each elementary func-
tion. HC4Revise is generally invoked as the revised procedure (subcontractor)
of HC4, an AC3-like propagation loop.
X-Newton computes an outer linear relaxation of the objective function and
the constraints, then computes a lower bound of the initial problem using LP
techniques (e.g. the simplex algorithm). 2n additional calls may contract the
domains of the variables.

4 Charibde: A Cooperative Approach

4.1 Hybridization of Stochastic and Deterministic Techniques

Our hybrid algorithm Charibde (Cooperative Hybrid Algorithm using Reliable

Interval-Based methods and Differential Evolution), written in OCaml [16],
combines a stochastic DE and a deterministic IBC for non-convex constrained
optimization. Although it embeds a stochastic component, Charibde is a fully
rigorous solver.
 Hybridization of Interval CP and Evolutionary Algorithms 451

Previous Work. Preliminary results of a basic version of Charibde were pub-

lished in 2013 [33] on classical multimodal problems (7 bound-constrained and 4
inequality-constrained problems) widely used in the Evolutionary Computation
community. We showed that Charibde benefited from the start of convergence
of the DE algorithm, and completed the proof of convergence faster than a stan-
dard IBC algorithm. We provided new optimal results for 3 problems (Rana,
Eggholder and Michalewicz).

Contributions. In this communication, we present two contributions:

1. we devised a new cooperative exploration strategy MaxDist that

– selects boxes to be explored in a novel manner;
– periodically reduces DE’s domain;
– restarts the population within the new (smaller) domain.
An example illustrates the evolution of the domain without loss of solutions;
2. we assess the performance of Charibde against state-of-the-art rigorous
(GlobSol, IBBA, Ibex) and non-rigorous (Couenne, BARON) solvers on a
benchmark of difficult problems.

Cooperative Scheme. Two independent parallel processes exchange bounds,

solutions and search-space via MPI message passing (Figure 2).

Differential Interval Branch and

Evolution Contract
updates
best individual best upper bound

population punctual solution

injected into
domain subspaces
reduce

Fig. 2. Cooperative scheme of Charibde

The cooperation boils down to three main steps:

1. whenever the DE improves its best evaluation, the best individual and its
evaluation are sent to the IBC to update the best known upper bound f˜;
2. whenever the IBC finds a better punctual solution (e.g. the center of a box),
it is injected into DE’s population;
3. the exploration strategy MaxDist periodically reduces the search-space of
DE, then regenerates the population in the new search-space.

Sections 4.2 and 4.3 detail the particular implementations of the DE (Algo-
rithm 1) and the IBC (Algorithm 2) within Charibde.
452 C. Vanaret et al.

Algorithm 1. Charibde: Differential Evolution

function DifferentialEvolution(f : objective function, C: system of con-
straints, D: search-space, NP : size of population, W : amplitude factor, CR:
crossover rate)
P ← initial population, randomy generated in D
f˜ ← +∞
repeat
(x, fx ) ← MPI ReceiveIBC()
Insert x into P
f˜ ← fx
Generate temporary population P  by crossover
P ← P
(xbest , fbest ) ← BestIndividual(P )
if fbest < f˜ then
f˜ ← fbest
MPI SendIBC(xbest , fbest )
end if
until termination criterion is met
return best individual of P
end function

Algorithm 2. Charibde: Interval Branch and Contract

function IntervalBranchAndContract(F : objective function, C: system of con-
straints, D: search-space, ε: precision)
f˜ ← +∞  best known upper bound
Q ← {D}  priority queue
while Q = ∅ do
(xDE , fDE ) ← MPI ReceiveDE()
f˜ ← min(f˜, fDE )
Extract a box X from Q
Contract X w.r.t. constraints  Algorithm 3
if X cannot be discarded then
if F (m(X)) < f˜ then  midpoint test
f˜ ← F (m(X))  update best upper bound
MPI SendDE(m(X), F (m(X)))
end if
Split X into {X1 , X2 }
Insert {X1 , X2 } into Q
end if
end while
return (f˜, x̃)
end function

4.2 Differential Evolution

Population-based metaheuristics, in particular DE, are endowed with mecha-

nisms that help escape local minima. They are quite naturally recommended to
 Hybridization of Interval CP and Evolutionary Algorithms 453

solve difficult multimodal problems for which traditional methods struggle to

converge. They are also capable of generating feasible solutions without any a
priori knowledge of the topology. DE has proven greatly beneficial for improving
the best known upper bound f˜, a task for which standard branch and bound
algorithms are not intrinsically intended.

Base Individual. In the standard DE strategy, all the current individuals have
the same probability to be selected as the base individual u. We opted for an
alternative strategy [23] that guarantees that all individuals of the population
play this role once and only once at each generation: the index of the base
individual is obtained by summing the index of the individual x and an offset
in {1, . . . , NP − 1}, drawn with uniform probability.

Bound Constraints. When a coordinate yi of y (computed during the

crossover) exceeds the bounds of the component Di of the domain D, the bounce-
back method [23] replaces yi with a valid coordinate yi that lies between the base
coordinate ui and the violated bound:

ui + ω(Di − ui ) if yi > Di
yi = (5)
ui + ω(Di − ui ) if yi < Di

where ω is drawn in [0, 1] with uniform probability.

Constraint Handling. The extension of evolutionary algorithms to con-

strained optimization has been addressed by numerous authors. We implemented
the direct constraint handling [23] that assigns to each individual a vector of eval-
uations (objective function and constraints), and selects the new individual y
(see Section 2) based upon simple rules:

– x and y are feasible and y has a lower or equal objective value than x;
– y is feasible and x is not;
– x and y are infeasible, and y does not violate any constraint more than x.

Rigorous Feasibility. Numerous NLP solvers tolerate a slight violation (relax-

ation) of the inequality constraints (e.g. g ≤ 10−6 instead of g ≤ 0). The evalu-
ation of a “pseudo-feasible” solution x (that satisfies such relaxed constraints)
is not a rigorous upper bound of the global minimum; roundoff errors may even
lead to absurd conclusions: f (x) may be lower than the global minimum, and
(or) x may be very distant from actual feasible solutions in the search-space.
To ensure that an individual x is numerically feasible (i.e. that the evaluations
of the constraints are nonpositive), we evaluate the constraints gi using interval
arithmetic. x is considered as feasible when the interval evaluations Gi (x) are
nonpositive, that is ∀i ∈ {1, . . . , m}, Gi (x) ≤ 0.
454 C. Vanaret et al.

Rigorous Objective Function. When x is a feasible point, the evaluation

f (x) may be subject to roundoff errors; the only reliable upper bound of the
global minimum available is F (x) (the right bound of the interval evalua-
tion). However, evaluating the individuals using only interval arithmetic is much
costlier than cheap floating-point arithmetic.
An efficient in-between solution consists in systematically computing the
floating-point evaluations f (x), and computing the interval evaluation F (x) only
when the best known approximate evaluation is improved. F (x) is then com-
pared to the best known reliable upper bound f˜: if f˜ is improved, F (x) is sent
to the IBC. This implementation greatly reduces the cost of evaluations, while
ensuring that all the values sent to the IBC are rigorous.

4.3 Interval Branch and Contract

Branching aims at refining the computation of lower bounds of the functions
using interval arithmetic. Two strategies may be found in the early literature:
– the variable with the largest domain is bisected;
– the variables are bisected one after the other in a round-robin scheme.
More recently, the Smear heuristic [10] has emerged as a competitive alternative
to the two standard strategies. The variable xi for which the interval quantity
∂F
∂xi (X)(Xi − xi ) is the largest is bisected.
Charibde’s main contractor is detailed in Algorithm 3. We exploit the con-
tracted nodes of HC4Revise to compute partial derivatives via automatic dif-
ferentiation [26]. HC4Revise is a revise procedure within a quasi-fixed point
algorithm with tolerance η ∈ [0, 1]: the propagation loop stops when the box
X is not sufficiently contracted, i.e. when the size of X becomes larger than a
fraction ηw0 of the initial size w0 . Most contractors include an evaluation phase
that yields a lower bound of the problem on the current box. Charibde thus
computes several lower bounds (natural, Taylor, LP) as long as the box is not
discarded. Charibde calls ocaml-glpk [18], an OCaml binding for GLPK (GNU
Linear Programming Kit). Since the solution of the linear program is computed
using floating-point arithmetic, it may be subject to roundoff errors. A cheap
postprocessing step [21] computes a rigorous bound on the optimal solution of
the linear program, thus providing a rigorous lower bound of the initial problem.
When the problem is subject to equality constraints hj (j ∈ {1, . . . , p}),
IBBA [22], Ibex [30] and Charibde handle a relaxed problem where each equality
constraint hj (x) = 0 (j ∈ {1, . . . , p}) is replaced by two inequalities:
−ε= ≤ hj (x) ≤ ε= (6)
ε= may be chosen arbitrarily small.

4.4 MaxDist: A New Exploration Strategy

The boxes that cannot be discarded are stored in a priority queue Q to be
processed at a later stage. The order in which the boxes are extracted determines
 Hybridization of Interval CP and Evolutionary Algorithms 455

Algorithm 3. Charibde: contractor for constrained optimization

function Contraction(in-out X: box, F : objective function, in-out f˜: best upper
bound, in-out C: system of constraints)
lb ← −∞  lower bound
repeat
w0 ← w(X)  initial size
FX ← HC4Revise(F (X) ≤ f˜)  evaluation of f /contraction
lb ← FX  lower bound by natural form
G ← ∇F (X)  gradient by AD
lb ← max(lb, SecondOrder(X, F, f˜, G))  second-order form
C ← HC4(X, C, η)  quasi-fixed point with tolerance η
if use linearization then
lb ← max(lb, Linearization(X, F, f˜, G, C))  simplex or X-Newton
end if
until X = ∅ or w(X) > ηw0
return lb
end function

the exploration strategy of the search-space (“best-first”, “largest first”, “depth-

first”). Numerical tests suggest that

– the “best-first” strategy is rarely relevant because of the overestimated range

(due to the dependency problem);
– the “largest first” strategy does not give advantage to promising regions;
– the “depth-first” strategy tends to quickly explore the neighborhood of local
minima, but struggles to escape from them.

We propose a new exploration strategy called MaxDist. It consists in extract-

ing from Q the box that is the farthest from the current solution x̃. The under-
lying ideas are to

– explore the neighborhood of the global minimizer (a tedious task when the
objective function is flat in this neighborhood) only when the best possible
upper bound is available;
– explore regions of the search-space that are hardly accessible by the DE
algorithm.

The distance between a point x and a box X is the sum of the distances between
each coordinate xi and the closest bound of Xi . Note that MaxDist is an adap-
tive heuristic: whenever the best known solution x̃ is updated, Q is reordered
according to the new priorities of the boxes.
Preliminary results (not presented here) suggest that MaxDist is competi-
tive with standard strategies. However, the most interesting observation lies in
the behavior of Q: when using MaxDist, the maximum size of Q (the maxi-
mum number of boxes simultaneously stored in Q) remains remarkably low (a
few dozens compared to several thousands for standard strategies). This offers
promising perspectives for the cooperation between DE and IBC: the remaining
456 C. Vanaret et al.

boxes of the IBC may be exploited in the DE to avoid exploring regions of the
search-space that have already been proven infeasible or suboptimal.
The following numerical example illustrates how the remaining boxes are
exploited to reduce DE’s domain through the generations. Let

(x + y − 10)2 (x − y + 10)2
min − −
(x,y)∈(X,Y ) 30 120
20 (7)
s.t. −y ≤0
x2
x2 + 8y − 75 ≤ 0

be a constrained optimization problem defined on the box X ×Y = [0, 10]×[0, 10]

(Figure 3a). The dotted curves represent the frontiers of the two inequality
constraints, and the contour lines of the objective function are shown in solid
dark. The feasible region is the banana-shaped set, and the global minimizer is
located in its lower right corner.
The initial domain of DE (which corresponds to the initial box in the IBC)
is first contracted with respect to the constraints of the problem. The initial
population of DE is then generated within this contracted domain, thus avoiding
obvious infeasible regions of the search-space. This approach is similar to that
of [11]. Figure 3b depicts the contraction (the black rectangle) of the initial
domain with respect to the constraints (sequentially handled by HC4Revise):
X × Y = [1.4142, 8.5674] × [0.2, 9.125].
Periodically, we compute the convex hull (Q) of the remaining boxes of Q
and replace DE’s domain with (Q). Note that
1. the convex hull (linear complexity) may be computed at low cost, because
the size of Q remains small when using MaxDist;
2. by construction, MaxDist handles boxes on the rim of the remaining domain
(the boxes of Q), which boosts the reduction of the convex hull.
Figures 3c and 3d represent the convex hull (Q) of the remaining subboxes in
the IBC, respectively after 10 and 20 DE generations. The population is then
randomly regenerated within the new contracted domain (Q). The convex
hull operation progressively eliminates local minima and infeasible regions. The
global minimum eventually found by Charibde with precision ε = 10−8 is f˜ =
f (8.532424, 0.274717) = −2.825296148; both constraints are active.

5 Experimental Results
Currently, GlobSol [15], IBBA [22] and Ibex [8] are among the most efficient
solvers in rigorous constrained optimization. They share a common skeleton of
interval branch and bound algorithm, but differ in the acceleration techniques.
GlobSol uses the reformulation-linearization technique (RLT), that introduces
new auxiliary variables for each intermediary operation. IBBA calls a contractor
similar to HC4Revise, and computes a relaxation of the system of constraints
 Hybridization of Interval CP and Evolutionary Algorithms 457

10.0

7.5

5.0
y

2.5

0.0

Fig. 3. Evolution of DE’s domain with the number of generations

using affine arithmetic. Ibex is dedicated to both numerical CSPs and con-
strained optimization; it embeds most of the aforementioned contractors (HC4,
3B, Mohc, CID, X-Newton). Couenne [5] and BARON [25] are state-of-the-art
NLP solvers. They are based on a non-rigorous spatial branch and bound algo-
rithm, in which the objective function and the constraints are over- and underes-
timated by convex relaxations. Although they perform an exhaustive exploration
of the search-space, they cannot guarantee a given precision on the value of the
optimum.
All five solvers and Charibde are compared on a subset of 11 COCONUT con-
strained problems (Table 1), extracted by Araya [3] for their difficulty: ex2 1 7,
ex2 1 9, ex6 2 6, ex6 2 8, ex6 2 9, ex6 2 11, ex6 2 12, ex7 2 3, ex7 3 5, ex14 1 7
and ex14 2 7. Because of numerical instabilities of the ocaml-glpk LP library
458 C. Vanaret et al.

(“assert failure”), the results of the problems ex6 1 1, ex6 1 3 and ex 6 2 10 are
not presented. The second and third columns give respectively the number of
variables n and the number of constraints m. The fourth (resp. fifth) column
specifies the type of the objective function (resp. the constraints): L is linear, Q
n and NL is nonlinear. The logsize of the domain D (sixth column)
is quadratic
is log( i=1 (Di − Di )).
The comparison of CPU times (in seconds) for solvers GlobSol, IBBA, Ibex,
Couenne, BARON and Charibde on the benchmark of 11 problems is detailed in
Table 2. Mean times and standard deviations (in brackets) are given for Charibde
over 100 runs. The numerical precision on the objective function ε = 10−8 and
the tolerance for equality constaints ε= = 10−8 were identical for all solvers.
TO (timeout) indicates that a solver could not solve a problem within one hour.
The results of GlobSol (proprietary piece of software) were not available for
all problems; only those mentioned in [22] are presented. The results of IBBA
were also taken from [22]. The results of Ibex were taken from [3]: only the
best strategy (simplex, X-NewIter or X-Newton) for each benchmark problem
is presented. Couenne and BARON (only the commercial version of the code is
available) were run on the NEOS server [13].

Table 1. Description of difficult COCONUT problems

Type
Problem n m f gi , hj Domain logsize
ex2 1 7 20 10 Q L +∞
ex2 1 9 10 1 Q L +∞
ex6 2 6 3 1 NL L −3 · 10−6
ex6 2 8 3 1 NL L −3 · 10−6
ex6 2 9 4 2 NL L −2.77
ex6 2 11 3 1 NL L −3 · 10−6
ex6 2 12 4 2 NL L −2.77
ex7 2 3 8 6 L NL 61.90
ex7 3 5 13 15 L NL +∞
ex14 1 7 10 17 L NL 23.03
ex14 2 7 6 9 L NL +∞

Charibde was run on an Intel Xeon E31270 @ 3.40GHz x 8 with 7.8 GB of

RAM. BARON and Couenne were run on 2 Intel Xeon X5660 @ 2.8GHz x 12
with 64 GB of RAM. IBBA and Ibex were run on similar processors (Intel x86,
3GHz). The difference in CPU time between computers is about 10% [4], which
makes the comparison quite fair.
The hyperparameters of Charibde for the benchmark problems are given in
Table 3; NP is the population size, and η is the quasi-fixed point ratio. The
amplitude W = 0.7, the crossover rate CR = 0.9 and the MaxDist strategy
are common to all problems. Tuning the hyperparameters is generally problem-
dependent, and requires structural knowledge about the problem: the population
size NP may be set according to the dimension and the number of local minima,
 Hybridization of Interval CP and Evolutionary Algorithms 459

Table 2. Comparison of convergence times (in seconds) between GlobSol, IBBA, Ibex,
Charibde (mean and standard deviation over 100 runs), Couenne and BARON on
difficult constrained problems
Rigorous Non rigorous
Problem GlobSol IBBA Ibex Charibde Couenne BARON
ex2 1 7 16.7 7.74 34.9 (13.3) 476 16.23
ex2 1 9 154 9.07 35.9 (0.29) 3.01 3.58
ex6 2 6 306 1575 136 3.3 (0.41) TO 5.7
ex6 2 8 204 458 59.3 2.9 (0.37) TO TO
ex6 2 9 463 523 25.2 2.7 (0.03) TO TO
ex6 2 11 273 140 7.51 1.96 (0.06) TO TO
ex6 2 12 196 112 22.2 8.8 (0.17) TO TO
ex7 2 3 TO 544 1.9 (0.30) TO TO
ex7 3 5 TO 28.91 4.5 (0.09) TO 4.95
ex14 1 7 TO 406 4.2 (0.13) 13.86 0.56
ex14 2 7 TO 66.39 0.2 (0.04) 0.01 0.02
Sum > 1442 TO 1312.32 101.26 TO TO

the crossover rate CR is related to the separability of the problem, and the
techniques based on linear relaxation have little influence for problems with few
constraints, but are cheap when the constraints are linear.

Table 3. Hyperparameters of Charibde for the benchmark problems

Problem NP Bisections Fixed-point ratio (η) LP X-Newton
ex2 1 7 20 RR 0.9  
ex2 1 9 100 RR 0.8 
ex6 2 6 30 Smear 0 
ex6 2 8 30 Smear 0 
ex6 2 9 70 Smear 0
ex6 2 11 35 Smear 0
ex6 2 12 35 RR 0 
ex7 2 3 40 Largest 0  
ex7 3 5 30 RR 0 
ex14 1 7 40 RR 0 
ex14 2 7 40 RR 0 

Charibde outperforms Ibex on 9 out of 11 problems, IBBA on 10 out of 11

problems and GlobSol on all the available problems. The cumulated CPU time
shows that Charibde (101.26s) improves the performances of Ibex (1312.32s) by
an order of magnitude (ratio: 13) on this benchmark of 11 difficult problems.
Charibde also proves highly competitive against non-rigorous solvers Couenne
and BARON. The latter are faster or have similar CPU times on some of the
11 problems, however they both timeout on at least five problems (seven for
Couenne, five for BARON). Overall, Charibde seems more robust and solves all
460 C. Vanaret et al.

the problems of the benchmark, while providing a numerical proof of optimality.

Surprisingly, the convergence times do not seem directly related to the dimen-
sions of the instances. They may be explained by the nature of the objective
function and constraints (in particular, Charibde seems to struggle when the
objective function is quadratic) and the dependency induced by the multiple
occurrences of the variables.
Table 4 presents the best upper bounds obtained by Charibde, Couenne
and BARON at the end of convergence (precision reached or timeout). Trun-
cated digits on the upper bounds are bounded (e.g. 1.2387 denotes [1.237, 1.238]
and −1.2387 denotes [−1.238, −1.237]). The incorrect digits of the global minima
obtained by Couenne and BARON are underlined. This demonstrates that non-
rigorous solvers may be affected by roundoff errors, and may provide solutions
that are infeasible or have an objective value lower than the global minimum
(Couenne on ex2 1 9, BARON on ex2 1 7, ex2 1 9, ex6 2 8, ex6 2 12, ex7 2 3
and ex7 3 5). For the most difficult instance ex7 2 3, Couenne is not capable
of finding a feasible solution with a satisfactory evaluation within one hour. It
would be very informative to compute the ratio between the size of the feasible
domain (the set of all feasible points) and the size of the entire domain. On the
other hand, the strategy MaxDist within Charibde greatly contributes to finding
an excellent upper bound of the global minimum, which significantly accelerates
the interval pruning phase.

Table 4. Best upper bounds obtained by Charibde, Couenne and BARON

Problem Charibde Couenne BARON
ex2 1 7 −4150.4101339298 −4150.4101273187 −4150.4101607987
ex2 1 9 −0.3750000075 −0.3750001543 −0.3750011110
ex6 2 6 −0.0000026032 0.0000007110 −0.0000026032
ex6 2 8 −0.027006350
49 −0.027006350
49 −0.0270063710
ex6 2 9 −0.0340661854 −0.034066184 −0.0340661910
ex6 2 11 −0.0000026732 −0.0000026732 −0.0000026732
ex6 2 12 0.289194740
39 0.28919475 0.2891916998
ex7 2 3 7049.2480205298 1050 7049.0202917076
ex7 3 5 1.2067169921 1.2068965 0.2398244887
ex14 1 7 0.000000010
09 0.0000000010 0
ex14 2 7 0.0000000087 0.0000000010 0

6 Conclusion

We proposed a cooperative hybrid solver Charibde, in which a deterministic

interval branch and contract cooperates with a stochastic differential evolution
algorithm. The two independent algorithms run in parallel and exchange bounds,
solutions and search-space in an advanced manner via message passing. The
domain of the population-based metaheuristic is periodically reduced by remov-
ing local minima and infeasible regions detected by the branch and bound.
 Hybridization of Interval CP and Evolutionary Algorithms 461

A comparison of Charibde with state-of-the-art interval-based solvers (Glob-

Sol, IBBA, Ibex) and NLP solvers (Couenne, BARON) on a benchmark of dif-
ficult COCONUT problems shows that Charibde is highly competitive against
non-rigorous solvers (while bounding the global minimum) and converges faster
than rigorous solvers by an order of magnitude.

References
1. Alliot, J.M., Durand, N., Gianazza, D., Gotteland, J.B.: Finding and proving
the optimum: cooperative stochastic and deterministic search. In: 20th European
Conference on Artificial Intelligence (ECAI 2012), Montpellier, France, August
27–31, 2012 (2012)
2. Araya, I., Trombettoni, G., Neveu, B.: Exploiting monotonicity in interval con-
straint propagation. In: Proc. AAAI, pp. 9–14 (2010)
3. Araya, I., Trombettoni, G., Neveu, B.: A contractor based on convex interval taylor.
In: Beldiceanu, N., Jussien, N., Pinson, É. (eds.) CPAIOR 2012. LNCS, vol. 7298,
pp. 1–16. Springer, Heidelberg (2012)
4. Araya, I., Trombettoni, G., Neveu, B., Chabert, G.: Upper bounding in inner
regions for global optimization under inequality constraints. Journal of Global
Optimization 60(2), 145–164 (2014)
5. Belotti, P., Lee, J., Liberti, L., Margot, F., Wächter, A.: Branching and bounds
tighteningtechniques for non-convex minlp. Optimization Methods & Software
24(4–5), 597–634 (2009)
6. Benhamou, F., Goualard, F., Granvilliers, L., Puget, J.F.: Revising hull and box
consistency. In: International Conference on Logic Programming, pp. 230–244. MIT
press (1999)
7. Blum, C., Puchinger, J., Raidl, G.R., Roli, A.: Hybrid metaheuristics in combina-
torial optimization: A survey. Applied Soft Computing 11(6), 4135–4151 (2011)
8. Chabert, G., Jaulin, L.: Contractor programming. Artificial Intelligence 173,
1079–1100 (2009)
9. Cotta, C., Troya, J.M.: Embedding branch and bound within evolutionary algo-
rithms. Applied Intelligence 18(2), 137–153 (2003)
10. Csendes, T., Ratz, D.: Subdivision direction selection in interval methods for global
optimization. SIAM Journal on Numerical Analysis 34(3), 922–938 (1997)
11. Focacci, F., Laburthe, F., Lodi, A.: Local search and constraint programming. In:
Handbook of metaheuristics, pp. 369–403. Springer (2003)
12. Gallardo, J.E., Cotta, C., Fernández, A.J.: On the hybridization of memetic algo-
rithms with branch-and-bound techniques. IEEE Transactions on Systems, Man,
and Cybernetics, Part B: Cybernetics 37(1), 77–83 (2007)
13. Gropp, W., Moré, J.: Optimization environments and the NEOS server. Approxi-
mation theory and optimization, 167–182 (1997)
14. Hansen, E.: Global optimization using interval analysis. Dekker (1992)
15. Kearfott, R.B.: Rigorous global search: continuous problems. Springer (1996)
16. Leroy, X., Doligez, D., Frisch, A., Garrigue, J., Rémy, D., Vouillon, J.: The objective
caml system release 3.12. Documentation and userâĂŹs manual. INRIA (2010)
17. Lhomme, O.: Consistency techniques for numeric csps. In: IJCAI, vol. 93,
pp. 232–238. Citeseer (1993)
18. Mimram, S.: ocaml-glpk (2004). http://ocaml-glpk.sourceforge.net/
19. Moore, R.E.: Interval Analysis. Prentice-Hall (1966)
462 C. Vanaret et al.

20. Moore, R.E.: On computing the range of a rational function of n variables over a
bounded region. Computing 16(1), 1–15 (1976)
21. Neumaier, A., Shcherbina, O.: Safe bounds in linear and mixed-integer linear pro-
gramming. Mathematical Programming 99(2), 283–296 (2004)
22. Ninin, J., Hansen, P., Messine, F.: A reliable affine relaxation method for global
optimization. Groupe d’études et de recherche en analyse des décisions (2010)
23. Price, K., Storn, R., Lampinen, J.: Differential Evolution - A Practical Approach
to Global Optimization. Natural Computing, Springer-Verlag (2006)
24. Puchinger, J., Raidl, G.R.: Combining metaheuristics and exact algorithms in com-
binatorial optimization: a survey and classification. In: Mira, J., Álvarez, J.R. (eds.)
IWINAC 2005. LNCS, vol. 3562, pp. 41–53. Springer, Heidelberg (2005)
25. Sahinidis, N.V.: Baron: A general purpose global optimization software package.
Journal of Global Optimization 8(2), 201–205 (1996)
26. Schichl, H., Neumaier, A.: Interval analysis on directed acyclic graphs for global
optimization. Journal of Global Optimization 33(4), 541–562 (2005)
27. Skelboe, S.: Computation of rational interval functions. BIT Numerical Mathemat-
ics 14(1), 87–95 (1974)
28. Sotiropoulos, D., Stavropoulos, E., Vrahatis, M.: A new hybrid genetic algo-
rithm for global optimization. Nonlinear Analysis: Theory, Methods & Applications
30(7), 4529–4538 (1997)
29. Storn, R., Price, K.: Differential evolution - a simple and efficient heuristic
for global optimization over continuous spaces. Journal of Global Optimization,
341–359 (1997)
30. Trombettoni, G., Araya, I., Neveu, B., Chabert, G.: Inner regions and interval
linearizations for global optimization. In: AAAI (2011)
31. Trombettoni, G., Chabert, G.: Constructive interval disjunction. In: Bessière, C.
(ed.) CP 2007. LNCS, vol. 4741, pp. 635–650. Springer, Heidelberg (2007)
32. Van Hentenryck, P.: Numerica: a modeling language for global optimization. MIT
press (1997)
33. Vanaret, C., Gotteland, J.-B., Durand, N., Alliot, J.-M.: Preventing premature con-
vergence and proving the optimality in evolutionary algorithms. In: Legrand, P.,
Corsini, M.-M., Hao, J.-K., Monmarché, N., Lutton, E., Schoenauer, M. (eds.) EA
2013. LNCS, vol. 8752, pp. 29–40. Springer, Heidelberg (2014)
34. Zhang, X., Liu, S.: A new interval-genetic algorithm. In: Third International
Conference on Natural Computation, ICNC 2007, vol. 4, pp. 193–197. IEEE (2007)
 A General Framework for Reordering Agents
Asynchronously in Distributed CSP

Mohamed Wahbi1(B) , Younes Mechqrane2 , Christian Bessiere3 ,

and Kenneth N. Brown1
1
Insight Centre for Data Analytics, University College Cork, Cork, Ireland
{mohamed.wahbi,ken.brown}@insight-centre.org
2
Mohammed V University-Agdal Rabat, Rabat, Morocco
ymechqrane@gmail.com
3
University of Montpellier, Montpellier, France
bessiere@lirmm.fr

Abstract. Reordering agents during search is an essential component

of the efficiency of solving a distributed constraint satisfaction prob-
lem. Termination values have been recently proposed as a way to sim-
ulate the min-domain dynamic variable ordering heuristic. The use of
termination values allows the greatest flexibility in reordering agents
dynamically while keeping polynomial space. In this paper, we propose
a general framework based on termination values for reordering agents
asynchronously. The termination values are generalized to represent var-
ious heuristics other than min-domain. Our general framework is sound,
complete, terminates and has a polynomial space complexity. We imple-
mented several variable ordering heuristics that are well-known in cen-
tralized CSPs but could not until now be applied to the distributed
setting. Our empirical study shows the significance of our framework
compared to state-of-the-art asynchronous dynamic ordering algorithms
for solving distributed CSP.

1 Introduction

Distributed artificial intelligence involves numerous combinatorial problems

where multiple entities, called agents, need to cooperate in order to find a con-
sistent combination of actions. Agents have to achieve the combination in a
distributed manner and without any centralization. Examples of such problems
are: traffic light synchronization [12], truck task coordination [18], target track-
ing in distributed sensor networks [11], distributed scheduling [16], distributed
planning [6], distributed resource allocation [19], distributed vehicle routing [13],
etc. These problems were successfully formalized using the distributed constraint
satisfaction problem (DisCSP) paradigm.
DisCSP are composed of multiple agents, each owning its local constraint
network. Variables in different agents are connected by constraints. Agents must
assign values to their variables distributively so that all constraints are satisfied.
To achieve this goal, agents assign values to their variables that satisfy their

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 463–479, 2015.
DOI: 10.1007/978-3-319-23219-5 33
464 M. Wahbi et al.

own constraints, and exchange messages to satisfy constraints with variables

owned by other agents. During the last two decades, many algorithms have been
designed for solving DisCSP. Most of these algorithms assume a total (priority)
order on agents that is static. However, it is known from centralized CSPs that
reordering variables dynamically during search improves the efficiency of the
search procedure. Moreover, reordering agents in DisCSP may be required in
various applications (e.g., security [24]).
Agile Asynchronous Backtracking (AgileABT) [1] is a polynomial space asyn-
chronous algorithm that is able to reorder all agents in the problem. AgileABT is
based on the notion of termination value, a tuple of agents’ domain sizes. Besides
implementing the min-domain ([10]) dynamic variable ordering heuristic (DVO),
the termination value acts as a timestamp for the orders exchanged by agents
during search. Since the termination of AgileABT depends on the form of the
termination value, the use of other forms of termination values (implementing
other heuristics) may directly affect the termination of the algorithm.
In this paper, we generalize AgileABT to get a new framework AgileABT(α),
in which α represents any measure used to implement a DVO (e.g., domain
size in min-domain). AgileABT(α) is sound and complete. We define a simple
condition on the measure α which guarantees that AgileABT(α) terminates. If
the computation of the measure α also has polynomial space complexity, then
AgileABT(α) has polynomial space complexity. This allows us to implement
for the first time in Distributed CSP a wide variety of DVOs that have been
studied in centralized CSP. To illustrate this, we implement a number of DVOs,
including dom/deg [3] and dom/wdeg [14], and evaluate their performance on
benchmark DisCSP problems.
The paper is organized as follows. Section 2 gives the necessary background
on distributed CSP and dynamic reordering materials. It then discusses the
Agile Asynchronous Backtracking algorithm for solving distributed CSP. Our
general framework is presented and analyzed in Section 3. We show our empirical
results in Section 4 and report related work in Section 5. Finally, we conclude
in Section 6.

2 Background
2.1 Distributed Constraint Satisfaction Problem
The Distributed Constraint Satisfaction Problem (DisCSP) is a 5-tuple
(A, X , D, C, ϕ), where A is a set of agents {A1 , . . . , Ap }, X is a set of variables
{x1 , . . . , xn }, D = {D1 , . . . , Dn } is a set of domains, where Di is the initial set
of possible values which may be assigned to variable xi , C is a set of constraints,
and ϕ : X → A is a function specifying an agent to control each variable. During
a solution process, only the agent which controls a variable can assign it a value.
A constraint C(X) ∈ C, on the ordered subset of variables X = (xj1 , . . . , xjk ),
is C(X) ⊆ Dj1 × · · · × Djk , and specifies the tuples of values which may be
assigned simultaneously to the variables in X. For this paper, we restrict atten-
tion to binary constraints. We denote by Ci ⊆ C all constraints that involve xi . A
 A General Framework for Reordering Agents 465

solution is an assignment to each variable of a value from its domain, satisfying

all constraints. Each agent Ai only knows constraints relevant to its variables
(Ci ) and the existence of other variables involved in these constraints (its neigh-
bors). Without loss of generality, we assume each agent controls exactly one
variable (p = n), so we use the terms agent and variable interchangeably and do
not distinguish between Ai and xi .
Each agent Ai stores a unique order, an ordered tuple of agents IDs, denoted
by λi . λi is called the current order of Ai . Agents appearing before Ai in λi are
the higher priority agents (predecessors) denoted by λ− i and conversely the lower
priority agents (successors) λ+ i are agents appearing after Ai in λi . We denote
by λi [k] (∀k ∈ 1..n) the ID of the agent located at position k in λi .

2.2 Asynchronous Backtracking - ABT

The first complete asynchronous search algorithm for solving DisCSP is Asyn-
chronous Backtracking (ABT) [2,27]. In ABT, agents act concurrently and asyn-
chronously, and do not have to wait for decisions of others. However, ABT
requires a total priority order among agents. Each agent tries to find an assign-
ment satisfying the constraints with what is currently known from higher prior-
ity agents. When an agent assigns a value to its variable, it sends out messages
to lower priority agents, with whom it is constrained, informing them about
its assignment. When no value is possible for a variable, the inconsistency is
reported to higher agents in the form of a no-good (an unfruitful value com-
bination). ABT computes a solution (or detects that no solution exists) in a
finite time. However, the priority order of agents is static and uniform across the
agents.

2.3 No-goods and Explanations

During a solution process, agents can infer inconsistent sets of assignments, called
no-goods. No-goods are used to justify value removals. A no-good ruling out value
vi from the domain of a variable xi is a clause of the form xj = vj ∧. . .∧xk = vk →
xi = vi . This no-good, which means that the assignment xi = vi is inconsistent
with the assignments xj = vj ∧. . .∧xk = vk , is used by agent Ai to justify removal
of vi from its domain Di . The left hand side (lhs) and the right hand side (rhs)
of a no-good are defined from the position of →. The variables in the lhs of
a no-good must precede the variable on its rhs in the current order because
the assignments of these variables have been used to filter the domain of the
variable in its rhs. These ordering constraints induced by a no-good are called
safety conditions in [9]. For example, the no-good xj = vj ∧ xk = vk → xi = vi
implies that xj ≺ xi and xk ≺ xi that is xj and xk must precede xi in the
variable ordering (i.e., xj , xk ∈ λ−
i ). We say that a no-good is compatible with
an order λi if all agents in its lhs appear before its rhs in λi .
The current domain of a variable xi , maintained by Ai , is composed by values
not ruled out by a no-good.1 The initial domain size (before search starts) of Ai
1
To stay polynomial, Ai keeps only one no-good per removed value.
466 M. Wahbi et al.

is denoted by d0i while its current domain size is denoted by di . Let Σi be the
conjunction of the left hand sides of all no-goods ruling out values from Di . We
explain the current domain size of Di by the following expression ei : Σi → di ,
called explanation of xi (ei ). Every explanation ei induces safety conditions:
{∀xm ∈ Σi , xm ≺ xi }. When all values of a variable xi are ruled out by some no-
goods (Σi → 0), these no-goods are resolved, producing a new no-good from Σi .
There are clearly many different ways of representing Σi as a directed no-good
(an implication). In standard backtracking search algorithms (like ABT), the
variable, say xt , that has the lowest priority in the current order (among variables
in Σi ) must change its value. xt is called the backtracking target and the directed
no-good is ngt : Σi \ xt → xt = vt . In AgileABT the backtracking target is not
necessarily the variable with the lowest priority within the conflicting variables
in the current order.

2.4 Agile Asynchronous Backtracking

In AgileABT, an order λ is always associated with a termination value τ. A

termination value is a tuple of positive integers (representing the sizes of the
domains of other agents seen from Ai ). When comparing two orders the strongest
order is that associated with the lexicographically smallest termination value.
The lexicographic order on agents IDs (<lex ) is used to break ties, the smallest
being the strongest.
In AgileABT, all agents start with the same order. Then, every agent Ai
is allowed to change the order asynchronously. In the following we describe
AgileABT by illustrating the computation performed within agent Ai . Ai can
change its current order λi only if it receives a stronger one from another agent or
if itself proposes a new order (λi ) stronger than its current order λi . Ai can only
propose new orders (λi ) when it tries to backtrack after detecting a dead-end
(Σi → 0).
In AgileABT, agents exchange the following types of messages to coordinate
the search (where Ai is the sender):
– ok? message is sent by Ai to all lower agents (λ+
i ) to ask whether its assign-
ment is acceptable. Besides the assignment, the ok? message contains an
explanation ei which communicates the current domain size of xi , the cur-
rent order λi , and the current termination value τi stored by Ai .
– ngd message is sent by Ai when all its values are ruled out by Σi . This
message contains a directed no-good, as well as λi and τi .
– order message is sent to propose a new order. This message includes the
order λi proposed by Ai accompanied by the termination value τi .

Each agent needs to compute the size of the domain of other variables to
build its termination value. Hence, each agent Ai stores a set Ei of explanations
sent by other agents. During search, Ai updates Ei to store new received expla-
nations and to remove those that are no more relevant to the search state or
not compatible with its current order λi . If ek ∈ Ei , Ai uses this explanation
 A General Framework for Reordering Agents 467

Algorithm 1. Computing termination value using heuristic α.

function T V α (λ)
1. τ is an array of length n;
2. for ( j ← 1 to n ) do τ[j] ← α(λ[j]) ;
3. return τ;

to justify the size dom(k) of the current domain of xk , i.e., dk . Otherwise, Ai

assumes that the size of the current domain of xk is equal to d0k .
In AgileABT, the termination value τi = [tvi1 , . . . , tvin ] computed by agent
Ai is such that tvik = dom(λi [k]), ∀k ∈ 1..n. τi depends on the order λi and the
domain sizes of agents given by the set of explanations Ei (Algorithm 1, using
T V α (λi ) with α = dom).
In standard backtracking search algorithms, the backtracking target is always
the variable that has the lowest priority among the variables in the detected
conflict (i.e., Σi ). AgileABT relaxes this restriction by allowing Ai to select the
target of backtracking xt among conflicting variables Σi . The only restriction for
selecting xt as a backtracking target is to find an order λi such that τi = T V α (λi )
with α = dom (Algorithm 1) is lexicographically smaller than the termination
value associated with the current order λi and xt is the lowest among variables
in Σi w.r.t. λi .
When a dead-end occurs, AgileABT iterates through all variables xt ∈ Σi ,
considering xt as the target of the backtracking, i.e., the directed no-good
is ngt : Σi \ xt → xt = vt . Ai then updates Ei to remove all explanations
containing xt (after backtracking xt assignment will be changed). Next, it
updates the explanation of xt by considering the new generated no-good ngt
(i.e., et ← [Σt ∪ lhs(ngt ) → dt − 1]). Finally, Ai computes a new order
(λi ) and its associated termination value (τi ) from the updated explanations
Ei . λi is obtained by performing a topological sort on the directed acyclic
graph (G) formed by safety conditions induced by the updated explanations
Ei (∀xm ∈ Σk |ek ∈ Ei , (xm , xk ) ∈ G) and τi is obtained from T V α (λi ) with
α = dom (Algorithm 1). Let λi be the strongest computed order over all possi-
ble targets in Σi . If the termination value τi associated to λi is lexicographi-
cally smaller than the τi associated to the current order λi , Ai reorders agents
according to λi and informs all agents about the new order λi and its associated
termination value τi . The backtracking target is that used when Ai computed
λi . If no λi stronger than λi exists, the backtracking target xt is the variable
that has the lowest priority among Σi in the current order λi . For more details
we refer the reader to [1,25].

Example of Running AgileABT

Figure 1 presents an example of a possible execution of AgileABT on a simple
problem. This problem (fig. 1a) consists of 5 agents with the following domains
∀i ∈ 1..5, Di = {1, 2, 3, 4} and 6 constraints among these agents c12 : x1 = x2 , c13 :
468 M. Wahbi et al.

∀i ∈ 1..4
A2 x2 Di = {1, 2, 3, 4}
.=
.=

A1 x1 .x1 = |x5 −2| x5 A5

.=

.≥
A3 x3 .< x4 A4

(a) The constraint graph (b) Explanations and no-goods maintained by A5

(c) A5 : updated explanations with x1 as target (d) A5 : selection of target (x1 )

Fig. 1. An example of a possible execution of AgileABT on a simple problem.

x1 = x3 , c15 : x1 = |x5 −2|, c25 : x2 = x5 , c34 : x3 < x4 , and c45 : x4 ≥ x5 . All agents
start with the same initial ordering λi = [1, 2, 3, 4, 5] associated with the termi-
nation value τi = [4, 4, 4, 4, 4] and values are chosen lexicographically. Consider
the situation in A5 after receiving ok? messages from other agents (fig. 1b).
On receipt, explanations e1 , e2 , e3 , and e4 are stored in E5 , and assignments
x1 = 1, x2 = 2, x3 = 2, and x4 = 3 are stored in A5 agent-view. After checking
its constraints (c15 , c25 , and c45 ), A5 detects a dead-end (D5 = ∅) where Σ5 :
{x1 = 1 ∧ x2 = 2 ∧ x4 = 3}. A5 iterates through all variables xt ∈ Σ5 , considering
xt as the target of the backtracking. Figure 1c shows the updates on the expla-
nations stored in A5 (E5 ) when it considers x1 as the target of the backtracking
(i.e., xt = x1 ). A5 updates E5 to remove all explanations containing x1 (i.e., e2
and e3 ) and considering the new generated no-good ng1 in the explanation of
x1 , i.e., e1 (fig. 1c, left). Finally, A5 computes a new order (λ5 ) and its associ-
ated termination value (τ5 ) from the updated explanations E5 . λ5 is obtained
by performing a topological sort on the directed acyclic graph formed by safety
conditions induced by the updated explanations E5 (fig. 1c, right). Figure 1d
presents the computed orderings and their associated termination values (by
topological sort) when considering each xt ∈ Σ5 as backtracking target. The
strongest computed order (e.g, λ5 = [3, 4, 2, 5, 1], τ5 = [4, 2, 4, 2, 3]) is that com-
puted when considering x1 as backtracking target. Since λ5 is stronger than λ5 ,
A5 changes its current order to λ5 and proposes this ordering to all other agents
 A General Framework for Reordering Agents 469

Algorithm 2. AgileABT(α): Procedures for changing the order by Ai .

procedure processOrder(λj , τj )
tv
1. if ( τj ≺tv τi ∨ (τj = τi ∧ λj <lex λi ) ) then changeOrder(λj , τj ) ;
procedure proposeOrder(args)
 
2. λi , τi  ← computeNewOrder(args);

3. if ( τi ≺tv τi ) then
4. changeOrder(λi , τi ) ;
5. sendMsg: order λi , τi  to all agents in A ;
procedure changeOrder(λ , τ  )

6. λi ← λ ; τi ← τ  ;
7. remove no-goods and explanations incompatible with τi ;

through order messages (i.e., order : λ5 , τ5 ). Then, A5 sends the no-good ng1
to agent x1 .
It has been proved that AgileABT is sound, complete and terminates [1,25].
The termination proofs of AgileABT are based on the fact that the termina-
tion value is a tuple of positive integers (representing the expected sizes of the
domains of other agents) and, as search progresses, these tuples can only decrease
lexicographically. Thus, any change to the form of the termination values (i.e.
implemented heuristic) may directly affect the termination of AgileABT.

3 Generalized AgileABT
In AgileABT, the termination value can be seen as an implementation of the dom
dynamic variable ordering heuristic. In this section, we generalize AgileABT
to get a new framework AgileABT(α), in which α represents any measure
used to implement a DVO. The original AgileABT [1] is then equivalent to
AgileABT(dom).
Due to space constraints, we only present in Algorithm 2 the pseudo-code of
AgileABT(α) related to the cases where an agent (Ai ) may change its current
order λi (i.e., calling procedure changeOrder) where ≺tv is an ordering on the
tv
termination values and = represents the equality. Ai can change its current
order λi and its associated termination value τi (procedure changeOrder, line 6)
in two cases. The first case is when Ai receives a stronger order λj associated
with the termination value τj from another agent Aj (processOrder, line 1 The
second case occurs when Ai itself proposes a new order (λi ) associated with a
termination value τi that is preferred (w.r.t. ≺tv ) to the termination value τi
associated to its current order λi (procedure proposeOrder, line 3 and 4).
The soundness, completeness and polynomial space complexity2
of AgileABT(α) are directly inherited from original AgileABT, i.e.,

2
If the computation of the measure α also has polynomial space complexity.
470 M. Wahbi et al.

AgileABT(dom). The only property that could be jeopardized is the termination

of the algorithm. In the following we define a sufficient condition on termination
values and the ordering ≺tv which guarantees that AgileABT(α) (Algorithm 2)
terminates. Next, we will discuss a condition on the measure α that allows the
termination values to obey the required condition.

Condition 1 The priority ordering ≺tv is a well-ordering on the range of func-

tion T V α .

Proposition 1. AgileABT(α) terminates if ≺tv obeys condition 1.

Proof. Following the pseudo-code of AgileABT(α), an agent Ai can only change

its current order in two cases (line 1 and 3). The termination values can only
decrease w.r.t. the well-ordering ≺tv , or remain the same and have a lexico-
graphically decreasing agent order (line 1). The agent order cannot decrease
lexicographically indefinitely and by condition 1 the termination values can-
not decrease indefinitely w.r.t. ≺tv . Therefore, AgileABT(α) cannot change the
order indefinitely. Once the order stops changing, all agents will eventually have
the same termination value and the same order to which it is attached (line 5,
Algorithm 2). This order corresponds to the strongest order computed in the
system so far. Since the agent order is now common and static, AgileABT(α)
will behave exactly like ABT, which terminates. 


To guarantee that AgileABT(α) terminates we need to define a well-ordering

≺tv on the termination values. Let α be the measure applied to the agents, and
let the function T V α be as defined in Algorithm 1. Let S be the range of α and
let ≺α be a total preference order on S.

Definition 1. Let λi and λj be two total agent orderings, τi = T V α (λi ) and

τj = T V α (λj ). The termination value τi is preferred (w.r.t. ≺tv ) to τj (i.e.,
τi ≺tv τj ) if and only if τi is lexicographically less than τj (w.r.t. ≺α ). In other
words, τi ≺tv τj iff ∃ k ∈ 1..n such that α(λi [k]) ≺α α(λj [k]) and ∀p ∈ 1..k−1
α(λi [p])=α(λj [p]).

Condition 2 The priority ordering ≺α is a well-ordering on the range of

measure α.

Proposition 2. AgileABT(α) terminates if ≺α obeys condition 2.

Proof. Suppose condition 2 is satisfied but AgileABT(α) does not terminate.

Then, by proposition 1, condition 1 is not satisfied. Therefore, ≺tv is not a
well-ordering on the range of T V α . In other words, we can obtain an infinite
decreasing sequence of termination values using a lexicographic comparison w.r.t.
≺α . Therefore we must have an infinite decreasing sequence of α(λ[k]) values,
for some k ∈ 1..n. But this contradicts condition 2 that forces ≺α to be a well-
ordering on S. Therefore AgileABT(α) must terminate. 

 A General Framework for Reordering Agents 471

Algorithm 3. Compute the wdeg of agent Ai (wdeg(i)).

// Filtering xj ∈ X by propagating C(X)
function revise(C(X), xj )
1. foreach ( vj ∈ Dj ) do
2. if ( ¬hasSupport(C(X), vj , xj ) ) then Dj ← Dj \ vj ;
3. if ( Dj = ∅ ) then weight[C] ← weight[C] + 1 ;
4. return Dj
= ∅;

procedure computeWeight()
5. wdeg ← 1;
6. foreach ( C(X) ∈ Ci | nbU nassigned(X) > 1 ) do wdeg ← wdeg + weight[C];
7. wdeg(i) ← min(wdeg, W ) ;

In the following, we consider a number of different heuristics that are known

to be effective in reducing search in centralized CSP, but which could not before
now be applied to distributed CSP. We show how the measures that inform
these heuristics can obey condition 2, and thus can be applied in the general
AgileABT(α) framework.

3.1 Neighborhood Based Variable Ordering Heuristics

In the first category we try to take into account the neighborhood

of each agent. We implemented three DVOs (dom/deg, dom/f deg and
dom/pdeg) to obtain respectively AgileABT(dom/deg), AgileABT(dom/f deg),
and AgileABT(dom/pdeg). In AgileABT(dom/deg), each agent Ai only requires
to know the degree deg(k) of each agent Ak in the problem. deg(k) (i.e.,
the number of neighbors of Ak ) can be obtained before the search starts as
is the case for d0k of each agent. Afterwards, Ai computes τi = T V dom/deg (λi )
using α(k) = dom(k)
deg(k) (Algorithm 1, line 2). In AgileABT(dom/f deg) and
AgileABT(dom/pdeg) each agent Ai is required to know the set of neighbors
of each agent Ak because it will need to compute the incoming degree pdeg(k)
and the outgoing degree f deg(k) of Ak for any proposed order. Again this
information can be known in a preprocessing step before running AgileABT(α).
Afterwards, Ai computes τi from T V dom/f deg (λi ) (resp. T V dom/pdeg (λi )) using
α(k) = fdom(k) dom(k)
deg(k) (resp. α(k) = pdeg(k) ) where the incoming degree pdeg(k) in λi


is the number of neighbors of Ak that appear before k in λi and the outgoing
degree f deg(k) in λi is the number of neighbors of Ak that appear after k in λi .

3.2 Conflict-Directed Variable Ordering Heuristic

The second category covers the conflict-directed variable ordering heuristic:

dom(k)
dom/wdeg. In order to compute τi using α(k) = wdeg(k) , each agent Ai in
AgileABT(dom/wdeg) requires to know the weighted degree wdeg(k) of each
other agent Ak . Ai maintains its weighted degree, wdeg(i), that it computes
472 M. Wahbi et al.

and the weighted degrees received from other agents, wdeg(k). In order to
compute the weighted degree, wdeg, (Algorithm 3) each agent Ai maintains a
counter weight[C] for each constraint C(X) in Ci . Whenever a domain (Dj where
xj ∈ X) is wiped-out while propagating C(X) (line 3, Algorithm 3), weight[C] is
incremented. Before assigning its variable and sending an ok? message to lower
priority agents, Ai computes its weighted degree, wdeg(i), by summing up (line
6, Algorithm 3) the weights of all constraints in Ci having at least two unassigned
variables ([14]). However, to guarantee that AgileABT(dom/wdeg) terminates
we only update wdeg(i) if the new computed weighted degree (wdeg) does not
exceed a limit W on which all agents agree beforehand (line 7, Algorithm 3). In
AgileABT(dom/wdeg), whenever Ai sends an ok? message it attaches to this
message the largest weighted degree computed so far wdeg(i).

3.3 Theoretical Analysis

Proposition 3. All measures α above are a well-ordering on a subset of Q
w.r.t. <.
Proof. We proceed by contradiction. Suppose there is an infinite decreasing
sequence of values of α(k). In all measures above, α(k) = dom(k)
ω(k) , for some ω(k).
dom(k) is the expected domain size of the agent Ak . It is obvious that dom(k)
is a well-ordering on N w.r.t. <, and so cannot decrease indefinitely. Therefore,
ω(k ) must increase indefinitely. ω(k ) is a positive integer whose value depends
on the measure used. Two cases were explored in this paper. The first case con-
cerns the family of degree-based heuristics (deg, pdeg, f deg). In this case, all of
the ω(k ) are greater than or equal to 1 and smaller than the number of agents
in the system (i.e., n) because an agent is at most constrained to n − 1 other
agents. Thus, 1 ≤ ω(k ) ≤ n − 1. The second case is related to the heuristic
wdeg. We have outlined in section 3.2 that an agent is not allowed to incre-
ment its weight when it has reached the limit W set beforehand (Algorithm 3,
line 7). Thus, 1 ≤ ω(k ) ≤ W . In both cases ω(k ) cannot increase indefinitely.
Therefore for all measures presented above, of the form α(k) = dom(k)
ω(k) , cannot
decrease indefinitely, and so α(k) is a well-ordering w.r.t. <. 


4 Empirical Analysis
In this section we experimentally compare AgileABT(α)3 using different DVO
heuristics to three other algorithms: ABT, ABT DO with nogood-triggered
heuristic (ABT DO-ng) [29] and ABT DO with min-domain retroactive heuris-
tic (ABT DO Retro(mindom)) [31]. All experiments were performed on the
DisChoco 2.0 platform [26],4 in which agents are simulated by Java threads
that communicate only through message passing.
3
For AgileABT(dom/wdeg), we fixed W = 1, 000. But, varying W made negligible
difference to the results.
4
http://dischoco.sourceforge.net/
 A General Framework for Reordering Agents 473

#ncccs 〈n=20, d=10, p1=0.2〉

4.0 ⋅10 5
ABT 3e+4
ABT_DO-ng
5
3.0 ⋅10 ABT_DO-Retro(mindom) 2e+4

AgileABT(dom)
1e+4
2.0 ⋅10
5 AgileABT(dom/deg)
AgileABT(dom/pdeg) 0.65
AgileABT(dom/fdeg)
1.0 ⋅10 5
AgileABT(dom/wdeg)

0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8

p2

Fig. 2. Sparse uniform binary random DisCSPs

When comparing distributed algorithms, the performance is evaluated using

two common metrics: the communication load and computation effort. Commu-
nication load is measured by the total number of exchanged messages among
agents during algorithm execution (#msg) [15]. Computation effort is measured
by the number of non-concurrent constraint checks (#ncccs) [28]. #ncccs is
the metric used in distributed constraint solving to simulate computation time,
but for dynamic reordering algorithms its variant generic #ncccs is used[30].
Algorithms are evaluated on three benchmarks: uniform binary random DisC-
SPs, distributed graph coloring problems, and composed random instances. All
binary table constraints in these problems are implemented using AC-2001 [4].

4.1 Uniform Binary Random DisCSPs

Uniform binary random DisCSPs are characterized by n, d, p1 , p2 , where n is
the number of agents/variables, d is the number of values in each domain, p1
is the network connectivity defined as the ratio of existing binary constraints
to possible binary constraints, and p2 is the constraint tightness defined as the
474 M. Wahbi et al.

#ncccs 〈n=20, d=10, p1=0.7〉

8.0 ⋅10 6
2.0e+6
ABT
ABT_DO-ng
6.0 ⋅10 6 1.5e+6 ABT_DO-Retro(mindom)
1.0e+6
AgileABT(dom)
AgileABT(dom/deg)
4.0 ⋅10 6
0.3
AgileABT(dom/pdeg)
AgileABT(dom/fdeg)
6
2.0 ⋅10 AgileABT(dom/wdeg)

0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55

p2

Fig. 3. Dense uniform binary random DisCSPs

ratio of forbidden value pairs to all possible pairs. We solved instances of two
classes of random DisCSPs: sparse problems 20, 10, 0.2, p2  and dense problems
20, 10, 0.7, p2 . We varied the tightness from 0.1 to 0.9 by steps of 0.05. For each
pair of fixed density and tightness (p1 , p2 ), we generated 20 instances, solved 5
times each. We report average over the 100 execution.
Figures 2 and 3 show the results on sparse respectively dense uniform binary
random DisCSPs. In sparse problems (Figure 2), AgileABT(α) outperforms all
other algorithms on both #msg and #ncccs. ABT with AC-2001 is signifi-
cantly the slowest algorithm but it requires fewer messages than ABT DO-
ng. Regarding the speedup, AgileABT(α) shows almost an order of magni-
tude improvement compared to ABT and they all improve on ABT DO-ng and
ABT DO Retro(mindom). Comparing AgileABT(α) algorithms, neighborhood
based heuristics (i.e., deg, pdeg and f deg) show an almost two-fold improvement
over dom and wdeg on #ncccs. This improvement is even more significant on
#msg. wdeg requires fewer messages than dom. In dense problems (Figure 3),
neighborhood based heuristics outperform all other algorithms both on #msg
and #ncccs. ABT requires almost half the #msg of ABT DO Retro(mindom),
 A General Framework for Reordering Agents 475

AgileABT(dom), and AgileABT(dom/wdeg). ABT DO-ng is always the algo-

rithm that requires more messages. AgileABT(dom/wdeg) shows poor perfor-
mance. It is slower and requires more messages than AgileABT(dom).

4.2 Distributed Graph Coloring Problems

Distributed graph coloring problems are characterized by n, d, p1 , where n, d

and p1 are as above and all constraints are binary difference constraints. We
report the average on 100 instances of two classes n = 15, d = 5, p1 = 0.65 and
n = 25, d = 5, p1 = 0.45 in Table 1. Again, AgileABT(α) using neighborhood
based DVO are by far the best algorithms for solving both classes. ABT DO-ng
shows poor performance on solving those problems. ABT DO Retro(mindom)
outperforms AgileABT(dom) in both classes. Comparing AgileABT(dom) to

Table 1. Distributed graph coloring problems

15, 5, 0.65 25, 5, 0.45

Algorithm
#msg #ncccs #msg #ncccs
AgileABT(dom/wdeg) 90,630 188,991 2,600,016 2,783,132
AgileABT(dom/f deg) 51,820 104,517 940,481 937,861
AgileABT(dom/pdeg) 47,949 89,514 454,998 434,540
AgileABT(dom/deg) 44,083 78,050 607,927 505,140
AgileABT(dom) 79,518 204,012 3,001,538 3,836,301
ABT DO Retro(mindom) 73,278 115,850 1,089,024 830,423
ABT DO-ng 157,873 282,737 4,547,565 3,639,791
ABT 58,817 288,803 1,626,901 3,836,391

Table 2. Composed random instances

Instances 25-1-25 25-1-40

Algorithm #msg #ncccs #msg #ncccs
AgileABT(dom/wdeg) 85,521 30,064 89,804 33,461
AgileABT(dom/f deg) 146,668 219,980 1,337,552 2,830,906
AgileABT(dom/pdeg) 57,079 16,043 54,667 17,704
AgileABT(dom/deg) 122,735 309,064 740,669 2,793,670
AgileABT(dom) 57,451 20,944 59,859 23,405
ABT DO Retro(mindom) 67,022 41,401 96,783 59,980
ABT DO-ng 1,329,257 1,614,960 > 108 > 109
ABT 2,850,137 22,042,094 9,429,088 72,524,742
476 M. Wahbi et al.

AgileABT(dom/wdeg), dom is slower than wdeg but it requires fewer messages.

In n = 15, d = 5, p1 = 0.65, only AgileABT(α) using neighborhood based DVO
outperform ABT on messages while other asynchronous dynamic ordering algo-
rithms require more messages.

4.3 Composed Random Instances

We also evaluate all algorithms on two sets of unsatisfiable composed random
instances used to evaluate the conflict-directed variable ordering heuristic in cen-
tralized CSP [7,20].5 Each set contains 10 different instances where each instance
is composed of a main (under-constrained) fragment and some auxiliary frag-
ments, each of which being grafted to the main one by introducing some binary
constraints. Each instance contains 33 variables and 10 values per variable, and
as before, each variable is controlled by a different agent. We solved each instance
5 times and present the average over 50 executions in Table 2. The results show
that AgileABT(dom/pdeg) outperforms all other algorithms in both classes. The
second best algorithm for solving these instances is AgileABT(dom). ABT shows
very poor performance on solving these problems followed by ABT DO-ng that
cannot solve instances in the second class (25-1-40) within the limits we fixed
for all algorithms (108 #msg and 109 #ncccs). Regarding AgileABT(α) DVOs,
wdeg seems to pay off on these instances compared to dom/deg and dom/f deg.
In 25-1-40, AgileABT(dom/deg) outperforms ABT DO Retro(mindom), but the
opposite happens for 25-1-25.

4.4 Discussion

Looking at all results together, we come to the straightforward conclusion that

AgileABT(α) with neighbourhood-based heuristics, namely deg, f deg and pdeg
perform very well compared to other techniques. We think that neighbourhood-
based heuristics perform well thanks to their ability to take into account the
structure of the problem [3]. Distinctly, among these three heuristics dom/pdeg
seems to be the best one because of the limited changes on the agent at the
first position. In dom/pdeg τ[1] = dom(λ[1]) because pdeg(λ[1]) = 1. Thus, the
number of order changes (cost on messages) in AgileABT(dom/pdeg) is reduced.
Note that the strength of AgileABT(α) is that it enables any ordering to be
identified and executed as the algorithm runs. However, each change invokes a
series of coordination messages, and so too many changes of order will have a
negative impact.
On the other hand the AgileABT(α) with the conflict-directed variable order-
ing heuristic, namely wdeg, shows a relatively poor performance. This fact can
be explained by the limited amount of constraint propagation performed by
DisCSP algorithms. Furthermore, asynchrony affects reception and treatment of
ok? and ngd messages and has a direct impact on the computation of weights
and new orders. For some instances of the coloring problem, the performance of
5
http://www.cril.univ-artois.fr/∼lecoutre/benchmarks.html
 A General Framework for Reordering Agents 477

the conflict-directed heuristic varies significantly from one execution to another,

indicating it is more sensitive to the asynchrony than the other heuristics.

5 Related Work
Bliek ([5]) proposed Generalized Partial Order Dynamic Backtracking
(GPODB), an algorithm that generalizes both Partial Order Dynamic Back-
tracking (PODB) [9] and Dynamic Backtracking (DBT) [8]. This generaliza-
tion was obtained by relaxing the safety conditions of PODB which results in
additional flexibility. AgileABT has some similarities with PODB and GPODB
because AgileABT also maintains a set of safety conditions. However, both
PODB and GPODB are subject to the same restriction: when a dead end occurs,
the backtracking target of the generated no-good must be selected such that the
safety conditions induced by the new no-good satisfy all existing safety condi-
tions. By contrast, AgileABT overcomes this restriction by allowing the violation
of existing safety conditions by relaxing some explanations.
Silaghi et al. proposed Asynchronous Backtracking with Reordering (ABTR)
[21–23]. In ABTR, a reordering heuristic identical to the centralized dynamic
backtracking [8] was applied to ABT. The authors in [29] proposed Dynamic
Ordering for Asynchronous Backtracking (ABT DO) [29]. When an ABT DO
agent assigns a value to its variable, it can reorder lower priority agents. In
other words, an agent on a position j can change order of the agents on posi-
tions j + 1 through n. The authors in [29] proposed three different ordering
heuristics to reorder lower priority agents: random, min-domain and nogood-
triggered inspired by dynamic backtracking [8]. Their experimental results show
that nogood-triggered (ABT DO-ng), where the generator of the no-good is
placed just after the target of the backtrack, is best.
A new kind of ordering heuristics for ABT DO is presented in [31] for reorder-
ing higher agents. These retroactive heuristics enable the generator of a no-good
(backtrack) to be moved to a higher position than that of the target of the back-
track. The resulting algorithm is called ABT DO Retro. The degree of flexibility
of these heuristics depends on the size of the no-good storage capacity K, which
is predefined. Agents are limited to store no-goods with a size equal to or smaller
than K. The space complexity of the agents is thus exponential in K. However,
the best of those heuristics, ABT DO Retro(mindom) [17,31], does not need
any no-good storage. In ABT DO Retro(mindom), the agent that generates a
no-good is placed between the last and the second last agents in the no-good if
its domain size is smaller than that of the agents it passes on the way up.

6 Conclusion
We proposed a general framework for reordering agents asynchronously in DisC-
SPs which can implement many different dynamic variable ordering heuristics.
Our general framework is sound, complete, and has a polynomial space complex-
ity. We proved that a simple condition on the measure used in the heuristics is
478 M. Wahbi et al.

enough to guarantee termination. We implemented several DVOs heuristics that

are well-known in centralized CSP but were not able to be used before in the
distributed CSP. Our empirical study shows the significance of these DVOs on
a distributed setting. In particular, it highlights the good performance of neigh-
borhood based heuristics. Future work will focus on designing new DVOs that
can be used in AgileABT(α).

Acknowledgments. This work is partly funded by Science Foundation Ireland (SFI)

under Grant Number SFI/12/RC/2289.

References
1. Bessiere, C., Bouyakhf, E.H., Mechqrane, Y., Wahbi, M.: Agile asynchronous back-
tracking for distributed constraint satisfaction problems. In: Proceedings of ICTAI
2011, Boca Raton, Florida, USA, pp. 777–784, November 2011
2. Bessiere, C., Maestre, A., Brito, I., Meseguer, P.: Asynchronous backtracking with-
out adding links: a new member in the ABT family. Artif. Intel. 161, 7–24 (2005)
3. Bessiere, C., Régin, J.C.: MAC and combined heuristics: two reasons toforsake FC
(and CBJ?) on hard problems. In: Freuder, E.C. (ed.) CP 1996. LNCS, vol. 1118,
pp. 61–75. Springer, Heidelberg (1996)
4. Bessiere, C., Régin, J.C.: Refining the basic constraint propagation algorithm. In:
Proceedings of IJCAI 2001, San Francisco, CA, USA, pp. 309–315 (2001)
5. Bliek, C.: Generalizing partial order and dynamic backtracking. In: Proceedings of
AAAI 1998/IAAI 1998, Menlo Park, CA, USA, pp. 319–325 (1998)
6. Bonnet-Torrés, O., Tessier, C.: Multiply-constrained dcop for distributed plan-
ning and scheduling. In: AAAI SSS: Distributed Plan and Schedule Management,
pp. 17–24 (2006)
7. Boussemart, F., Hemery, F., Lecoutre, C., Sais, L.: Boosting systematic search by
weighting constraints. In: Proceedings of ECAI 2004, pp. 146–150 (2004)
8. Ginsberg, M.L.: Dynamic Backtracking. JAIR 1, 25–46 (1993)
9. Ginsberg, M.L., McAllester, D.A.: GSAT and dynamic backtracking. In: KR,
pp. 226–237 (1994)
10. Haralick, R.M., Elliott, G.L.: Increasing tree search efficiency for constraint satis-
faction problems. Artif. Intel. 14(3), 263–313 (1980)
11. Jung, H., Tambe, M., Kulkarni, S.: Argumentation as distributed constraint sat-
isfaction: applications and results. In: Proceedings of AGENTS 2001, pp. 324–331
(2001)
12. Junges, R., Bazzan, A.L.C.: Evaluating the performance of dcop algorithms in
a real world, dynamic problem. In: Proceedings of AAMAS 2008, Richland, SC,
pp. 599–606 (2008)
13. Léauté, T., Faltings, B.: Coordinating logistics operations with privacy guarantees.
In: Proceedings of the IJCAI 2011, pp. 2482–2487 (2011)
14. Lecoutre, C., Boussemart, F., Hemery, F.: Backjump-based techniques versus
conflict-directed heuristics. In: Proceedings of IEEE ICTAI 2004, pp. 549–557
(2004)
15. Lynch, N.A.: Distributed Algorithms. Morgan Kaufmann Series (1997)
16. Maheswaran, R.T., Tambe, M., Bowring, E., Pearce, J.P., Varakantham, P.: Taking
DCOP to the real world: efficient complete solutions for distributed multi-event
scheduling. In: Proceedings of AAMAS 2004 (2004)
 A General Framework for Reordering Agents 479

17. Mechqrane, Y., Wahbi, M., Bessiere, C., Bouyakhf, E.H., Meisels, A., Zivan, R.:
Corrigendum to “Min-Domain Retroactive Ordering for Asynchronous Backtrack-
ing”. Constraints 17, 348–355 (2012)
18. Ottens, B., Faltings, B.: Coordination agent plans trough distributed constraint
optimization. In: Proceedings of MASPLAN 2008, Sydney, Australia (2008)
19. Petcu, A., Faltings, B.V.: A value ordering heuristic for local search in distributed
resource allocation. In: Faltings, B.V., Petcu, A., Fages, F., Rossi, F. (eds.) CSCLP
2004. LNCS (LNAI), vol. 3419, pp. 86–97. Springer, Heidelberg (2005)
20. Roussel, O., Lecoutre, C.: Xml representation of constraint networks: Format
XCSP 2.1. CoRR (2009)
21. Silaghi, M.C.: Framework for modeling reordering heuristics for asynchronous back-
tracking. In: IEEE/WIC/ACM International Conference on Intelligent Agent Tech-
nology, IAT 2006, pp. 529–536, December 2006
22. Silaghi, M.C.: Generalized dynamic ordering for asynchronous backtracking on
DisCSPs. In: Proceedings of DCR 2006 (2006)
23. Silaghi, M.C., Sam-Haroud, D., Faltings, B.: ABT with Asynchronous Reordering.
In: 2nd Asia-Pacific IAT (2001)
24. Silaghi, M.C., Sam-Haroud, D., Calisti, M., Faltings, B.: Generalized english auc-
tions by relaxation in dynamic distributed CSPs with private constraints. In:
Proceedings of DCR 2001, pp. 45–54 (2001)
25. Wahbi, M.: Algorithms and Ordering Heuristics for Distributed Constraint Satis-
faction Problems. John Wiley & Sons, Inc. (2013)
26. Wahbi, M., Ezzahir, R., Bessiere, C., Bouyakhf, E.H.: DisChoco 2: a platform
for distributed constraint reasoning. In: Proceedings of workshop on DCR 2011,
pp. 112–121 (2011). http://dischoco.sourceforge.net/
27. Yokoo, M., Durfee, E.H., Ishida, T., Kuwabara, K.: Distributed constraint sat-
isfaction for formalizing distributed problem solving. In: Proceedings of ICDCS,
pp. 614–621 (1992)
28. Zivan, R., Meisels, A.: Parallel Backtrack search on DisCSPs. In: Proceedings of
DCR 2002 (2002)
29. Zivan, R., Meisels, A.: Dynamic Ordering for Asynchronous Backtracking on DisC-
SPs. Constraints 11(2–3), 179–197 (2006)
30. Zivan, R., Meisels, A.: Message delay and DisCSP search algorithms. AMAI 46(4),
415–439 (2006)
31. Zivan, R., Zazone, M., Meisels, A.: Min-Domain Retroactive Ordering for Asyn-
chronous Backtracking. Constraints 14(2), 177–198 (2009)
 Automatically Generating Streamlined
Constraint Models with Essence and Conjure

James Wetter(B) , Özgür Akgün, and Ian Miguel

School of Computer Science, University of St Andrews, St Andrews, UK

{jpw3,ozgur.akgun,ijm}@st-andrews.ac.uk

Abstract. Streamlined constraint reasoning is the addition of unin-

ferred constraints to a constraint model to reduce the search space, while
retaining at least one solution. Previously, effective streamlined models
have been constructed by hand, requiring an expert to examine closely
solutions to small instances of a problem class and identify regularities.
We present a system that automatically generates many conjectured reg-
ularities for a given Essence specification of a problem class by examin-
ing the domains of decision variables present in the problem specification.
These conjectures are evaluated independently and in conjunction with
one another on a set of instances from the specified class via an auto-
mated modelling tool-chain comprising of Conjure, Savile Row and
Minion. Once the system has identified effective conjectures they are
used to generate streamlined models that allow instances of much larger
scale to be solved. Our results demonstrate good models can be identified
for problems in combinatorial design, Ramsey theory, graph theory and
group theory - often resulting in order of magnitude speed-ups.

1 Introduction
The search space defined by a constraint satisfaction problem can be vast, which
can lead to impractically long search times as the size of the problem instance
considered grows. An approach to mitigating this problem is to narrow the focus
of the search onto promising areas of the search space using streamliners [11].
Streamliners take the form of uninferred additional constraints (i.e. constraints
not proven to follow from the original problem statement) that rule out a sub-
stantial portion of the search space. If a solution lies in the remainder then it can
typically be found more easily than when searching the full space. Previously,
streamlined models have been produced by hand [11,13,15,16], which is both
difficult and time-consuming. The principal contribution of this paper is to show
how a powerful range of streamliners can be generated automatically.
Our approach is situated in the automated constraint modelling system Con-
jure [2]. This system takes as input a specification in the abstract constraint
specification language Essence [7,8]. Figure 1 presents an example specifica-
tion, which asks us to partition the integers 1 . . . n into k parts subject to a set
of constraints. Essence supports a powerful set of type constructors, such as set,
multi set, function and relation, hence Essence specifications are concise and

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 480–496, 2015.
DOI: 10.1007/978-3-319-23219-5 34
 Automatically Generating Streamlined Constraint Models 481

language Essence 1.3

given k , l , n : int (1..)

find p : partition ( numParts k ) from int (1.. n )

such that
forAll s in parts ( p ) .
forAll start : int (1.. n - l +1) .
forAll width : int (1..( n - start +1)/( l -1)) .
!( forAll i : int (0.. l -1) .
( start + i * width ) in s )

Fig. 1. Essence specification of the Van Der Waerden Number Problem [24] (see
Appendix A). The specification describes a certificate for the lower bound on W (k, l),
and will be unsatisfiable if n is equal to W (k, l).

highly structured. Existing constraint solvers do not support these abstract deci-
sion variables directly. Therefore we use Conjure to refine abstract constraint
specifications into concrete constraint models, using constrained collections of
primitive variables (e.g. integer, Boolean) to represent the abstract structure.
Our method exploits the structure in an Essence specification to produce
streamlined models automatically, for example by imposing streamlining con-
straints on or between the parts of the partition in the specification in Figure 1.
The modified specification is refined automatically into a streamlined constraint
model by Conjure. Identifying and adding the streamlining constraints at this
level of abstraction is considerably easier than working directly with the con-
straint model, which would involve first recognising (for example) that a certain
collection of primitive variables and constraints together represent a partition
— a potentially very costly step.
Our system contains a set of rules that fire when their preconditions match a
given Essence specification to produce candidate streamliners. Using Conjure
to refine the streamlined specifications into constraint models, solved with Sav-
ile Row1 [19] and Minion2 [10], candidates are evaluated against instances of
the specified class. Effective streamliners are combined to produce more powerful
candidates. As we will show, high quality streamlined models can be produced
in this way, in some cases resulting in a substantial reduction in search effort.
Herein we focus on satisfaction problems. Optimisation is an important future
consideration but requires different treatment: streamliners may allow us to find
good solutions quickly but exclude the optimal solution to the original model.
The rest of this paper is structured as follows. Following a summary of related
work, we give some background on the Essence language. Section 4 describes
in detail our approach to generating streamliners automatically, then Section 5
discusses combining streamliners to produce yet more effective models. We con-
clude following a discussion of discovering streamliners in practice.

1
http://savilerow.cs.st-andrews.ac.uk
2
http://constraintmodelling.org/minion/
482 J. Wetter et al.

2 Related Work
Colton and Miguel [5] and Charnley et al [4] used Colton’s HR system [6] to con-
jecture the presence of implied constraints from the solutions to small instances
of several problem classes, including quasigroups and moufang loops. The Otter
theorem prover3 was used to prove the soundness of these conjectures. If proven
sound, the implied constraints were added to the model to aid in the solution of
larger instances of the same problem class.
Streamlined constraint reasoning differs from the approach of Charnley et
al in that the conjectured constraints are not typically proven to follow from a
given constraint model. Rather, they are added in the hope that they narrow
the focus of the search onto an area containing solutions. When first introduced
by Gomes and Sellmann [11] streamlined constraint reasoning was used to help
construct diagonally ordered magic squares and spatially balanced experiment
designs. For the magic squares the additional structure enforced was regularity
in the distribution of numbers in the square. That is, the small numbers are
not all similarly located, and likewise for large numbers. The spatially balanced
experiment designs were forced to be self-symmetric: the permutations repre-
sented by the rows of the square must commute with each other. For both of
these problems streamlining led to huge improvements in solve time, allowing
much larger instances to be solved.
Kouril et al. refer to streamlining as “tunneling” [13]. They describe the addi-
tional constraints as tunnels that allow a SAT solver to tunnel under the great
width seen early in the search tree when computing bounds on Van de Waerden
numbers. They used simple constraints that force or disallow certain sequences
of values to occur in the solutions. Again this led to a dramatic improvement in
run time of the solver, allowing much tighter bounds to be computed.
Le Bras et al. used streamlining to help construct graceful double wheel
graphs [15]. Constraints forcing certain parts of the colouring to form arithmetic
sequences allowed for the construction of colourings for much larger graphs.
These constraints led to the discovery of a polynomial time construction for
such colourings, proving that all double wheel graphs are graceful.
Finally Le Bras et. al. made use of streamlining constraints to compute new
bounds on the Erdős discrepancy problem [16]. Here constraints enforcing peri-
odicity in the solution, the occurrence of the improved Walters sequence, and
a partially multiplicative property improved solver performance, allowing the
discovery of new bounds.
In all of these examples streamliners proved very valuable, but were generated
by hand following significant effort by human experts. In what follows, we will
show that the structure recognised and exploited by these experts is often present
in abstract constraint specifications.

3
http://www.cs.unm.edu/∼mccune/otter/
 Automatically Generating Streamlined Constraint Models 483

language Essence 1.3

given v , b , r , k , lambda : int (1..)

where v = b , r = k
letting Obj be new type of size v ,
Block be new type of size b

find bibd : relation ( symmetric ) of ( Obj * Block )

such that
forAll o : Obj . | bibd (o , _ )| = r ,
forAll bl : Block . | bibd (_ , bl )| = k ,
forAll o1 , o2 : Obj .
o1 != o2 -> | bibd ( o1 , _ ) intersect bibd ( o2 , _ )| = lambda

Fig. 2. Essence specification of the square (ensured by the where statement) Balanced
Incomplete Block Design Problem [20]. Streamliner added as an Essence annotation
shown underlined.

3 Background: Essence
The motivation for abstract constraint specification languages, such as Zinc [18]
and Essence is to address the modelling bottleneck: the difficulty of formulating
a problem of interest as a constraint model suitable for input to a constraint
solver. An abstract constraint specification describes a problem above the level
at which constraint modelling decisions are made. An automated refinement
system, such as Conjure, can then be used to produce a constraint model from
the specification automatically.
An Essence specification, such as those given in Figures 1 and 2, identifies:
the input parameters of the problem class (given), whose values define a prob-
lem instance; the combinatorial objects to be found (find); and the constraints
the objects must satisfy (such that). In addition, an objective function may
be specified (min/maximising — not shown in these examples) and identifiers
declared (letting). Abstract constraint specifications must be refined into con-
crete constraint models for existing constraint solvers. Our Conjure system
employs refinement rules to convert an Essence specification into the solver-
independent constraint modelling language Essence . From Essence we use
Savile Row to translate the model into input for a particular solver while
performing solver-specific model optimisations.
A key feature of abstract constraint specification languages is the support for
abstract decision variables with types such as set, multiset, relation and function,
as well as nested types, such as set of sets and multiset of relations. This allows
the problem structure to be captured very concisely. As explained below, this
clarity of structure is a good basis for the conjecture of streamlining constraints.

4 From Conjectures to Streamlined Specifications

This section presents the methods used to generate streamlined models automat-
ically. The process is driven by the decision variables in an Essence specification,
484 J. Wetter et al.

such as the partition in Figure 1. For each variable, the system forms conjec-
tures of possible regularities that impose additional restrictions on the values of
that variable’s domain. Since the domains of Essence decision variables have
complex, nested types, these restrictions can have far-reaching consequences for
constraint models refined from the modified specification. The intention is that
the search space is reduced considerably, while retaining at least one solution.
Multiple conjectures found to produce successful streamlined specifications indi-
vidually can be combined in an attempt to produce a single more sophisticated
streamliner. Currently conjectures are formed about each variable independently;
an important future direction is to make conjectures across multiple variables.

4.1 Exploiting Essence Domain Annotations

Essence allows domains to be annotated to restrict the set of acceptable values.
For example, a function variable domain may be restricted to injective functions,
or a partition variable domain may be restricted to regular partitions. Hence, the
simplest source of streamliners is the systematic annotation of the decision vari-
ables in an input specification. This sometimes retains solutions to the original
problem while improving solver performance. Consider the Balanced Incomplete
Block Design problem (BIBD, Figure 2), where the decision variable is a rela-
tion. For square BIBDs we might consider a streamliner requiring a symmetric
relation, achieved simply by adding the symmetric annotation as shown.
Figure 3 summarises an experiment with this streamliner on a set of satisfi-
able square BIBD instances. Original and streamlined specifications were refined
with Conjure, using the Compact heuristic [1] to select one model. For each
instance Savile Row was used to prepare the resulting model for Minion,
which was used to find a solution. Streamlining uniformly resulted in a solvable
instance, and as seen in Figure 3a in all but one instance search size is equiv-
alent or reduced and as seen in Figure 3b the corresponding execution time is
sometimes reduced.

1e+08 1000

1e+07
100
Symmetric Relation Search Nodes

Symmetric Relation Search Time

1e+06
10
100000

10000 1

1000
0.1
100
0.01
10

1 0.001
1 10 100 1000 10000 100000 1e+06 1e+07 1e+08 0.001 0.01 0.1 1 10 100 1000
Original Specification Search Nodes Original Specification Search Time

(a) Search nodes. (b) Search time.

Fig. 3. Search effort to find the first solution to satisfiable square BIBD instances
where v = b ≤ 40: original vs streamlined specification.
 Automatically Generating Streamlined Constraint Models 485

4.2 Conjecture-Forming Rules

The existing Essence domain annotations are, however, of limited value. They
are very strong restrictions and so often remove all solutions to the original
problem when added to a specification. In order to generate a larger variety of
useful conjectures we employ a small set of rules, categorised into two classes:
1. First-order rules add constraints to reduce the domain of a decision variable
directly.
2. Higher-order rules take other rules as arguments and use them to reduce the
domain of a decision variable.
The full list of rules is shown in Table 1. A selection of the first-order rules is
given in Figure 4, and a selection of the higher-order rules are given in Figure 5.
We define four first-order rules that operate on integer domains directly:
‘odd’, ‘even’, ‘lowerHalf’ and ‘upperHalf’. Each restricts an integer domain to
an appropriate subset of its values. We define six first-order rules for function
domains. Two of these constrain the function to be monotonically increasing (or
decreasing). The other four place the largest (or smallest) element first (or last).
Functions mapping a pair of objects to a third object of the same type can be
viewed as binary operators on these objects. We define three first-order rules
to enforce such functions to be commutative, non-commutative, and associa-
tive respectively. Finally, we define a first-order rule for partitions called ‘quasi-
regular’. Partition domains in Essence can have a regular annotation, however
this can be too strong. The ‘quasi-regular’ streamlining rule posts a soft version

Table 1. The rules used to generate conjectures. Rows with a softness parameter
specify a family of rules each member of which is defined by an integer parameter.
Class Trigger Domain Name Softness
odd{even} no
int
lower{upper}Half no
monotonicIncreasing{Decreasing} no
function int --> int
largest{smallest}First{Last} no
First-order
commutative no
function (X,X) --> X associative no
non-commutative no
partition from X quasi-regular yes
all no
most yes
set of X
half no
approxHalf yes
range no
Higher-order
defined no
function X --> Y
pre{post}fix yes
allBut yes
function (X,X) --> Y diagonal no
partition from X parts no
486 J. Wetter et al.

Name odd Name lowerHalf

Input X: int Input X: int(l..u)
Output X % 2 = 1 Output X < l + (u - l) / 2

Name monotonicIncreasing Name largestFirst

Input X: function int --> int Input X: function int(l..u) --> int
Output Output
forAll i in defined(X) . forAll i in defined(X) .
forAll j in defined(X) . X(min(defined(X)) <= X(i)
i < j -> X(i) <= X(j)

Name commutative Name quasi-regular

Input X: function (D, D) --> D Input X: partition from _
Output Output
forAll (i,j) in defined(X) . minPartSize(X,
X((i,j)) = X((j,i)) |participants(X)|/|parts(X)| - k)
/\
maxPartSize(X,
|participants(X)|/|parts(X)| + k)

Fig. 4. A selection of the first-order streamlining rules.

Name all Name most

Parameter R (another rule) Parameter R (another rule)
Input X: set of _ Parameter k (softness)
Output forAll i in X . R(i) Input X: set of _
Output
k >= sum i in X . toInt(!R(i))

Name range Name prefix

Parameter R (another rule) Parameter R (another rule)
Input X: function _ --> _ Parameter k (softness)
Output R(range(X)) Input X: function int(l..h) --> _
Output R(restrict(X, ‘int(l..h-k)‘))

Name parts Name diagonal

Parameter R (another rule) Parameter R (another rule)
Input X: partition of _ Input
Output R(parts(X)) X: function (D1, D1) --> D2
Output
{ R(X )
@ find X : function D1 --> D2
such that
forAll i : D1 .
(i,i) in defined(X) -> X (i) = X((i,i)),
forAll i : D1 .
i in defined(X ) -> X (i) = X((i,i))
}

Fig. 5. A selection of the higher-order streamlining rules.

 Automatically Generating Streamlined Constraint Models 487

of the regularity constraint, which takes an integer parameter, k, to control the

softness of the constraint. In our experiments we varied the value of k between 1
and 3. Larger values of k will make the constraint softer as they allow the sizes
of parts in the partition to be k-apart.
Higher-order rules take another rule as an argument and lift its operation
to a decision variable with a nested domain. For example, the ‘all’ rule for sets
applies a given streamlining rule to all members of a set, if it is applicable.
We define three other higher-order rules that operate on set variables: ‘half’,
‘most’ and ‘approxHalf’, the last two with softness parameters. For a set of
integers, applying the ‘half’ rule with the ‘even’ rule as the parameter – denoted
‘half(even)’ – forces half of the values in the set to be even. The parameter to the
higher-order rule can itself be a higher order rule, so for a set of set of integers
‘all(half(even))’ constrains half of the members of all inner sets to be even.
The ‘defined’ and ‘range’ rules for functions use the defined and range oper-
ators of Essence to extract the corresponding sets from the function variable.
Once the set is extracted the parameter rule R can be directly applied to it. The
‘prefix’ and ‘postfix’ rules work on functions that map from integers by focusing
on a prefix or postfix of the integer domain. The ‘parts’ rule views a partition
as a set of sets and opens up the possibility of applying set rules to a partition.
The ‘diagonal’ rule introduces an auxiliary function variable. The auxiliary
variable represents the diagonal of the original variable, and it is channelled into
the original variable. Once this variable is constructed the streamlining rule taken
as a parameter, R, can be applied directly to the auxiliary variable. Similarly the
‘allBut’ rule introduces an auxiliary function variable that represents the original
variable restricted to an arbitrary subset (of fixed size) of its domain. This is
similar to the ‘prefix’ rule but allows the ignored mappings of the function to
fall anywhere in the function’s domain rather than only at the end.
It is important to note that allowing higher-order rules to take other higher-
order rules as parameters naively can lead to infinite recursion; such as ‘pre-
fix(prefix(prefix(...)))’ or ‘prefix(postfix(prefix(..)))’. We considered two ways of
mitigating this problem: 1) using a hard-coded recursion limit 2) only allow-
ing one application of the same higher-order rule in a chain of rule applications
and at the same level. Using a hard-coded recursion limit has two drawbacks.
It still allows long lists of useless rule applications like applying ‘prefix’ repeat-
edly. It can also disallow useful but long sequences of rule applications. Instead,
we implemented a mechanism where the same higher-order rule can only be
applied once at the same level; that is, a rule can only be applied more than
once if the domain of its input is a subcomponent of the domain of the input
of the previous application. This disallows ‘prefix(prefix(...))’, but allows ‘pre-
fix(range(half(prefix(even),1)),1)’ which works on a decision with the following
type function int --> (function int --> int).
488 J. Wetter et al.

In order to apply these rules to Essence problem specifications, we extract

a list of domains from the top level decision variables. For each domain, we find
the list of all applicable rules. Each application of a streamlining rule results in
a conjecture that can be used to streamline the original specification.
To see an example of an effective constraint generated by this system consider
the problem of generating certificates for lower bounds on Van Der Waerden
numbers shown in Figure 1. Here only one decision variable is present, a partition
of integers, p. First the rule parts is triggered so p is treated as a set of set of
integers. Next, the rule all is triggered such that all the members of the outer
set are restricted, then another rule approxHalf is triggered so approximately
half the members of each inner set are restricted. Finally lowerHalf is triggered
so the domain of the integer has its upper bound halved. The complete rule
that is being applied is ‘parts(all(approxHalf(lowerHalf, i)))’, and the resulting
constraint is:

forAll s in parts(p) .
|s|/2 + i >= sum x in s . toInt(x <= n/2) /\
|s|/2 - i <= sum x in s . toInt(x <= n/2)

where i is the parameter given to approxHalf. This constraint enforces that

each part of the partition consists of approximately half ‘small’ numbers and
half ‘large’ numbers. Figure 8 shows that this constraint with i = 2 drastically
reduces the number of search nodes explored when finding a single solution to
the problem.
Initially these rules were applied to the variable domains declared as finds in
the original specification. It was observed that the wheel like structures in the
graceful graph labeling problems were not exposed in the function variables used
to represent the labeling. In order to extract such structures a preprocessing step
is performed that introduces restricted function variables to the rule system in
order to generate a wider range of streamlining constraints.
This preprocessing step is triggered when a function variable, f : A → B,
is present in the find declarations. For each such variable the constraints are
checked for an application of the function, f (a), quantified over a strict subset
of the domain, a ∈ A ⊂ A. The quantification can be , ∀ or ∃. If such an
expression is present in the constraints the conjecture generation rules are also
applied to a restricted version of the function, f |A .
This system is capable of generating novel streamliners that have not been
previously reported, such as the ‘quasi-regular’ rule. In addition it generates
some streamliners very similar to some of those previously seen in the literature.
For example appliying ‘restrict(range(all(odd)),C1 )’ to the colouring function in
the gracful double wheel labelling specification defines the same solution set as
‘C1 is odd’ in La Bras et. al. [16]. Although it should be noted that the rules
presented here do not generate all streamliners previously reported.
 Automatically Generating Streamlined Constraint Models 489

Table 2. Number of conjectures generated for a set of problem classes. The Total
column lists all conjectures generated for each class, the Retain column lists the number
of conjectures that retain at least one solution and the Improve column lists the number
of conjectures that improve solver performance.

Problem Class Total Retain Improve

Wheel Graphs [9] 1479 1299 136
Double Wheel Graphs [15] 2142 1942 466
Graceful Graph Colouring [21]
Helm Graphs [3] 1428 1296 85
Gear Graphs [17] 1428 1214 70
QGE3 593 570 51
QGE4 593 572 34
Quasigroup Existence [23] QGE5 593 582 24
QGE6 593 569 44
QGE7 593 555 32
Equidistant Frequency Permutation Arrays [12] 560 377 2
Van Der Waerden Numbers [24] 433 364 14
Schur Numbers [22] 437 419 0
Langfords Problem [14] 357 228 24

4.3 Experimental Analysis

To evaluate this system it was used to streamline several different problem

classes, consisting of graceful graph colouring problems [21], quasigroup exis-
tence problems [23], equidistant frequency permutation arrays [12], Van Der
Waerden numbers [24] with k = 3 and l = 4, Schur numbers [22] and Langford’s
problem [14] (see Appendix A). The instances used to evaluate the performance
of the streamlined models were obtained by increasing the size of the integer
parameters until the original specification was unable to compute one solution
in under 100,000 search nodes. Experiments were run on a 32-core AMD Opteron
6272 at 2.1 GHz and took around 5 hours per problem class to complete.
Table 2 shows the total number of conjectures generated, the number of
conjectures that retain solutions to the original problem and the number of
conjectures that reduce the total number of search nodes explored after the
specification is refined, tailored and solved for all instances.
Figure 6 shows the search nodes and solve time to find the first solution for
the original specification vs the best streamlined model, where the quality metric
compares the number of instances solved in under 100,000 search nodes and
breaks ties by comparing the total number of search nodes across all instances.
Both the size of the search and search time are often reduced by orders of
magnitude by the streamlining constraints.
 490 J. Wetter et al.
Best Streamlined Specification Search Nodes

100000 100

Best Streamlined Specification solve Time

Wheels
Double Wheels
Gears
10000 10 Helms
EFPA
Langford
1000 1 VdW
QGE3
QGE4
QGE5
100 0.1 QGE6
QGE7

10 0.01

1 0.001
1 10 100 1000 10000 100000 0.001 0.01 0.1 1 10 100
Original Specification Search Nodes Original Specification Solve Time

Fig. 6. The size and execution time of search required to find the first solution for
a collection of problem classes for both the original specification and the streamlined
specification that resulted in the smallest cumulative search size across instances. The
generated conjectures often result in order of magnitude reduction in search size for
harder problem instances.

5 Identifying Effective Combinations of Conjectures

It has previously been observed that applying several streamlining constraints
to a model simultaneously can result in larger performance gains that any of the
constraints in isolation [15]. In order to find such combinations of constraints we
must consider the power set of constraints that retain solutions to the original
problem. In this section we investigate finding powerful combinations of con-
straints automatically with the use of pruning and make a comparison between
depth first search and breadth first search of the lattice of constraints.
For many of the problems considered here a large number of singleton con-
jectures that retain solutions are generated (see Table 2) resulting in power sets
too large to be exhaustively explored in practice. Two forms of pruning were
used to reduce the number of combinations to be considered:
1. if a set of conjectures fails all supersets are excluded from consideration (see
Figure 7),
2. trivially conflicting conjectures are not combined, for example we avoid forc-
ing a set to simultaneously contain only odd numbers and contain only even
numbers. We associate a set of tags with each of the rules in order to imple-
ment this pruning. Rules applied to the same variable that share tags are
not combined. This also removes the possibility of combining two different
conjectures that differ only by a softness parameter.
These pruning rules only remove combinations that are sure to fail, or are equiv-
alent to a smaller set of conjectures.
Even with this pruning the number of combinations to consider was found
to be too large to allow exhaustive enumeration. Therefore a traversal of the
lattice allowing good combinations to be identified rapidly is desired. Here we
 Automatically Generating Streamlined Constraint Models 491

ABCD

ABC ABD ACD BCD

AB AC AD BC BD CD

A B C D
Fig. 7. The power set of singleton conjectures can be explored to identify combinations
that result in powerful streamlined specifications. If small sets of conjectures that fail
to retain solutions are identified all super sets can be pruned from the search vastly
reducing the number of vertices to be explored.

experimented with depth first search (DFS) and breadth first search (BFS) of the
lattice. In order to guide both the searches the singleton conjectures were ordered
from best to worst, where the quality metric compared the number of instances
solved in under 100,000 search nodes and ties were broken by comparing the
total search nodes across all instances.
In order to compare the two approaches they were each allowed to evaluate
500 combinations of conjectures in addition to the singleton conjectures already
evaluated for three problem classes (Van der Waerden numbers, graceful helms
graphs and graceful double wheel graphs). The set of instances used for evalua-
tion was augmented by increasing the integer parameters until the best singleton
streamliner was unable to solve the instance in under 100,000 search nodes. Each
combination of conjectures was evaluated by refining, tailoring and solving for
the first solution using Conjure’s compact heuristic and default settings for
Savile Row and Minion. The experiments were performed on 32-core AMD
Opteron 6272 at 2.1 GHz taking approximately 6 hours for each problem class.
Figures 8 to 10 show the best set of conjectures found by this process, where
the quality metric compares the number of instances solved and ties were broken
by comparing the total number of search nodes.
Figure 8 shows three singleton conjectures that were found to produce an effec-
tive streamlined specification of the Van Der Waerden numbers problem, one of
which results from the chained application of four rules, whereas another results
directly from a single rule. Figure 9 shows two conjectures being combined for the
graceful helm graph problem, one of which results from the chained application
of three rules. Figure 10 shows the combination of two conjectures for the grace-
ful colouring of double wheel graphs. In all cases the combination of streamliners
results in better performance than any of the streamliners in isolation.
492 J. Wetter et al.

100000 10

10000 1

Time (s)
Nodes

1000 0.1

basic basic
100 parts(all(half(odd))) p 0.01 parts(all(half(odd))) p
parts(all(approxHalf(lowerHalf,2) p parts(all(approxHalf(lowerHalf,2) p
quasiRegular(3) p quasiRegular(3) p
combined combined
10 0.001
35 40 45 50 55 60 35 40 45 50 55 60
n n

Fig. 8. Combining singleton conjectures to produce a more effective streamlined model

for Van der Waerden numbers. The instances have k = 3, l = 4 and n varies. The
first conjecture ensures odd number are evenly distributed between the parts of the
partition. The second conjecture ensures the ‘small’ numbers are evenly distributed
between the parts of the partition. The third conjecture ensures the sizes of the partition
vary from each other by at most three.

On one problem, Van Der Waerden numbers, DFS performed very well, reach-
ing a powerful set of three conjectures within the first three models evaluated.
On the other problems DFS performed poorly, unable to beat the best pair found
by BFS within the allotted resource budget. In all three cases DFS failed to find
an improved model after the first five models it considered.
The poor performance of DFS can be attributed to two factors. First, the two
best singleton conjectures do not always produce the best pair of conjectures,
even when they retain solutions in combination. A better heuristic would need
some notion of complementary conjectures. Second, far more combinations are
pruned from the search space if failing sets are detected early. Consider the
lattice of conjecture sets shown in Figure 7. If conjecture C and D fail to retain
solutions when used in combination so will {A,C,D}, {B,C,D} and {A,B,C,D}.
A breadth first traversal would be guaranteed to detected this failure early and
would consequently never evaluate these three models. Alternatively a depth
first traversal would detect this failure late, and would therefore waste time
evaluating the supersets of {C,D}, all of which fail to retain solutions.

6 Discussion: Generating Streamliners in Practice

In this section, we consider the process of generating and selecting streamliners
when presented with a new problem class of interest. Our methodology is a close
analogue of that adopted by human experts in manual streamliner generation.
Given an Essence specification of the problem class, we begin by identifying
suitable instances with which to evaluate candidate streamliners. These instances
must be satisfiable and solvable in reasonable time so that they can be used in
the evaluation of a large set of candidate streamlined specifications. This set of
instances can be selected manually, or generated automatically by attempting
to solve candidates using the basic specification — satisfiable instances solved
within a budget are kept for streamliner evaluation.
 Automatically Generating Streamlined Constraint Models 493

100000 10

10000 1

Time (s)
Nodes

1000 0.1

100 basic 0.01 basic

restrict(prefix(monotonicDecreasing,3),W1) e restrict(prefix(monotonicDecreasing,3),W1) e
restrict(smallestLast,W1) c restrict(smallestLast,W1) c
combined combined
10 0.001
5 10 15 20 25 5 10 15 20 25
n n

Fig. 9. Combining singleton conjectures to produce a more effective streamlined model

for Graceful Helm Graphs. The parameter n is the size of the wheel. The first conjecture
requires that the differences between the labels of the vertices on the inner loop are
in decreasing order, except the last 3. The second requires that the smallest label
occurring on the inner loop is the last vertex.

100000 10

1
10000
Time (s)
Nodes

0.1

1000

basic 0.01 basic

prefix(smallestLast,1) c prefix(smallestLast,1) c
restrict(prefix(range(all(odd)),3),W1) c restrict(prefix(range(all(odd)),3),W1) c
combined combined
100 0.001
4 5 6 7 8 9 4 5 6 7 8 9
n n

Fig. 10. Combining singleton conjectures to produce a more effective streamlined

model for Graceful Double Wheel Graphs. The parameter n defines the size of sin-
gle wheel. The first conjecture requires that the last vertex in the outer loop of the
graph takes the largest value. The second requires that the difference between adjacent
vertices on the inner loop are odd numbers, except the last 3.

Streamliners are then generated, combined and evaluated against the set of
test instances, as described in Sections 4 and 5. This is a costly process, in the
same way that a considerable effort is expended by human experts in manual
streamliner generation. However, streamliners are generated for use over the
entire problem class. Under the assumption that our problem class has infinitely
many elements, the cost of streamliner discovery is amortised over all instances
not used in the streamliner evaluation process and becomes negligible.

7 Conclusion

Streamliner generation has been the exclusive province of human experts, requir-
ing substantial effort in examining the solutions to instances of a problem class,
manually forming conjectures as to good streamliners, and then testing their
efficacy in practice. In this paper we have demonstrated for the first time the
494 J. Wetter et al.

automated generation of effective streamliners, achieved through the exploitation

of the structure present in abstract constraint specifications written in Essence.
In future work we will expand our set of streamliner generation methods and
explore streamliner generation in further, more complex problem classes.

Acknowledgments. This work is supported by UK EPSRC grant EP/K015745/1.

We thank Ian Gent, Chris Jefferson and Peter Nightingale for helpful comments.

A Problem Descriptions
Van Der Waerden Numbers. Van Der Waerden’s theorem states that given
any positive integers k and l, there exists a positive integer n such that for any
partition of the integers 1 to n into k parts at least one of the parts contains an
arithmetic sequence of length l. The Van Der Waerden number, W (k, l), is the
lowest such n [24]. The Essence specification studied here describes a certificate
that the given n = W (k, l).
Schur Numbers. Given a positive integer r, there exists a positive integer s
such that for any partition of the integers 1 to s at least one part is not sum free.
Alternatively at least one part is a super set of {x, y, z} such that x + y = z.
Schur’s number, S(r), is the smallest such s [22]. The Essence specification
studied here describes a certificate that the given s = S(r)
Graceful Graphs. Given a graph with n edges a graceful labelling assigns
each node in the graph a label between 0 and n such that no label is used more
than once and that every pair of adjacent nodes can be uniquely identified by
the absolute difference of their labels. A graceful graph is a graph that permits
a graceful labelling [21]. Several classes of graph have been investigate in this
context including wheels [9], double wheels [15], helms [3] and gears [17].
Quasigroup Existsence. Given a positive integer n, does there exist a quasi-
group (latin square) of size n such that an additional side constraint is met.
These side constraints are: QGE3 - ∀a, b ∈ g (a · b) · (b · a) = a, QGE4 -
∀a, b ∈ g (b · a) · (a · b) = a, QGE5 - ∀a, b ∈ g ((b · a) · b) · b = a, QGE6 -
∀a, b ∈ g (a · b) · b = a · (a · b), QGE7 - ∀a, b ∈ g (b · a) · b = a · (b · a) [23].
Equidistant Frequency Permutation Arrays. Given v, q, λ and d, con-
struct a set of v codewords such that each code word is of length q · λ and
contains λ occurrence of each symbol in the set {1, 2, . . . , q}. Each pair of code
words must be of hamming distance d [12].
Langford’s Problem. Given any positive integer n, arrange copies of the
numbers between 1 and n such that for all k in {1 . . . n} there are k digits
between occurrences of k [14].
 Automatically Generating Streamlined Constraint Models 495

References
1. Akgun, O., Frisch, A.M., Gent, I.P., Hussain, B.S., Jefferson, C., Kotthoff, L.,
Miguel, I., Nightingale, P.: Automated symmetry breaking and model selection in
Conjure. In: Schulte, C. (ed.) CP 2013. LNCS, vol. 8124, pp. 107–116. Springer,
Heidelberg (2013)
2. Akgün, O., Miguel, I., Jefferson, C., Frisch, A.M., Hnich, B.: Extensible automated
constraint modelling. In: AAAI 2011: Twenty-Fifth Conference on Artificial Intel-
ligence (2011)
3. Ayel, J., Favaron, O.: Helms are graceful. Progress in Graph Theory (Waterloo,
Ont., 1982), pp. 89–92. Academic Press, Toronto (1984)
4. Charnley, J., Colton, S., Miguel, I.: Automatic generation of implied constraints.
In: ECAI, vol. 141, pp. 73–77 (2006)
5. Colton, S., Miguel, I.: Constraint generation via automated theory formation. In:
Walsh, T. (ed.) Proceedings of the Seventh International Conference on Principles
and Practice of Constraint Programming, pp. 575–579 (2001)
6. Colton, S.: Automated Theory Formation in Pure Mathematics. Ph.D. thesis, Uni-
versity of Edinburgh (2001)
7. Frisch, A.M., Jefferson, C., Hernandez, B.M., Miguel, I.: The rules of constraint
modelling. In: Proc. of the IJCAI 2005, pp. 109–116 (2005)
8. Frisch, A.M., Harvey, W., Jefferson, C., Martı́nez-Hernández, B., Miguel, I.:
Essence: A constraint language for specifying combinatorial problems. Constraints
13(3), 268–306 (2008). doi:10.1007/s10601-008-9047-y
9. Frucht, R.: Graceful numbering of wheels and related graphs. Annals of the New
York Academy of Sciences 319(1), 219–229 (1979)
10. Gent, I.P., Jefferson, C., Miguel, I.: Minion: A fast scalable constraint solver.
In: ECAI, vol. 141, pp. 98–102 (2006)
11. Gomes, C., Sellmann, M.: Streamlined constraint reasoning. In: Wallace, M. (ed.)
CP 2004. LNCS, vol. 3258, pp. 274–289. Springer, Heidelberg (2004)
12. Huczynska, S., McKay, P., Miguel, I., Nightingale, P.: Modelling equidistant fre-
quency permutation arrays: an application of constraints to mathematics. In:
Gent, I.P. (ed.) CP 2009. LNCS, vol. 5732, pp. 50–64. Springer, Heidelberg (2009)
13. Kouril, M., Franco, J.: Resolution tunnels for improved SAT solver performance. In:
Bacchus, F., Walsh, T. (eds.) SAT 2005. LNCS, vol. 3569, pp. 143–157. Springer,
Heidelberg (2005)
14. Langford, C.D.: Problem. The Mathematical Gazette, 287–287 (1958)
15. Le Bras, R., Gomes, C.P., Selman, B.: Double-wheel graphs are graceful.
In: Proceedings of the Twenty-Third international Joint Conference on Artificial
Intelligence, pp. 587–593. AAAI Press (2013)
16. Le Bras, R., Gomes, C.P., Selman, B.: On the erdős discrepancy problem. In:
O’Sullivan, B. (ed.) CP 2014. LNCS, vol. 8656, pp. 440–448. Springer, Heidelberg
(2014)
17. Ma, K., Feng, C.: On the gracefulness of gear graphs. Math. Practice Theory 4,
72–73 (1984)
18. Marriott, K., Nethercote, N., Rafeh, R., Stuckey, P.J., de la Banda, M.G., Wallace, M.:
The design of the zinc modelling language. Constraints 13(3) (2008). doi:10.1007/
s10601-008-9041-4
19. Nightingale, P., Akgün, Ö., Gent, I.P., Jefferson, C., Miguel, I.: Automatically
improving constraint models in Savile Row through associative-commutative com-
mon subexpression elimination. In: O’Sullivan, B. (ed.) CP 2014. LNCS, vol. 8656,
pp. 590–605. Springer, Heidelberg (2014)
496 J. Wetter et al.

20. Prestwich, S.: CSPLib problem 028: Balanced incomplete block designs. http://
www.csplib.org/Problems/prob028
21. Rosa, A.: On certain valuations of the vertices of a graph. In: Theory of Graphs
Internat. Symposium, Rome, pp. 349–355 (1966)
22. Schur, I.: Über die kongruenz xm + y m ≡ z m (mod p). Jahresber. Deutsch. Math.
Verein 25, 114–117 (1916)
23. Slaney, J., Fujita, M., Stickel, M.: Automated reasoning and exhaustive search:
Quasigroup existence problems. Computers & Mathematics with Applications
29(2), 115–132 (1995)
24. Van der Waerden, B.L.: Beweis einer baudetschen vermutung. Nieuw Arch. Wisk
15(2), 212–216 (1927)
Application Track
 Constraint-Based Local Search
for Finding Node-Disjoint Bounded-Paths
in Optical Access Networks

Alejandro Arbelaez(B) , Deepak Mehta, and Barry O’Sullivan

Insight Centre for Data Analytics, University College Cork, Cork, Ireland
{alejandro.arbelaez,deepak.mehta,barry.osullivan}@insight-centre.org

Abstract. Mass deployment of fibre access networks is without doubt

one of the goals of many network operators around the globe. The Pas-
sive Optical Network has been proposed as a solution to help deliver
mass deployment of fibre-to-the-home (FTTH), by reducing the cost per
customer compared to current commercially available technologies. A
major failure in the access network (e.g., fibre cut, amplifier failure, or
other equipment can fail) might affect tens of thousands of customers.
Therefore, protecting the network from such failures is one of the most
important requirements in the deployment of FTTH solutions. In this
paper we use a constraint-based local search approach to design reliable
passive optical networks via node-disjoint paths whereby each customer
is protected against any one node or link failure. We experiment with a
set of very large size real-world networks corresponding to Ireland, Italy
and the UK and demonstrate the tradeoff between cost and resiliency.

1 Introduction
Continuous growth in the amount of data transferred within national and global
networks in the last decade necessitates new infrastructures and data transfer
technologies [1]. In line with this, the goal of the DISCUS project [2] is to develop
an end-to-end network design that can provide high-speed broadband capability
of at least three orders-of-magnitude greater than today’s networks to all users,
reduce energy consumption by 95%, and remain economically viable.1 The archi-
tecture is based on maximising the sharing of network infrastructure between
customers by deploying a Long-Reach Passive Optical Network (LR-PON) [3] in
the access part. All network points have equal bandwidth and service capability
including core bandwidth (10Gbs to 100Gbs) delivered to the access network.
LR-PONs provide an economically viable solution for fibre-to-the-home net-
work architectures [4]. In a LR-PON fibres are distributed from metro nodes
(MNs) to exchange sites (ESs) through cables that form a tree distribution net-
work. Typically due to signal loss the maximum distance from a MN to an ES is
up to 90 km, and the maximum distance from an ES to the customers is up to
10 km. A major fault occurrence in a LR-PON would be a complete failure of a
1
http://www.discus-fp7.eu

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 499–507, 2015.
DOI: 10.1007/978-3-319-23219-5 35
500 A. Arbelaez et al.

MN, which could affect tens of thousands of customers. A dual homing protec-
tion mechanism is recommended for LR-PONs whereby customers are connected
to two MNs, so that whenever a single MN fails all customers are still connected
to a back-up. Simply connecting two MNs to an ES is not sufficient to guarantee
the connectivity because if a link or a node is used is common in the routes of
fibre going from ES to its 2 MNs then both MNs would be disconnected.
Broadly speaking, there are two protection strategies for dual homing. One
approach is protecting links in the tree distribution network, e.g., a fibre cut,
amplifier failure, or other equipment that can fail. This strategy, known as the
edge-disjoint solution, allows switching to an alternative path whenever a link in
the distribution network fails. Alternatively, protection can be provided at the
node level with node-disjoint paths between MNs and ESs, when an entire ES
fails and all adjacent links to the node will be affected. This strategy, known
as node-disjoint solution, provides a stronger protection mechanism and allows
switching to an alternative path whenever a node fails. The selection of one
protection mechanism over another is a matter of choice for the broadband
provider, generally higher protection means higher deployment cost.
Figure 1 shows an example with two MNs F1 and F2 and the set of ESs
{a, b, c, d, e, f }. Black and gray coloured edges are used to show the trees corre-
sponding to the facilities F1 and F2 respectively, and maximum allowed distance
λ=12. This example shows three scenarios satisfying the length constraint, how-
ever, each scenario shows a different level of resiliency. Dashed lines indicate
edges leading to a conflict, i.e., violating node or edge disjointness, and gray
nodes indicate the conflicting node when violating node-disjointness. The total
solution cost for Figure 1(a) is 46. The indicated solution satisfies the length con-
straint (i.e., the distance from the MNs to any ES is at most λ=12) and it also sat-
isfies node-disjoint constraints and consequently edge-disjoint constraints. Here
we observe that failure of any single link or node would not affect the remain-
ing ESs. Figure 1(b) does not satisfy node-disjoint constraints but edge-disjoint
constraints. Here we observe that a single node failure could disconnect one or
more ESs from the MNs F1 and F2 , e.g., f would be disconnected from F1 and
F2 whenever a fails. Nevertheless, the solution is resilient to any single facility
or single link failure. Figure 1(c) shows a solution that violates both node and
edge disjoint constraints. In this example the link e, f  appears in both trees
and the path from f is neither edge-disjoint nor node-disjoint but it is resilient
to a failure of a single MN.
Network operators tend to evaluate multiple options. Their choices of pro-
viding resiliency depends on many factors, e.g., captial and operational costs,
rural or sparse regions, business or residential customers etc. Ideally they would
prefer a tool where one can try to impose different configurations and get feed-
back quickly. In particular we focus on node-disjointness as the other options
can be seen as relaxation of the node-disjoint constraints. The problem is to
determine the routes of cable fibres such that there are two node-disjoint paths
from each ES to its two MNs, the length of each path must be below a given
threshold and the total cable length required for connecting each ES to two MNs
 Constraint-Based Local Search for Finding Node-Disjoint Bounded-Paths 501

F1 F2 F1 F2 F1 F2

5 7 5 7 5 7
4 8 4 8 4 8

a 3 a 3 a 3
b 2 b 2 b 2
4 2 6 4 2 6 4 2 6
3 c 3 c 3 c
d e d e d e
1 1 2 1 1
1
f f f

(a) Node and (b) Not Node but (c) Neither Node
Edge disjoint Edge Disjoint nor Edge Disjoint

Fig. 1. Passive Optical Networks with and without node and edge disjoint protection.

should be minimised. The bottleneck is that the instances of the network sizes
can be very huge e.g., UK has approximately 5000 ESs while Italy has more
than 10,000. Solving very large instances of this problem in limited time is very
challenging. We therefore apply a constraint-based local search and develop an
efficient algorithm to handle node-disjoint constraints for selecting best move
during LS.

2 Constraint-Based Local Search

Constraint-based local search has been recently introduced as an alternative
solution to efficiently tackle many network design problems [5,6] ranging from
telecommunications to transportations, and VLSI circuit design.
The local search (LS) algorithm starts with an initial solution and iteratively
improves the solution by performing small changes. A move operator guides the
algorithm towards better solutions. [5] defines the node and subtree operator,
the algorithm moves a given node (resp. a subtree and the nodes emanating
from it) from one location to another in the tree by improving the objective.
These operators have been used to find distance-constrained and edge-disjoint
paths. Empirical results suggest that the subtree operator outperforms the node
operator. [6] defines the arc operator: the algorithm moves a given arc from one
location to another in the tree, this operator has been used in the context of
the routing and wavelength assignment problem [7] and to find edge disjoint
paths in a given graph [8]. We remark that node and subtree move-operators
are specifically designed for tree structures whereas the arc operator is more
general and it can be applied to more general graphs. Nevertheless, the former is
more efficient when we are dealing with trees, while the latter is more expensive.
In this paper we limit our attention to using the subtree move-operator as it
provides a better time complexity than that of the arc operator, i.e., O(n) vs.
O(n3 ) when considering the distance constraint. We refer the reader to [5] for a
detailed analysis of the time complexities of the local search operators.
Other related work includes [9] where the authors proposed a dedicated algo-
rithm to tackle the Distance Constrainted Minimum Spanning Tree problem.
502 A. Arbelaez et al.

The authors studied two move-operators to tackle the distance constraint, infor-
mally speaking nodes are sequentially added in the tree using Prim’s algorithm,
and a backup route is used whenever the move cannot satisfy the length con-
straint. This framework is difficult to extend with side constraints, and therefore
the algorithms cannot be extended to solve our problem with node and edge
disjointed paths. For instance, the operators rely on a pre-computed alternative
route from the root to any node. However that route might violate disjointedness.

2.1 Node-Disjointness
The LS algorithm uses the move-operators in two different phases. In the inten-
sification phase a subtree is selected and it is moved to another location that
provides the maximal improvement in the objective function. In the perturba-
tion phase both a subtree and its new location are selected randomly in order
diversify the search by adding noise to the current incumbent solution. In order
to complete the intensification and diversification phases, the sub-tree opera-
tor requires four main functions: removing a subtree from the current solution;
checking whether a given solution is feasible or not; inserting a subtree in the
solution; finding the best location of a subtree in the current solution.
Additionally, the LS algorithm is equipped with an incremental way of main-
taining information related to checking constraints and computing the cost of
the assignment. For the length constraint it is necessary to maintain for each
node ej the length of the path from MN i to ES (or node in the tree) j, and
the length of the path emanating from ej down to the farthest leaf from ej .
These two lengths are updated after removing and inserting a subtree in the
tree. For enforcing the edge-disjoint constraint in the two paths of a client, we
must maintain the predecessors of each client in the two paths associated to the
MNs, and require that the two predecessors must be different [5]. For enforcing
the node-disjoint constraint between the paths of two clients it is necessary to
maintain all the nodes occurring in each path except the source and target. This
is done by maintaining the transitive graph.
Let M be the set of facilities or MCs. Let E be the set of clients or ESs.
Let Ei ⊆ E be the set of clients associated with MN mi ∈ M . We use N to
denote the set of nodes, which is equal to M ∪ E. We use Ti to denote the tree
network associated with MN i. We also use Ni ⊆ N = Ei ∪ {mi } to denote the
set of nodes in Ti . Let λ be the maximum path-length from a facility to any of
its clients.
Let ej be a node in the tree Ti associated with facility mi . Let epj be the
predecessor of ej , and let Lej be a list of all nodes in the subtree starting at ej of
tree Ti . We now describe the complexities of remove, insert, feasibility check and
best location operations performed during search with respect to node-disjoint
constraint:
Remove. To remove a subtree it is necessary to update the transitive graph.
This is done by removing all the transitive links between any node in the path
from the MN mi down to ej and all the elements in Lej . In Figure 2(a) gray lines
show the links that must be removed from the transitivity graph, i.e., links that
 Constraint-Based Local Search for Finding Node-Disjoint Bounded-Paths 503

1 1

5
2 2
3 3
4 5 10 11
4

8 9 10 11 6 7 8 9 6 7

(a) Remove the subtree (b) Insert the subtree

rooted at node 5 rooted at 5 between node 1
and node 3

Fig. 2. Removing/Inserting a subtree while maintaining transitive links.

reach any node in the subtree emanating from node 5. Thus, the time-complexity
of removing a subtree rooted at ej from the tree Ti is quadratic with respect to
the number of clients of facility mi .
Insert. The insertion involves adding the transitive links between each node in
the path from the MN down to ej and each element of Lej . Figure 2(b) shows the
links that must be added in the transitivity graph; black links denote existing
links in the tree and gray links denote the new links added by transitivity in the
graph. Therefore, a subtree rooted at ej can be inserted in Ti in quadratic time
complexity with respect to the number of clients of facility mi .
Feasibility. To verify the consistency of the node-disjoint constraint we need
to check that any node occurring in the subtree starting from the node ej is
not transitively connected to any node in the path from the MN down to the
potential predecessor of ej in any other tree corresponding to other MNs. This is
done by checking the occurrence of the links corresponding to the pairs of nodes
occurring in the transitive graph. As the number of links can be quadratic, the
time complexity is quadratic. Additionally, if the node ej is breaking an existing
arc ep , eq , it is necessary to check that ej is not transitively connected to any
node in the subtree emanating from eq .
Best Location. Selecting the best location for a given subtree involves travers-
ing all the nodes of the tree associated with the MN and selecting the one that
maximises the reduction in the cost. As the complexity of feasibility is quadratic,
the time complexity of finding the best location is cubic in the number of nodes
of a given tree.
The complexities for feasibility checking and best location described above do
not make any assumption on the order in which different locations are explored.
In the following we will show that the amortized complexities can be reduced if
the locations in the tree are traversed in a particular order.
For a given subtree rooted at ej there are two types of feasible locations. The
first is a node-type location, say ek , of the tree and the feasibility check is whether
the parent of ej can be ek or not. This type of location is denoted by (ek , ∅). The
second is a link-type location, say epl , el , of the tree and the feasibility check
is whether we can break the link epl , el  by making epl and ej the parents of ej
and el respectively. This type of location is denoted by (epl , {el }). For node-type
504 A. Arbelaez et al.

Algorithm 1. BestLocation(ej , Ti , G)
1: Remove subtree rooted at ej from Ti , update Ti and G
2: Nlocations ← ∅, Llocations ← ∅
3: Let s be a stack, let q be a queue
4: s.push(root(Ti ))
5: while s = ∅ do
6: ek ← s.pop()
7: trans ← {ek , el ) | el ∈ Lej ∧ ek , el  ∈ G}
8: if trans = ∅ then
9: Nlocations ← Nlocations ∪ {(ek , ∅)}
10: for all ek , el  ∈ Ti do
11: Llocations ← Llocations ∪ (ek , {el })
12: s.push(el )
13: end for
14: if out(ek , Ti ) = 0 then
15: q.append((ek , check))
16: while q = ∅ do
17: (el , status) ← q.remove()
18: if epl , el  ∈ Llocations then
19: if ej , el  ∈ G||status =rem then
20: Llocations ← Llocations \{epl , el }
21: q.append((epl , rem))
22: else
23: q.append((epl , check))
24: Best ← SelectBest(ej , NLocations ∪ Llocations)
25: return Best

locations we explore the tree in a depth-first manner, whereas for link-type loca-
tions we explore the tree in a bottom-up manner.
Algorithm 1 shows the pseudo-code to compute the feasible set of locations
for a given subtree rooted at ej and then select the best location in a tree Ti with
respect to a transitive graph G. The algorithm maintains the set of feasible node-
type locations with respect to the node disjoint constraint in the set Nlocations,
and employs a depth-first search algorithm with a stack s (Lines 5-18). It explores
the successors of a node ek only when there is no existing link in the transitive
graph between the node ek and any node in Lej . Notice that the time complexity
of the first while loop is quadratic in the number of the nodes of the tree. Also
notice that in the first while loop, the algorithm updates the set Llocations by
adding all the candidate links starting from the node ek which will be filtered in
the next while loop. Moreover, for each leaf node, when the outgoing degree of
node ek is 0 (i.e. out(ek , Ti ) = 0), the node is appended to the queue with the
status check as the feasible set of link locations follows a bottom-up approach.
In the second while loop, the algorithm filters the set of candidate links to
determine the feasible set by employing a bottom-up approach with a queue
q. It explores a node el only if its predecessor is a valid parent candidate for
the node ej of a given subtree. If there is a link from ej to el in the transitive
 Constraint-Based Local Search for Finding Node-Disjoint Bounded-Paths 505

graph or its status is marked as rem then (epl , el ) is not a feasible link-location
and subsequently all the links involved in the path from the facility to el would
not be feasible link-locations. Therefore, the queue is appended with (epl , rem)
and the set Llocations is filtered. Otherwise, the status of the predecessor of
el is set to check and appended to the q. Notice that the time complexity of
second while loop is linear in the number of links of the tree. Thus, the amortized
time complexity of feasibility checking is linear and, consequently, the amortized
complexity of BestLocation is quadratic.
We remark that computing fji for checking the length constraint can be
merged with the verification of the transitive graph with the same worst-case
time complexity i.e., linear. However we still incrementally maintain fji and bij
for each node in the graph in order to amortize the complexity of the operation
to depend only in the number of nodes in the subtree enumerating from ej .

3 Empirical Evaluation
All experiments were performed on a 4-node cluster, each node features 2 Intel
Xeon E5430 processors at 2.66Ghz and 12 GB of RAM memory. We report the
median cost solution of 11 independent executions of the LS algorithm with a
1-hour time limit for each experiment. We would like to point out that CPLEX
ran out of memory when we tried to create and solve the MIP model for nearly
all instances. Hence, we limit our attention to the CBLS algorithm.
To evaluate the performance of the proposed LS we use 3 real networks from
three EU countries: Ireland with 1121 ESs, the UK with 5393 ESs, and Italy
with 10708 ESs. For each dataset we use [10] to identify the position of the
MNs and computed four instances for each country. Ireland with 18, 20, 22, and
24 MNs; the UK with 75, 80, 85, and 90 MNs; and Italy with 140, 150, 160,
and 170 MNs. We set λ=90 km. Cable fibre cost is a dominant factor in the
deployment of LR-PONs, therefore we report the estimated total cost in euro
to deploy edge and node disjoint distribution networks. To this end, we use the
subtree operator to optimise the total cable link distance and then we extract
the actual fibre cost of the solutions. [11] suggests that on average four fibres will
be needed to provide broadband services to all customers for each ES, and the
cost of the fibres are defined as follows: {12, e2430}, {24, e2430}, {48, e3145},
{96, e4145}, {144, e5145}, {192, e6145}, {240, e7145}, {276, e7859}, the left
part indicates the number of fibres per cable and the respective cost per km.
In Table 1 we report results. As expected the cost of node-disjointness is
higher than that of edge-disjointness as the former provides stronger protection.
However, it is worth pointing out that the total cable link distance of deployment
of a fully node disjoint solution is only up to 13.7% more expensive for Ireland,
16.3% for the UK, and 19.7% for Italy, with respect the best known edge-disjoint
solutions. And with a gap between 22% and 31% with respect to the LBs,2
2
We recall that the lower bounds are valid for both edge and node disjoint paths.
Therefore we expect the node disjoint solution to be much closer to the actual
optimal solution.
506 A. Arbelaez et al.

Table 1. Results for Ireland, UK and Italy to provide Node and Edge disjoint paths.

Cable Link distance Fibre Cost

Country |M | Protection GAP (%) in Mill e
LB
Node Edge Node Edge Node Edge
18 14809 19824 17107 25.78 13.43 55 49
Ireland 20 14845 19508 16819 24.27 11.73 54 48
|E|=1121 22 14990 19076 16711 22.18 10.29 53 48
24 14570 18452 16163 22.23 9.85 51 46
75 54720 78131 65377 30.31 16.30 223 197
UK 80 54975 77269 64565 29.20 14.85 220 194
|E|=5393 85 55035 75799 63517 28.00 13.35 216 191
90 55087 74376 62163 25.93 11.38 212 188
140 76457 111434 89418 31.43 14.49 335 293
Italy 150 76479 109954 88255 30.45 13.34 329 288
|E|=10708 160 76794 109708 88336 30.05 13.06 326 286
170 77013 108288 87405 29.09 11.88 321 282

which is used as the reference baseline for comparison. Finally, we would like
to highlight the cost in euro of deployment both the edge and node disjoint
alternatives. We foresee that the cost of deployment node disjoint protection in
LR-PON in Ireland, the UK, and Italy is respectively 9.9%, 11%, and 12% more
expensive (on average) than the corresponding edge disjoint alternative.

4 Conclusions and Future Work

We have extended an existing constraint-based local search framework for finding

the Distance-Constrained Node-Disjoint paths between a given pairs of nodes.
The efficiency of our approach is demonstrated by experimenting with a set of
problem instances provided by network operators in Ireland, the UK, and Italy.
We have seen that providing node-disjointness increases the cable link distance
cost up to 13% for Ireland, 16% for the UK, and 19% for Italy with respect
to the best known solutions with a weaker protection mechanism. Addition-
ally, we provided a cost estimation in euro to deploy both protection solutions.
Our experiments indicate that implementing node disjointness is between 9%
and 11% more expensive with respect to the best known solutions with weaker
protection.

Acknowledgments. We would like to thank both eircom and Telecom Italia for pro-
viding their optical network data. We would also like to thank atesio for providing the
Italian reference network on which the Italian experiments have been carried out. This
work is supported by the EU Seventh Framework Programme (FP7/ 2007-2013) under
grant agreement no. 318137 (DISCUS) and by Science Foundation Ireland (SFI) under
grant 08/CE/I1423. The Insight Centre for Data Analytics is also supported by SFI
under Grant Number SFI/12/RC/2289.
 Constraint-Based Local Search for Finding Node-Disjoint Bounded-Paths 507

References
1. Benhamiche, A., Mahjoub, R., Perrot, N.: On the design of optical OFDM-based
networks. In: Pahl, J., Reiners, T., Voß, S. (eds.) INOC 2011. LNCS, vol. 6701,
pp. 1–6. Springer, Heidelberg (2011)
2. Ruffini, M., Wosinska, L., Achouche, M., Chen, J., Doran, N., Farjady, F., Mon-talvo,
J., Ossieur, P., O’Sullivan, B., Parsons, N., Pfeiffer, T., Qiu, X.Z., Raack, C., Rohde,
H., Schiano, M., Townsend, P., Wessaly, R., Yin, X., Payne, D.B.: Discus: An end-to-
end solution for ubiquitous broadband optical access. IEEE Communications Maga-
zine 52(2), 24–32 (2014)
3. Davey, R., Grossman, D., Rasztovits-Wiech, M., Payne, D., Nesset, D., Kelly, A.,
Rafel, A., Appathurai, S., Yang, S.H.: Long-reach passive optical networks. Journal
of Lightwave Technology 27(3), 273–291 (2009)
4. Payne, D.B.: FTTP deployment options and economic challenges. In: Proceedings of
the 36th European Conference and Exhibition on Optical Communication (ECOC
2009) (2009)
5. Arbelaez, A., Mehta, D., O’Sullivan, B., Quesada, L.: Constraint-based local search
for edge disjoint rooted distance-constrainted minimum spanning tree problem.
In: CPAIOR 2015, pp. 31–46 (2015)
6. Pham, Q.D., Deville, Y., Van Hentenryck, P.: LS(Graph): a constraint-based local
search for constraint optimization on trees and paths. Constraints 17(4), 357–408
(2012)
7. Mukherjee, B.: Optical WDM Networks (Optical Networks). Springer-Verlag New
York Inc., Secaucus (2006)
8. Blesa, M., Blum, C.: Finding edge-disjoint paths in networks: An ant colony
optimization algorithm. Journal of Mathematical Modelling and Algorithms 6(3),
361–391 (2007)
9. Ruthmair, M., Raidl, G.R.: Variable neighborhood search and ant colony opti-
mization for the rooted delay-constrained minimum spanning tree problem.
In: Schaefer, R., Cotta, C., Kolodziej, J., Rudolph, G. (eds.) PPSN XI. LNCS,
vol. 6239, pp. 391–400. Springer, Heidelberg (2010)
10. Ruffini, M., Mehta, D., O’Sullivan, B., Quesada, L., Doyle, L., Payne, D.B.:
Deployment strategies for protected Long-Reach PON. Journal of Optical Com-
munications and Networking 4, 118–129 (2012)
11. DISCUS: Project Deliverable 2.6, Architectural optimization for different geotypes
 Open Packing for Facade-Layout Synthesis
Under a General Purpose Solver

Andrés Felipe Barco1(B) , Jean-Guillaume Fages2 , Elise Vareilles1 ,

Michel Aldanondo1 , and Paul Gaborit1
1
Université de Toulouse, Mines d’Albi, Route de Teillet Campus Jarlard,
81013 Albi Cedex 09, France
abarcosa@mines-albi.fr
2
COSLING S.A.S., 2 Rue Alfred Kastler, 44307 Nantes Cedex 03, France

Abstract. Facade-layout synthesis occurs when renovating buildings to

improve their thermal insulation and reduce the impact of heating on the
environment. This interesting problem involves to cover a facade with a
set of disjoint and configurable insulating panels. Therefore, it can be
seen as a constrained rectangle packing problem, but for which the num-
ber of rectangles to be used and their size are not known a priori. This
paper proposes an efficient way of solving this problem using constraint
programming. The model is based on an open variant of the DiffN global
constraint in order to deal with an unfixed number of rectangles, as well
as a simple but efficient search procedure to solve this problem. An empir-
ical evaluation shows the practical impact of every choice in the design
of our model. A prototype implemented in the general purpose solver
Choco is intended to assist architect decision-making in the context of
building thermal retrofit.

1 Introduction

Currently buildings energetic consumption represents more than one third of

the total energy consumption in developed countries [4,6,16]. One strategy for
reducing such energy consumption lies on buildings thermal retrofit achieved
either by an internal or an external insulation [9]. Among several options [9], an
external insulation may be based on covering the entire building with an envelope
made out of rectangular wood-made panels [7,23]. However, some difficulties
are present when targeting such renovation in industrial proportions, e.g. in a
country. These difficulties include slow conception using by hand configuration,
human scheduling and craft assembly. In consequence, it is essential to assist
this massive retrofit of buildings with decision support systems [10].
A crucial aspect of the retrofit automation lies in its facade layout-synthesis.
Simply stated, given a rectangular facade surface and an undetermined number

The authors wish to acknowledge the TBC Générateur d’Innovation company, the
Millet and SyBois companies and all partners in the CRIBA project, for their involve-
ment in the construction of the CSP model.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 508–523, 2015.
DOI: 10.1007/978-3-319-23219-5 36
 Open Packing for Facade-Layout Synthesis Under a General Purpose Solver 509

of rectangular panels, find a solution on how to determine the number of panels,

assign size to them and place them over the facade. The family of these problems
is called layout synthesis [13] and are by nature combinatorial problems [5,24].
Three characteristics make this problem novel and interesting:

1. Unlike most works [12,13], the number of panels to be allocated in a non-

overlapping fashion is not known a priori.
2. Some rectangular areas inside facades (existing windows and doors) are
meant to be completely overlapped by panels. Each area must be covered
by one and only one panel in which the corresponding hole will be manufac-
tured once the layout definition is done.
3. Facades have specific areas to attach panels that are strong enough to sup-
port their weight.

Due to the rectangular geometry of panels and facades, this problem may be
treated as a constrained two-dimensional packing problem [8]. By doing this, we
may take advantage of the great body of literature on two-dimensional packing
while exploiting technological tools, such as general purpose constraint solvers,
to tackle the problem. Indeed, Constraint Programming (CP) is, arguably, the
most used technology at the crossroads of Artificial Intelligence and Operations
Research to address combinatorial problems. CP provides a declarative language
and efficient solvers to model and solve complex decision problems where vari-
ables are subject to various constraints. It is known to solve efficiently packing
problems [3] having, among other abstractions, the geometrical constraint GEOST
[2]. However, as we only deal with rectangular shapes, the constraint GEOST [2]
seems too complex for our need and would bring an unnecessary risk from a
software maintenance point of view. Instead, we use the simpler and well known
non overlapping DiffN global constraint [3]. Moreover, we exploit the possibil-
ities, provided by general purpose CP solvers, to implement ad hoc constraints
and search procedures that fit the problem structure. Thus, we consider a CP
solver as an algorithm integration framework for the development of a decision
support application.
Nevertheless, not having a predefined number of rectangles becomes a draw-
back given that the great majority of constraint programming environments
implement global constraints and search engines with a fixed set of variables. In
fact, performing filtering and searching using an unfixed number of variables, i.e.,
a dynamically changing problem, is an active research topic in the constraint pro-
gramming community. In [1], the author solves the problem of unknown variables
by dynamically adding variables while exploring the search tree. In essence, it
introduces a setup in which constraints may be deactivated to be replaced with
new activated constraints involving more or less variables. Nonetheless, even
though the idea seems simple, a good implementation is intricate. Instead, our
work is inspired from [22], where the authors introduce open global constraints.
An open global constraint is an extension of an existing global constraint that
includes a set variable (or an array of binary variables) to indicate the subset of
decision variables the former constraint holds on. In other words, some decision
510 A.F. Barco et al.

variables of the problem become optional (see [11,19] and Section 4.4.16 in [18]
for further information).
The aim of this paper is to propose a solution to the facade-layout synthe-
sis problem as a two-dimensional packing problem with optional rectangles. We
do so by using an open variant of the DiffN constraint [3] to handle rectan-
gles that are potentially in the solution. Also, we present a simple yet efficient
search heuristic which captures the problem structure. The proposed solutions
are implemented using the open-source constraint environment Choco [17]. An
empirical evaluation shows the practical impact of every contribution and pro-
vides a better understanding of the solving process. The paper is divided as
follows. In Section 2 the facade-layout elements are introduced. In Section 3, the
constraint-based definition of the problem is presented. In Section 4 we provide
technical details of our implementation. In Section 5, a search heuristic that cap-
tures the problem structure is presented. Afterwards, in Section 6, we show some
performance evaluation of our methods. Finally, some conclusions are discussed
in Section 7.

2 Retrofit Industrialization
This work is part of project called CRIBA (for its acronym in French of Con-
struction and Renovation in Industrialized Wood Steel) [7]. This project focuses
on the industrialization of energetic renovation for residential buildings. The
challenge, very ambitious, is to have a building energetic performance under
25kW h/m2 /year after the renovation. The complete renovation (internal and
external retrofit) started at the beginning of 2015 with the internal part only.
The industrialization is based on an external new thermal envelope which
wraps the whole buildings. The envelope is composed of prefabricated rectan-
gular panels comprising insulation and cladding, and sometimes including in
addition, doors, windows and solar modules. As a requirement for the reno-
vation, facades have to be strong enough to support the weight added by the
envelope.
Within CRIBA several tools, needed to industrialize the renovation process,
will be developed:
a. a new method for three-dimensional building survey and modeling (building
information model),
b. a configuration system for the design of the new thermal envelope (topic of
this paper), and
c. a working site planning model with resource constraints.
This section introduces the problem of facade layout-synthesis from the
industrial point of view.

2.1 Elements
Facades. A facade is represented by a two-dimensional coordinate plane (see
Figure 1), with origin of coordinates (0,0) at the bottom-left corner of the facade,
and contains rectangular zones defining:
 Open Packing for Facade-Layout Synthesis Under a General Purpose Solver 511

– Perimeter of facade with it size (height and width in meters).

– Frames (existing windows and doors over the facade) play an important role
as they are meant to be completely overlapped by one and only one panel.
Frames are defined with:
• Origin point (x,y) with respect to origin of facade.
• Width and height (in meters).
– Supporting areas. As the layout problem must deal with a perpendicular
space plan, gravity must be considered. It turns out that some areas over
the facade have load bearing capabilities that allow us to attach panels.
Supporting areas have well-defined:
• Origin point (x,y) with respect to origin of facade.
• Width and height (in meters).

Fig. 1. Facades: Frames and supporting areas.

Rectangular panels. Panels are rectangular (see Figure 2), of varying sizes and
may include different equipment (solar modules, window-holes, shutters, etc.).
These panels are designed one at a time in the process of layout synthesis and
manufactured in the factory prior to shipment and installation on the building
site. These panels have a well-defined:
– Size (height and width in meters). Height and width are constrained by a
given lower and upper bound provided that is consequence of environmental
or manufacturing limitations.
– Thickness and insulation. Thermal performance of a given panel depends
on several properties: Size, thickness and insulation type. Consider that the
smaller the thickness of the panel the better quality should be the insulation
in order to reach performance objectives.
– New frames (such as new doors and new windows). Given internal structure
of rectangular panels, new frames must respect a parameterizable minimum
distance (Δ) with respect to panel’s borders.
– Cost depending mainly on size and attached equipment (in Euros).
– Thermal performance (in watts per meter square-kelvins, w.m−2 .k −1 )
depending on size, chosen thickness and insulation type.
512 A.F. Barco et al.

Fig. 2. Rectangular parameterizable panel.

2.2 Problem Features

As mentioned in the introduction, there are three key issues reflected from
the industrial scenario. The unfixed number of panels is the most problematic.
Figure 3 depicts restrictions of the last two issues. Partially overlapping a frame,
as is the case of panel P 1 in Figure 3.1, is forbidden due to manufacturing limi-
tations. Also, due to the internal structure of panels, frames’ border and panels’
border must be separated by a minimum distance and thus panel P 2 is not valid.
Additionally, in oder to attach panels, the corners of each panel must match a
supporting area, thus, the panel P 3 is not valid. Figure 3.2 presents a valid
layout plan composed of six panels where all requirements are fulfilled.

Fig. 3. Ill-configured panels and layout-plan solution.

Frames mandatory overlapping is a complex requirement for the retrofit.

Essentially, a frame (e.g. window or door) should be completely overlapped by
one and only one panel. Figure 4 presents two cases in which panels have conflicts
w.r.t. frames and possible solutions for them.
 Open Packing for Facade-Layout Synthesis Under a General Purpose Solver 513

Fig. 4. Frame mandatory overlapping.

2.3 Assumptions
The following assumptions have been taken into account for the present work.
First, all supporting areas are strong enough in such a way that the problem
only deals with the placement of panel’s corners over supporting areas. In other
words, there is unlimited load-bearing capabilities in supporting areas and no
capabilities in the remaining of the facade area. Second, in order to compute
costs and thermal performance, we assume panel’s thickness to be a constant
and we consider only one insulation type. And third, we assume that costs and
thermal performance depends only and are in proportion with panel’s size.

2.4 Cost Function

In the industrial scenario the ranking is made w.r.t. cost and thermal perfor-
mance of the layout plan. The cost of a layout plan depends on the price of
isolated panels plus attached equipment. In this work however, we do not take
into account attached equipment as it depends on user’s preferences and not
in the layout plan. Experts have provided us a way to compute the cost of an
insulation envelope w.r.t. panels’ size. Formula 1 expresses this knowledge.


N
cf ac = (wi × hi ) + (α − wi − hi ) (1)
i=1

where wi and hi represent the width and height, respectively, of panel i,

and α is a factor provided by the manufacturer that depends on the particular
manufacturing process. As the formula express it, the term (α−wi −hi ) decreases
with the size of the panel. Thus manufacturing large panels is less costly, globally,
than manufacturing small ones.
Now, from a thermal performance point of view having large panels is good
because it minimizes the joints between panels. In fact, due to the thermal char-
acteristics of the retrofit, the less panels, the better, because most of the thermal
transfer is present in the intersection of two consecutive panels (junctions) plus
facade perimeter. Therefore, the performance of a layout plan depends on the
length of junctions between two consecutive panels. Computing the length of
514 A.F. Barco et al.

the junctions for a given envelope is straightforward, formula (2) expresses this
knowledge:


N
ttcf ac = wf ac + hf ac + (wi + hi ) (2)
i=1

where ttcf ac stands for thermal transfer coefficient for facade f ac, wf ac and
hf ac are the facade width and height, respectively, wi and hi represent the width
and height of panel i, respectively, and N is the number of panels composing
the envelope. According to formulas 1 and 2, using large panels is appropriated
to reduce costs and improve thermal insulation. As the larger the panels the
smaller the number of used panels, our optimization function lies on minimizing
the number of used panels.

3 Facade-Layout Synthesis as a CSP

In this section we introduce a constraint satisfaction model for the facade-
layout synthesis problem. Recall that a constraint satisfaction problem (CSP)
is described in terms of a set of variables V, a collection of potential values D
for each variable and a set of relations C over those variables, referred to as
constraints [14,15]. A CSP solution is an assignment of values for each variable
in such a way that every constraint in C is satisfied.
Let F denote the set of frames and S the set of supporting areas. Let oe.d and
le.d denote the origin and length, respectively, of a given entity e in the dimension
d, with d ∈ {1, 2}. For instance, of r.1 denotes the origin in the horizontal axis
and lf r.1 denotes the width of frame f r. Additionally, lbd and ubd respectively
denote the minimum and maximum size, in dimension d, for all rectangles.

3.1 Variables

At first glance, we know that decision variables will be related to rectangles

(i.e., panels). One of the major difficulties for tackling the problem using CP
is that the number of rectangles to be allocated is unfixed. Therefore, we first
heuristically bound this number from above. Let n denote an estimate of the
maximum number of rectangles to cover the facade. Given the lower  and upper 
1 ×f ac2
bounds for rectangles’ size and the facade size, we consider n = f ac lb1 ×lb2 .
We then create a set of n optional rectangle variables, each one referring to a
panel that may or may not belong to the solution. Each rectangles 0 ≤ p ≤ n is
described by its presence, origin and size attributes:

– bp ∈ {0, 1} indicates whether or not rectangles p is used in the solution.

– op.d ∈ [0, of ac.d ] is the origin of rectangles p in dimension d.
– lp.d ∈ [lbp.d , ubp.d ] is the length of rectangles p in dimension d.
 Open Packing for Facade-Layout Synthesis Under a General Purpose Solver 515

Note that, as a rectangle is already defined by an array of integer variables

(its coordinates and size), it is more natural to extend it with a fifth binary
variable representing its presence in the solution than introducing a set variable
to represent all present rectangles [22]. Domains for each variable depends on
the variable semantics. For instance, the origin op.1 is in the interval [0, lf ac.1 ].
Thus, for dealing with attaching points for rectangles, we use auxiliary variables
to represent op.2 + lp.2 and assign as domain those points over the facade where
supporting areas exists. Thus, these variables are instantiated using a domain
with holes, hence avoiding posting a constraint to deal with attaching points.
Finally, we introduce an objective variable z representing
 the number of
f ac1 ×f ac2
rectangles that are used in a solution. We then have ub1 ×ub2 ≤ z ≤ n.

3.2 Business Constraints

In order to configure the layout of a given facade we use constraints to ensure

relations over the variables representing entities. We shall begin the description
of four constraints that are related to the problem specifications.

(C1) Manufacturing and transportation limitations constrain panel’s size with a

give lower bound lb and upper bound ub in one or both dimensions:

∀p, 0 ≤ p < n, d ∈ {1, 2} lbd ≤ lp.d ≤ ubd

(C2) The entire facade area must be covered with panels:


n−1 
bp × lp.1 × lp.2 = lf ac.1 × lf ac.2
p=0

It is worth noticing that this constraint will lead to a very weak filtering,
because the domain of the l variables may be large and the constraint allows
panel overlaps. Therefore, it will be strengthened by the search heuristic.

(C3) Any two distinct panels that both belong to the solution do not overlap:

OpenDiffN(b, o, l)

This corresponds to the Open [22] variant of the DiffN [3] constraint, i.e.
a generalization of DiffN to handle optional rectangles.

(C4) Each frame over the facade must be completely overlapped by one and only
one panel. Additionally, frame borders and panel borders must be separated
by a minimum distance denoted by Δ:

∀f ∈ F, ∃p, 0 ≤ p < n, d ∈ {1, 2} |

op.d + Δ ≤ of.d ∧ of.d + lf.d ≤ op.d + lp.d + Δ

This constraint is implemented as a dedicated global constraint and will

be discussed further.
516 A.F. Barco et al.

3.3 Symmetry-Breaking Constraints

As we want to present to the end-user (e.g. an architect) a diverse set good if not
optimal solutions, we must avoid to enumerate symmetrical ones. The following
constraints break symmetries for rectangles in the solution as well as unused
rectangles.

(C5) Panels are ordered:

LexChainLessEq({{(1 − bp ), op,1 , op,2 }| 0 ≤ p < n})
This lexicographic constraint [21] ensures that priority is given to use the
first rectangles and that rectangles that are used in the solution are ordered
geometrically.
(C6) Unused panels are arbitrarily fixed:
∀p, 0 ≤ p < n, ∀d ∈ {1, 2}, bp = 0 ⇒ op,d = 0 ∧ lp,d = lbd
In order to avoid wasting time on unused rectangles, we may fix their origin
variables to the first possible attachment point as well as its size variables
to their minimum values. In the (C6) constraint network, it is assumed
that there is a valid supporting area at point (0,0).

4 Implementation
This section provides details on the model implementation. Basically, our solu-
tion follows the approach in [22] where a set variable is used to handle decision
variables that are potentially in the solution. In this work however, as rectan-
gle variables are already composed of several integer attributes, we found more
natural to use an extra binary variable per rectangle instead of a set variable.
Intuitively, an open constraint with boolean variables may be implemented fol-
lowing traditional filtering algorithms and may be enhanced by targeting the
structure of the problem.

4.1 An Open Constraint for Rectangle Non-Overlapping (C3)

As can be seen in the literature, the OpenDiffN constraint has already been
implemented (see No-Overlap with optional rectangles in Section 4.4.16 in [18]
for instance) but we consider it is necessary to provide a brief description of its
behavior. The filtering algorithm of the OpenDiffN checks whether two panels
that are part of the solution, i.e., whose bi is equal 1, do overlap, and proceeds to
domain filtering to prevent overlaps, as traditional DiffN propagators do. Con-
versely, if two panels do overlap in space, then domain filtering on the boolean
variables ensures that at least one of the two panels is not used. The overall fil-
tering is strengthened by a constructive disjunction algorithm that computes an
attaching point for the bottom left corner of each rectangle, that is valid (from
the packing point of view only) with respect to already fixed panels.
 Open Packing for Facade-Layout Synthesis Under a General Purpose Solver 517

4.2 A Constraint Dedicated to Frame Covering (C4)

The C4 constraint is propagated using a dedicated approach. The filtering algo-
rithm is pretty simple and works as follows: For every frame, two support panels
(i.e. panels the frame fits in) are computed. In case no support panel is found
then the solver fails. In case only a single panel is found, then a filtering proce-
dure is applied to enforce it to cover the frame. Finally, in case two panels have
been found, then no propagation is triggered because we do not know which
panel will be used to cover the facade.

4.3 Embedding Symmetry-Breaking

The lexicographic constraint has a strong influence on the model. It enables
to output different solutions, to reduce the size of the search tree but that is
not all: It is possible to speed up the other global constraints by taking that
information into account while filtering. For instance, any for loop seeking all
used rectangles (bi = 1) can stop as soon as one undetermined rectangle (bi =
{0, 1}) has been found because further rectangles are either undetermined or
unused. Thus, it is possible to exploit the problem structure to improve the
implemented constraints.
We will now see how to exploit the problem structure within search.

5 The Search Heuristic

The search heuristic is responsible of bounding rectangle’s decision variables
when propagation cannot infer more information. Our heuristic is described in
Algorithm 1. It is a constructive approach that considers each rectangle i one
by one and uses the following variable selection priority: bi , oi1 , oi2 , li1 and
finally li2 . We apply a traditional binary branching scheme over stated variables
[20]. It means that, instead of iterating over domain values, the heuristic assigns
a value to a variable and removes that value from the variable domain upon
backtracking.
The originality of our method is that some decisions cannot be negated:
Instruction d.once() in line 27 tells the solver not to try different values on
failure. For instance, if o1 = 1 and the node fails, it will not try to propagate
o1 = 1 and compute a new decision. Instead, it will backtrack once more (to the
decision associated with the size of the previous rectangle).
The heuristic implements the following key design choices:
1. We set the b variables to 1 in order to arrive rapidly to solutions.
2. The position variables o1 and o2 are fixed to their lower bounds in order to
leave no uncovered places between the considered rectangle and previously
placed rectangles. In short, the real decision variables are b, l1 and l2 . But
o1 and o2 should be set in a deterministic way without backtracking. As
rectangles are ordered, trying a larger value would lead to a hole on the
facade, which is forbidden. Note that this is only possible because the filtering
518 A.F. Barco et al.

Algorithm 1. Dedicated Search Heuristic

1 def Function getBranchingDecision:
2 int r ← −1; // compute the first unfixed rectangle;
3 for i ← 0 to n do
4 if |dom(bi )| + |dom(oi1 )| + |dom(oi2 )| + |dom(li1 )| + |dom(li2 )| > 5 then
5 r ← i; break;

6 if r == −1 then
7 return null; // all rectangles are fixed (a solution has been found)
8 // Find the next variable-value assignment to perform
9 IntegerVariable var, int val;
10 if |dom(bi )| > 1 then
11 var ← bi ;
12 val ← 1;
13 else if |dom(oi1 )| > 1 then
14 var← oi1 ;
15 val ← dom(oi1 ).lb;
16 else if |dom(oi2 )| > 1 then
17 var← oi2 ;
18 val ← dom(oi2 ).lb;
19 else if |dom(li1 )| > 1 then
20 var← li1 ;
21 val ← dom(li1 ).ub;
22 else
23 var← li2 ;
24 val ← dom(li2 ).ub;
25 Branch d = new Branch(var, val);
26 if var == oi1 ∨ var == oi2 then
27 d.once(); // prevents the solver from trying different values upon
backtracking
28 return d;

is strong enough: The lower bound is indeed a valid support from the packing
point of view.
3. The size variables l1 and l2 are set to their upper bounds in order to try
rectangles as large as possible and thus cover the largest possible area. This
enables to get a first solution that is very close to the optimal value.

Overall, the search mechanism combines two different principles. First, it is

based on a greedy constructive approach that is efficient but limited. Second, it
uses a customized (some decisions cannot be negated) backtracking algorithm
to explore alternatives.
 Open Packing for Facade-Layout Synthesis Under a General Purpose Solver 519

6 Evaluation
In this section we empirically evaluate our model, which has been implemented in
Java, with the Choco solver [17]. We consider a dataset of 20 instances, generated
from realistic specifications. The total area of the facade ranges from 104 to 106
to test the scalability of the approach. Panels are generated using a lower bound
of 20 (pixels) and an upper bound of 150 (pixels), for both width and height.

6.1 A Two-Step Approach

In a first experiment we want to evaluate whether or not the maximum number of
used panels is a good approximation of the optimum. Figure 5 presents the number
of used panels and the number of optional panels for every instance. The maximum
number of panels, which represents the worst case scenario where panels’ size lower
bounds are used, is never reached. We can see that the first solution is actually very
far from the optimum. Further, this maximum number is an upper bound far to
high: For a facade of size 2300 × 575, the solver handles 3220 optional panels to
compute a first solution that uses only 96 panels. This means that we create too
many useless variables that will slow down the solving process. Therefore, we set
up a 2-step approach: First a solution is computed using our model. Second, we
create a new model with the value of the previous solution as maximum number
of panels. Then, we enumerate all optimal solutions.

6.2 Impact of Symmetry Breaking

In a second experiment, we measure the impact of adding symmetry-breaking con-
straints. More precisely we compare the time to find a first solution (Figure 6.a)
and the number of computed solutions (Figure 6.b) with and without symmetry-
breaking constraints. Because of the huge amount of solutions, we use a time limit

Used panels versus facade surface

3500
Number of used panels
3000 Maximum number of optional panels

2500
Number of panels

2000

1500

1000

500

0
0⋅100 2⋅105 4⋅105 6⋅105 8⋅105 1⋅106 1⋅106 1⋅106
Facade surface

Fig. 5. Maximum number of optional panels and number of used panels in the first
solution, for every instance.
520 A.F. Barco et al.

of 60 seconds. As we can see it on Figure 6, symmetry-breaking constraints speed

up the search. Moreover, it enables to skip thousands of solutions that are identi-
cal for the end user. This is even more important than saving computational time.
Note that it seems that the number of solutions decreases when the facade area
increases: this is due to the time limit. As the problem gets bigger, the solving
process gets slower and enumerates less solutions in a given time.

Time of first solution versus facade surface Number of solutions versus facade surface
20 2500
Without symmetry breaking Without symmetry breaking
18 With symmetry breaking With symmetry breaking
16 2000

Number of Solutions
14
Times in seconds

12 1500
10
8 1000
6
4 500
2
0 0
0⋅100 2⋅105 4⋅105 6⋅105 8⋅105 1⋅106 1⋅106 1⋅106 0⋅100 2⋅105 4⋅105 6⋅105 8⋅105 1⋅106 1⋅106 1⋅106
Facade surface Facade surface

(a) (b)

Fig. 6. Time to reach a first solution and number of computed solutions, with a 60
seconds time limit.

6.3 Search Comparison

In regard to different search heuristics, we have not found any well-suited for
addressing the problem. Actually, well-known black-box search strategies such
as domOverWDeg, impact-based search or activity-based search, do not per-
form well given the problem particularities. These heuristics are designed to
solve problems in a blind way, when we have no expert knowledge of the prob-
lem. In our case, we mix very different kind of variables (booleans, positions,
sizes) that we are able to group by rectangles and order. Introducing random-
ness on either the variable selection or the value selection may be disastrous. In
particular, using arbitrary values for o1 and o2 makes a huge amount of possi-
bilities for uncovered places. Nonetheless, in order to validate the relevance of
our dedicated heuristic, we have tested 16 predefined heuristics from Choco on
a small facade (400 × 100). We present the results for those ones that threw
at least one solution. These strategies are: domOverWDeg which selects the vari-
able with the smallest ratio |d(x)|
w(x) , where |d(x)| denotes the domain size of a
variable x and w(x) its weighted degree; lexico_LB which chooses the first non-
instantiated variable, and assigns it to its lower bound; lexico_Split which
chooses the first non-instantiated variable, and removes the second half of its
domain; maxReg_LB which chooses the non-instantiated variable with the largest
difference between the two smallest values in its domain, and assigns it to its
lower bound; minDom_LB which chooses the first non-instantiated variable with
the smallest domain size, and assigns it to its lower bound and; minDom_MidValue
which chooses the first non-instantiated variable with the smallest domain size,
 Open Packing for Facade-Layout Synthesis Under a General Purpose Solver 521

and assigns it to the value closest to the middle of its domain. The last entry is
our own search heuristic. Recall that variables are ordered as b, o1 , o2 , l1 and l2 .

Table 1. Heuristic comparison on a 400 × 100 (pixels) facade.

Strategy First solution time (s) Total time (s) #nodes #solutions
domOverWDeg 18.77 19.94 1412897 66
lexico LB 0.03 0.22 2380 66
lexico Split 0.03 0.16 441 66
maxReg LB 0.03 0.22 2380 66
minDom LB 0.74 19.96 1411183 66
minDom MidValue 43.43 47.32 4755206 66
dedicated 0.03 0.85 10978 66

Table 2. Heuristic comparison on a 400 × 200 (pixels) facade with a 3−minute time
limit
Strategy First solution time (s) #nodes #solutions
domOverWDeg - 7286594 0
lexico LB - 5772501 0
lexico Split - 4966920 0
maxReg LB - 5490088 0
minDom LB - 11961712 0
minDom MidValue - 11157755 0
dedicated 0.039 3499527 726

Tables 1 and 2 respectively provide the results for a 400 × 100 and 400 × 200
instance. Although some predefined heuristics have a good performance on the
first (small) instance, none of them scales. In fact, no predefined search heuristic
finds a solution for a facade with size 400 × 200 in reasonable computational
time whereas our dedicated heuristic already finds 726 different solutions in
180 seconds. Our heuristic clearly outperforms the others. More importantly, it
enables to always output a solution fast, which is critical for the user.

7 Conclusions
This paper presents a solution to the facade-layout synthesis problem treated as
a two-dimensional packing problem. Although there exists a great body of lit-
erature on two-dimensional packing, our industrial scenario includes three char-
acteristics never considered simultaneously: Its deals with the allocation of an
unfixed number of rectangular panels that must not overlap, frames over the
facade must be overlapped by one and only one panel, and facades have specific
areas to attach panels. Thus, as far as we know, no support system nor design
system is well-suited for addressing such particularities.
We have used constraint satisfaction and constraint programming as under-
lying model and solving technique. Constraint programming technology is well-
suited for addressing this industrial problem because on the one hand, an objec-
tive function is identified, namely minimize number of panels, and on the other
522 A.F. Barco et al.

hand, the building of a prototype using an open constraint programming envi-

ronment is much easier, thanks to all the pre-defined constraints, search and
provided abstractions. The modeling decisions was made by a four-person team
whereas the development was carried out by a two-person team with knowl-
edge on open constraint programming environments (e.g. Choco, Gecode, finite
domain module of Mozart-Oz). The development was done in one month of
work.
Considering that the number of panels is not know in advance we have used
a variant of the DiffN constraint to handle optional rectangles by means of
boolean variables. We have implemented a constraint for the mandatory frame
overlapping and a dedicated search heuristic that takes advantage of the problem
structure and thus is able to enumerate optimal solutions w.r.t. the number of
used panels. Our proposed solutions have been implemented in the Choco solver
and demonstrate the validity of our method. In particular, our model takes
the benefit of both a greedy and a tree-search approach in order to output
several good solutions very fast, which was the most critical requirement from
the client. This highlights the importance of the great flexibility of Constraint
Programming.
Future directions of this work include the definition of different heuristics
that take into account aesthetic aspects, the inclusion of weight constraints over
panels and supporting areas as well as variables for panel’s thickness and isolation
type that have an impact on the layout solution. In addition, the packing of
different (convex) shapes should be stressed.

References
1. Barták, R.: Dynamic global constraints in backtracking based environments.
Annals of Operations Research 118(1–4), 101–119 (2003)
2. Beldiceanu, N., Carlsson, M., Poder, E., Sadek, R., Truchet, C.: A generic geomet-
rical constraint kernel in space and time for handling polymorphic k -dimensional
objects. In: Bessière, C. (ed.) CP 2007. LNCS, vol. 4741, pp. 180–194. Springer,
Heidelberg (2007)
3. Beldiceanu, N., Carlsson, M., Demassey, S., Poder, E.: New filtering for the cumu-
lative constraint in the context of non-overlapping rectangles. Annals of Operations
Research 184(1), 27–50 (2011)
4. The Energy Conservation Center: Energy Conservation Handbook. The Energy
Conservation Center, Japan (2011)
5. Charman, P.: Solving space planning problems using constraint technology (1993)
6. U.S. Green Building Council: New Construction Reference Guide (2013)
7. Falcon, M., Fontanili, F.: Process modelling of industrialized thermal renovation
of apartment buildings. In: eWork and eBusiness in Architecture, Engineering and
Construction, pp. 363–368 (2010)
8. Imahori, S., Yagiura, M., Nagamochi, H.: Practical algorithms for two-dimensional
packing. Chapter 36 of Handbook of Approximation Algorithms and Metaheuris-
tics (Chapman & Hall/Crc Computer & Information Science Series) (2007)
9. Jelle, B.P.: Traditional, state-of-the-art and future thermal building insulation
materials and solutions - properties, requirements and possibilities. Energy and
Buildings 43(10), 2549–2563 (2011)
 Open Packing for Facade-Layout Synthesis Under a General Purpose Solver 523

10. Juan, Y.-K., Gao, P., Wang, J.: A hybrid decision support system for sustain-
able office building renovation and energy performance improvement. Energy and
Buildings 42(3), 290–297 (2010)
11. Laborie, P.: IBM ILOG CP optimizer for detailed scheduling illustrated on three
problems. In: van Hoeve, W.-J., Hooker, J.N. (eds.) CPAIOR 2009. LNCS, vol.
5547, pp. 148–162. Springer, Heidelberg (2009)
12. Lee, K.J., Hyun, W.K., Lee, J.K., Kim, T.H.: Case- and constraint-based project
planning for apartment construction. AI Magazine 19(1), 13–24 (1998)
13. Robin, S.: Liggett. Automated facilities layout: past, present and future. Automa-
tion in Construction 9(2), 197–215 (2000)
14. Montanari, U.: Networks of constraints: Fundamental properties and applications
to picture processing. Information Sciences 7, 95–132 (1974)
15. Olarte, C., Rueda, C., Valencia, F.D.: Models and emerging trends of concurrent
constraint programming. Constraints 18(4), 535–578 (2013)
16. Pérez-Lombard, L., Ortiz, J., Pout, C.: A review on buildings energy consumption
information. Energy and Buildings 40(3), 394–398 (2008)
17. Prud’homme, C., Fages, J.G.: An introduction to Choco 3.0, an open source java
constraint programming library. In: International workshop on CP Solvers: Mod-
eling, Applications, Integration, and Standardization, Uppsala, Sweden (2013)
18. Schulte, C., Tack, G., Lagerkvist, M.Z.: Modeling and programming with Gecode
(2010)
19. Schutt, A., Feydy, T., Stuckey, P.J.: Scheduling optional tasks with explanation.
In: Schulte, C. (ed.) CP 2013. LNCS, vol. 8124, pp. 628–644. Springer, Heidelberg
(2013)
20. Smith, B.M.: Modelling for constraint programming (2005)
21. van Hoeve, W.-J., Hooker, J.N. (eds.): CPAIOR 2009. LNCS, vol. 5547. Springer,
Heidelberg (2009)
22. van Hoeve, W.-J., Régin, J.-C.: Open constraints in a closed world. In: Beck, J.C.,
Smith, B.M. (eds.) CPAIOR 2006. LNCS, vol. 3990, pp. 244–257. Springer, Heidel-
berg (2006)
23. Vareilles, E., Barco Santa, A.F., Falcon, M., Aldanondo, M., Gaborit, P.: Config-
uration of high performance apartment buildings renovation: a constraint based
approach. In: 2013 IEEE International Conference on Industrial Engineering and
Engineering Management (IEEM), pp. 684–688, December 2013
24. Zawidzki, M., Tateyama, K., Nishikawa, I.: The constraints satisfaction prob-
lem approach in the design of an architectural functional layout. Engineering
Optimization 43(9), 943–966 (2011)
 Power Capping in High Performance
Computing Systems

Andrea Borghesi1,2(B) , Francesca Collina1,2 , Michele Lombardi1,2 ,

Michela Milano1,2 , and Luca Benini1,2
1
DISI, University of Bologna, Bologna, Italy
2
DEI, University of Bologna, Bologna, Italy
{andrea.borghesi3,michele.lombardi2,michela.milano,luca.benini}@unibo.it,
francesca.collina@gmail.com

Abstract. Power consumption is a key factor in modern ICT infrastruc-

ture, especially in the expanding world of High Performance Computing,
Cloud Computing and Big Data. Such consumption is bound to become
an even greater issue as supercomputers are envisioned to enter the Exas-
cale by 2020, granted that they obtain an order of magnitude energy
efficiency gain. An important component in many strategies devised to
decrease energy usage is “power capping”, i.e., the possibility to con-
strain the system power consumption within certain power budget. In
this paper we propose two novel approaches for power capped workload
dispatching and we demonstrate them on a real-life high-performance
machine: the Eurora supercomputer hosted at CINECA computing cen-
ter in Bologna. Power capping is a feature not included in the commercial
Portable Batch System (PBS) dispatcher currently in use on Eurora. The
first method is based on a heuristic technique while the second one relies
on a hybrid strategy which combines a CP and a heuristic approach.
Both systems are evaluated and compared on simulated job traces.

1 Introduction

Supercomputer peak performance is expected to reach the ExaFLOP level in

2018-2020 [14][15], however energy efficiency is a key challenge to be addressed
to reach this milestone. Today’s most powerful Supercomputer is Tianhe-2 which
reaches 33.2 PetaFlops with 17.8 MWatts of power dissipation [13]. Exascale
supercomputers built upon today’s technology would led to an unsustainable
power demand (hundreds of MWatts) while according to [9] an acceptable range
for an Exascale supercomputer is 20MWatts; for this goal, current supercom-
puter systems must obtain significantly higher energy efficiency, with a limit
of 50GFlops/W. Today’s most efficient supercomputer achieves 5.2 GFlops/W,
thus we still need to close an order of magnitude gap to fulfill the Exascale
requirements.
Almost all the power consumed by HPC systems is converted into heat. In
addition to the power strictly needed for the computation - which measures
only the computational efficiency - the cooling infrastructure must be taken into

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 524–540, 2015.
DOI: 10.1007/978-3-319-23219-5 37
 Power Capping in High Performance Computing Systems 525

account, with its additional power consumption. The extra infrastructure needed
for cooling down the HPC systems has been proved to be a decisively limiting
factor for the energy performance [3]; a common approach taken to address this
problem is the shift from air cooling to the more efficient liquid cooling [10].
Hardware heterogeneity as well as dynamic power management have started
to be investigated to reduce the energy consumption [16][2]. These low-power
techniques have been derived from the embedded system domain where they
have proven their effectiveness [1]. However, a supercomputer is different from
a mobile handset or a desktop machine. It has a different scale, it cannot be
decoupled by the cooling infrastructures and its usage mode is peculiar: it is
composed by a set of scientific computing applications which run on different
datasets with a predicted end time [7]. Finally supercomputers are expensive (6
orders of magnitude more than an embedded device [19]) making it impossible
for researchers to have them on their desk. These features have limited the
development of ad-hoc power management solutions.
Current supercomputers cooling infrastructures are designed to withstand
power consumption at the peak performance point. However, the typical
supercomputer workload is far below the 100% resource utilization and also
the jobs submitted by different users are subject to different computational
requirements[29]. Hence, cooling infrastructures are often over-designed. To
reduce overheads induced by cooling over-provisioning several works suggest to
optimize job dispatching (resource allocation plus scheduling) exploiting non-
uniformity in thermal and power evolutions [8][20][21][23]. Currently most of
these works are based on simulations and model assumptions which unfortu-
nately are not mature enough to be implemented on HPC systems in production,
yet.
With the goal of increasing the energy efficiency, modern supercomputers
adopt complex and hybrid cooling solutions which try to limit the active cool-
ing (chiller, air conditioner) by allowing direct heat exchange with the ambient
(Free-cooling). Authors in [12] show for the 2013’s top GEEN500 supercom-
puter that the cooling costs increase four times when the ambient temperature
moves from 10 C to 40 C. Moreover the authors show that for a given ambient
temperature there is a well-defined maximum power budget which guarantees
efficient cooling. With an ambient temperature of 10 C the power budget which
maximizes the efficiency is 45KWatt while at 40 C is 15KWatt. Unfortunately,
today’s ITs infrastructure can control its power consumption only reducing the
power and performance of each computing node. This approach is not suitable
for HPC system where users require to execute their application in a portion
of the machine with guaranteed performance. A solution commonly adopted in
HPC systems is power capping[25][17][15][27][22], which means forcing a super-
computer not to consume more than a certain amount of power at any given
time. In this paper we study a technique to achieve a power capping by acting
on the number of job entering the system.
526 A. Borghesi et al.

We propose two different methods to enforce power capping constraints on a

real supercomputer: 1) a Priority Rules Based algorithm and 2) a novel hybrid
approach which combines a CP and a heuristic technique.

2 System Description and Motivations for Using CP

The Eurora Supercomputer: As described in [6] Eurora has a heterogeneous
architecture based on nodes (blades). The system has 64 nodes, each with 2
octa-core CPUs and 2 expansion cards configured to host an accelerator module:
currently, 32 nodes host 2 powerful NVidia GPUs, while the remaining ones are
equipped with 2 Intel MIC accelerators. Every node has 16GB of installed RAM
memory. A few nodes (external to the rack) allow the interaction between Eurora
and the outside world, in particular a login node connects Eurora to the users
and runs the job dispatcher (PBS). A key element of the energy efficiency of the
supercomputer is a hot liquid cooling system, i.e. the water inside the system
can reach up to 50◦ C. This obviously allows to save a lot of energy, since no
power is used to actively cool down the water; furthermore the heat in excess
can be reused as an energy source for other activities.

The PBS Dispatcher: The tool currently used to manage the workload on Eurora
system is PBS [28] (Portable Batch System), a proprietary job scheduler by
Altair PBS Works which has the task of allocating computational activities,
i.e. batch jobs, among available computing resources. The main components of
PBS are a server (which manages the jobs) and several daemons running on the
execution hosts (i.e. the 64 nodes of Eurora), which track the resource usage and
answer to polling requests about the host state issued by the server component.
Jobs are submitted by the users into one of multiple queues, each one char-
acterized by different access requirements and by a different estimated waiting
time. Users submit their jobs by specifying 1) the number of required nodes;
2) the number of required cores per node; 3) the number of required GPUs and
MICs per node (never both of them at the same time); 4) the amount of required
memory per node; 5) the maximum execution time. All processes that exceed
their maximum execution time are killed. The main available queues on the
Eurora system are called debug, parallel, and longpar. PBS periodically selects
a job to be executed considering the current state of the nodes - trying to find
enough available resources to start the job. If there are not enough available
resources the job is returned to its queue and PBS considers the following can-
didate. The choices are guided by priority values and hard-coded constraints
defined by the Eurora administrators with the aim to have a good machine uti-
lization and small waiting times. For more detailed information regarding the
system see [7].

Why CP? In its current state, the PBS system works mostly as an on-line
heuristic, incurring the risk to make poor resource assignments due to the lack
of an overall plan; furthermore, it does not include a power capping feature yet.
 Power Capping in High Performance Computing Systems 527

Other than that, as we shown in our previous work the lack of an overall plan may
lead to poor resource assignments. The task of obtaining a dispatching plan on
Eurora can be naturally framed as a resource allocation and scheduling problem,
for which CP has a long track of success stories. Nevertheless since the problem
is very complex and we are constrained by very strict time limits (due to the
real-time requirements of a supercomputer dispatcher) it is simply not possible
to always find a optimal solutions. Therefore, we used CP in combination with
multiple heuristic approaches in order to quickly find the resource allocation and
schedule needed.

3 Problem Definition
We give now a more detailed definition of the problem. Each job i enters the sys-
tem at a certain arrival time qi , by being submitted to a specific queue (depend-
ing on the user choices and on the job characteristics). By analyzing existing
execution traces coming from PBS, we have determined an estimated waiting
time for each queue, which applies to each job it contains: we refer to this value
as ewti .
When submitting the job, the user has to specify several pieces of information,
including the maximum allowed execution time Di , the maximum number of
nodes to be used rni , and the required resources (cores, memory, GPUs, MICs).
By convention, the PBS system considers each job as if it was divided into a
set of exactly rni identical “job units”, to be mapped each on a single node.
Formally, let R be a set of indexes corresponding to the resource types (cores,
memory, GPUs, MICs), and let the capacity of a node k ∈ K for resource r ∈ R
be denoted as capk,r . Let rqi,r be the requirement of a unit of job i for resource r.
The dispatching problem at time τ requires to assign a start time si ≥ τ to each
waiting job i and a node to each of its units. All the resource and power capacity
limits should be respected, taking into account the presence of jobs already in
execution. Once the problem is solved, only the jobs having si = τ are actually
dispatched. The single activities have no deadline or release time (i.e. they do
not have to end within or start after a certain date), nor the global makespan is
constrained by any upper bound.
Along with the aforementioned resources in this problem we introduce an
additional finite resource, the power. This will allow us to model and enforce the
power capping constraint. The power capping level, i.e. the maximal amount of
power consumed by the system at any given time, is specified by the user and
represents the capacity of the fake power resource; the sum of all the power con-
sumed by the running job plus the sum of power values related to idle resources
must never exceed the given limit throughout the execution of the whole sched-
ule. Another important thing to notice is that the power “resource” has not a
linear behaviour in terms of the computational workload: the first activation of
a core in a node causes greater power consumptions for the remaining cores in
that node.
The goal is to reduce the waiting times, as a measure of the Quality of Service
guaranteed to the supercomputer users.
528 A. Borghesi et al.

4 The PRB Approach

First, we implemented a dispatcher which belongs to a class of scheduling tech-
niques known in the literature as Priority Rules Based scheduling (PRB)[18].
The main idea underlying this approach is to order a set of tasks which need to
be scheduled, constructing the ordered list by assigning priority for each task.
Tasks are selected in the order of their priorities and each selected task is assigned
to a node; even the resource are ordered and the ones with higher priority are
preferred - if available. This is an heuristic technique and it is obviously not able
to guarantee an optimal solution but has the great advantage of being extremely
fast.
The jobs are ordered w.r.t to their expected wait times, with the “job
demand” (job requirements multiplied by the job estimated duration) used to
break ties. Therefore, jobs which are expected to wait less have higher priority,
subsequently jobs with smaller requirements and shorter durations are preferred
over heavier and longer ones. The mapper selects one job at time and maps it on
a available node with sufficient resources. The nodes are ordered using two crite-
ria: 1) at first, more energy efficient nodes are preferred (i.e. cores that operate at
higher frequencies also consume more power) 2) in case of ties, we favour nodes
based on their current load (nodes with fewer free resources are preferred1 ).
The PRB algorithm proceeds iteratively trying to dispatch all the activities
that need to be run and terminates only when there are no more jobs to dispatch.
We suppose that at time t = 0 all the resources are fully available, therefore the
PRB algorithm starts by simply trying to fit as many activities as possible on
the machine, respecting all resource constraints and considering both jobs and
nodes in the order defined by the priority rules. Jobs that cannot start at time
0 are scheduled at the first available time slot. At each time-event the algorithm
will try to allocate and start as many waiting jobs as possible and it will keep
postponing those whose requirements cannot be met yet.
The algorithm considers and enforces constraints on all the resources of the
system, including power.
Algorithm 1 shows the pseudo code of the PRB algorithm. Lines 1-6 initial-
ize the algorithm; J is the set of activities to be scheduled and R is the set of
nodes (ordered with the orderByRules() function which encapsulates the prior-
ity rules). Then the algorithm proceeds while there are still jobs to be scheduled;
at every iteration we try to start as many jobs as possible (line 8). Each job unit
is considered (line 10) and the availability of resources on every node is taken into
account; the function checkAvailability(rni , r) (line 12) returns true if there are
enough available resources on node r to map unit rni . If it is possible the unit is
then mapped and the system usage is updated (updateU sages(rni , R), line 13),
vice versa we register that at least a job unit could not be mapped (line 16). If all
the units of a job have been mapped (line 17-21), then the job can actually start
and is removed from the pool of jobs that still need to be scheduled. Conversely,

1
This criterion should decrease the fragmentation of the system, trying to fit as many
job as possible on the same node.
 Power Capping in High Performance Computing Systems 529

Algorithm 1. PRB algorithm

1 time ← 0;
2 startT imes ←− ∅;
3 endT imes ←− ∅;
4 runningJobs ←− ∅;
5 orderByRules(R);
6 orderByRules(J);
7 while J = ∅ do
8 for j ∈ J do
9 canBeM apped ← true;
10 for rni ∈ j do
11 for r ∈ R do
12 if checkAvailability(rni , r) then
13 updateU sages(rni , R);
14 break;
15 else
16 canBeM apped ← f alse

17 if canBeM apped = true then

18 J ←− J − {j};
19 runningJobs = runningJobs ∪ {j};
20 startT imes(j) ← time;
21 endT imes(j) ← time + duration(j);
22 else
23 undoU pdates(j, R);

24 orderByRules(R);
25 orderByRules(J);
26 time ← min(endT imes);

if the job cannot start we must undo the possible changes made to the system
usage we made in advance (updateU sages(rni , R), line 23). Finally, after having
dealt with all schedulable jobs we reorder the activities and nodes (the activity
pool is changed and the nodes order depends on the usages), lines 24-25, and
compute the closest time point where some used resource becomes free, following
time-event, i.e. the minimal end time of the running activities (line 26).
It is important to note that since in this problem we have no deadline on the
single activities nor a constraint on the global makespan, the PRB algorithm will
always find a feasible solution, for example delaying the least important jobs until
enough resources become available due to the completion of previously started
tasks.

5 Hybrid Approach
As previously discussed in [7] the task of obtaining a proactive dispatching plan
on Eurora can be naturally framed as a resource allocation and scheduling
530 A. Borghesi et al.

problem. Constraint Programming (CP) techniques have shown great success

when dealing with this kind of problem, thanks to the expressive and flexible
language used to model the problem and powerful algorithms to quickly find
good quality solutions[5]. In this Section we describe the mixed approach we
used to solve the resource allocation and scheduling problem under power cap-
ping constraints.
The key idea of our method is to decompose the allocation and scheduling
problem in two stages: 1) obtain a schedule using a relaxed CP model of the
problem 2) find a feasible mapping using a heuristic technique. Since we used
a relaxed model in the first stage, the schedule obtained may contain some
inconsistencies; these are fixed during the mapping phase, thus we eventually
obtain a feasible solution, i.e. a feasible allocation and schedule for all the jobs.
This two stages are repeated n times, where n has been empirically chosen after
an exploratory analysis, keeping in mind the trade-off between the quality of
the solution and the computational time required to find one. To make this
interaction effective, we devised a feedback mechanism between the second and
the first stage, i.e. from the infeasibilities found during the allocation phase we
learn new constraints that will guide the search of new scheduling solutions at
following iterations.
In this work we implemented the power capping requirements as an addi-
tional constraint: on top of the finite resources available in the system such as
CPUs or memory, we treat the power as an additional resource with its own
capacity (i.e. the user-specified power cap), which we cannot “over-consume” at
any moment[11]. In this way, the power used in the active nodes (i.e. those on
which a job is running) summed to the power consumed by the idle nodes will
never exceed the given threshold.
The next sections will describe in more detail the two stages of the decom-
posed approach.

5.1 Scheduling Problem

The scheduling problem consists in deciding the start times of a set of activities
i ∈ I satisfying the finite resource constraints and the power capping constraint.
Since all the job-units belonging to the same jobs must start at the same time,
during the scheduling phase we can overlook the different units since we need
only the start time for each job. Whereas in the actual problem the resources
are split among several nodes, the relaxed version we use in our two-stages
approach considers all the resources of the same type (cores, memory, GPUs,
MICs) as a pool of resource with a capacity CapTr which is the sum of all the
single resource capacities, Capr = k∈K capk,r ∀r ∈ R. As mentioned before
T

the power is considered as an additional resource type of the system, so we have


a set of indexes R corresponding to the resource types (cores, memory, GPUs,
MICs plus the power); the overall capacity CapTpower is equal to the user-defined
power cap.
The CP model employs other relaxations: 1) only the power required by
running jobs (active power) is considered, i.e. we use a lower bound of the total
 Power Capping in High Performance Computing Systems 531

power consumed in the system, 2) we assume that the jobs always run on the
nodes which require less power and 3) we overlook the non-linear behaviour
of the power consumption. These relaxations may produce infeasible solutions,
since a feasible schedule must take into account that the resource are actually
split among heterogeneous nodes and consider also the idle nodes for the power
capping. The following stage of our method will naturally fix these possible
infeasibilities during the allocation phase.
We define the scheduling model using Conditional Intervals Variables
(CVI)[24]. A CVI τ represents an interval of time: the start of the interval is
referred to as s(τ ) and its end as e(τ ); the duration is d(τ ). The interval may or
may not be present, depending on the value of its existence expression x(τ ) (if
not present it does not affect the model). CVIs can be subject to several different
constraints, among them the cumulative constraint[4] to model finite capacity
resources. 
∀r ∈ R cumulative(τ, reqr , CapTr ) (1)
where τ is the vector with all the interval vars, where reqr are the job require-
ments for resource r - using past execution traces we learned a simple model to
estimate the power consumed by a job based on its requirements. As mentioned
in section 3 this model is not linear. The cumulative constraints in 1 enforce that
at any given time, the sum of all job requirements will not exceed the available
capacity (for every resource type).
With this model it would be easy to define several different goals, depending
on the metric we optimize. Currently we use as objective function the weighted
queue time, i.e. we want to minimize the sum of the waiting times of all the jobs,
weighted on estimated waiting time for each job (greater weights to job which
should not wait long):
 max ewti
min (s(τi ) − qi ) (2)
ewti
i∈I

To solve the scheduling model we implemented a custom search strategy

derived from the Schedule Or Postpone strategy [26]. The criteria used to select
an activity among all the available ones at each decision node follows the priority
rules used in the heuristic algorithm, thus preferring jobs that can start first and
whose resource demand is lower. This strategy proved to be very effective and
able to rapidly find good solutions w.r.t. the objective function we are consider-
ing in this problem. With different goals we should change the search strategy
accordingly (as with the priority rules).

5.2 Allocation Problem

The allocation problem consists in mapping each job unit on a node. Further-
more, in our approach the allocation stage is also in charge of fixing the infeasibil-
ities generated at the previous stage. In order to solve this problem we developed
an algorithm which falls in the PRB category. Typical PRB schedulers decide
532 A. Borghesi et al.

both mapping and start times, whereas in our hybrid approach we need only to
allocate the jobs.
The behaviour of this algorithm (also referred as mapper ) is very close to
the PRB one described in Section 4 and in particular the rules used to order
jobs and resources are identical. The key difference is that now we already know
the start and end times of the activities (at least the possible ones, they may
change if any infeasibility is detected). This algorithm proceeds by time-step: at
each time event t it considers only the jobs that should start at time t according
to the relaxed CP model described previously, while the simple PRB algorithm
considers at each time-event all the activities that still need to be scheduled.
During this phase the power is also taken into account, again seen as a finite
resource with capacity defined by the power cap; here we consider both active
and idle nodes powers. If the job can be mapped somewhere in the system the
start time t from the previous stage is kept, otherwise - if there are not enough
resources available to satisfy the requirements - the previously computed start
time is discarded and the job will become eligible to be scheduled at the next
 
time-step t . At the next time event t all the jobs that should start are consid-
ered, plus the jobs that possibly have been postponed due to scarce resources at
the previous time-step. Through this postponing we are fixing the infeasibilities
inherited from the relaxed CP model.
Again, since in this problem we have no constraints on the total duration of
a schedule, it is always possible to delay a job until the system will have enough
available resources to run it, thus this method is guaranteed to find a feasible
solution.

5.3 Interaction between the Stages

We designed a basic interaction mechanism between the two stages with the
goal to find better quality solutions. The main idea is to exploit the information
regarding the infeasibilities found during the allocation phase to lead the search
of new solutions for the relaxed scheduling problems. In particular whenever we
detect a resource over-usage at time tau which requires a job to be postponed
during the second stage of our model we know that the set of job running at
time τ is a Conflict Set (not minimal), i.e. not all activities in the set can run
concurrently. A possible solution for a conflict set is for example to introduce
precedence relationships among activities (thus eliminating jobs from the conflict
set) until the infeasibility is resolved.
In our approach we use the conflict set detected in the mapping phase to
generate a new set of constraints which impose that not all jobs in the set will
run at the same time. We then introduce in the CP model a fake cumulative
constraint for each conflict set. The jobs included in such cumulative constraint
are those included in the conflict set, each of them with a “resource” demand
of one; the capacity not to be exceeded is given by the conflict set size minus
one. These cumulative constraints enforce that the jobs involved will not run at
the same time. This mechanism does not guarantee yet that the new solution
found by the CP scheduler will be feasible since the conflict sets we detect are
 Power Capping in High Performance Computing Systems 533

not minimal, nevertheless it provides a way to help the CP model to produce

solutions which will require less “fixing” by the mapper.
In conjunction with the additional cumulative constraint at each iteration
we also cast a further constraint on the objective variable in order to force the
new solution to improve in comparison to the previous one.

6 Added Value of CP

The dispatcher we realized is currently a prototype: it will eventually be deployed

on the Eurora supercomputer, but this requires still considerable development
and research effort (at the same time a previous version without power capping
of this model[7] has already successfully been implemented). At this stage we are
interested in investigating the kind of impact that introducing a power capping
feature may have on the dispatcher behaviour and performance.
The dispatcher we realized can work in two different modes: off-line, i.e. the
resource allocation and the schedule are computed before the actual execution,
and on-line, i.e. allocation and scheduling decisions are taken at run-time, upon
the arrival of new tasks in the system. Clearly the actual implementation on the
supercomputer would require the on-line strategy since the workload is submitted
by users and not statically decided a priori. At the same time we decided to use
the off-line strategy to perform the experiments described in this paper since we
wanted to test our approaches with reproducible conditions. We also needed our
techniques to be stable in order to implement them on a production system and
the off-line strategy allows us to better verify that.
The proposed methods were implemented using or-tools2 , Google’s software
suite for combinatorial optimization. We performed an evaluation of all our
approaches on PBS execution traces collected from Eurora in a timespan of sev-
eral months. From the whole set of traces we extracted different batches of jobs
submitted at various times and we used them to generate several job instances
of different size, i.e. the number of jobs per instance (a couple of hundreds of
instances for each size). Since in this experimental evaluation we were concerned
only in the off-line approach the real enter queue times were disregarded and in
our instances we assume that all jobs enter the system at the same time. The
main performance metric considered is the time spent by the jobs in the queues
while waiting their execution to begin (ideally as low as possible).

6.1 Evaluation of Our Models

We decided to make our experiments using two artificial versions of the Eurora
machine: A) a machine composed with 8 nodes and B) a machine with 32 nodes.
In addition to the real traces (set base) we generated two more sets of instances,
one which is composed by especially computationally intensive jobs, in terms
of resource requested (highLoad ), and one which presents jobs composed by

2
https://developers.google.com/optimization/
534 A. Borghesi et al.

an unusually high number of job-units (manyUnits). These additional groups

were generated using selected subsets of the original traces. From these sets we
randomly selected subsets of smaller instances with dimension of 25, 40, 50, 60,
80, 100, 150 and 200 jobs; for each size we used 50 different instances in our
tests. The instances of dimension 25 and 50 were run on the smaller machine A
while the remaining instances executed on machine B.
On each instance we ran the PRB algorithm (PRB ), the hybrid approach
with no feedback iteration (DEC noFeedBack ) and the hybrid approach with
feedback (DEC feedBack ). We tested the hybrid approach both with and with-
out the interaction between the two layers because the method without feedback
is much faster than the one with the interaction, therefore better suited to a real-
time application as a HPC dispatcher. The CP component of the decomposed
method has a timeout which forces the termination of the search phase. The
timeout is set to 5 seconds; if the solver finds no solution within the time limit,
the search is restarted with a increased timeout (we multiply by 2 the previous
timeout), until we reach a maximum value of 60 seconds - which is actually never
reached in our experiments. The PRB is extremely fast as finding a solution
takes only a fraction of second even with the larger instances; DEC noFeedBack
requires up to 3-4 seconds with the larger instances (but usually a lower qual-
ity solution is found in less than a second) and the DEC noFeedBack requires
significant larger times to compute a solution due to the multiple iterations, in
particular up to 15-20 seconds with the instances of 200 jobs.
Each experiment was repeated with different values of power capping to
explore how the bound on the power influences the behaviour of the developed
dispatcher.
At every run we measured the average weighted queue time of the solution,
that is the average time spent waiting by the jobs, weighted with the expected
wait time (given by the queue of the job). As a unit of measure we use the
Expected Wait Time, (EWT), i.e. the ratio between the real wait time and the
expected one. A EWT value of 1 tells us that a job has waited exactly as long
as it was expecting; values smaller than 1 indicate that the job started before
the expected start time and values larger than one that the job started later
than expected. To evaluate the performance of the different approaches we then
compute the ratio between the average weight queue time obtained by PRB and
by the two hybrid methods; finally we plot these ratios in the figures presented
in the rest of the section. Since the best average queue times are the lowest
ones, it is clear that if the value of the ratio goes below one the PRB approach
is performing better, while when the the value is above one then the hybrid
approaches are obtaining better results.
The following figures will show the ratios in the y-axis while the x-axis will
specify the power capping level; we only show significant power capping levels,
i.e. the one larger enough to allow a solution to the problem, hence the x-scale
range may vary.

Machine with 8 nodes: Figures 1, 2 and 3 show the results of the experiments
with the machine with 8 nodes; each figure corresponds respectively to the base
 Power Capping in High Performance Computing Systems 535

workload set of instances (both size 25 and 50 jobs), the highLoad case and
finally the manyUnits case. The solid lines represent the ratios between the
average queue times obtained by PRB and those obtained by DEC feedBack ;
conversely, the dashed line is the ratio between PRB and DEC noFeedBack. As
we can see in Figure 1 with an average workload the hybrid approaches usually
outperform the heuristic algorithm, markedly in the 25 jobs case and especially
at tighter power capping. With less tight power capping levels usually the hybrid
approaches and PRB offer more similar results; this is reasonable since when the
power constraints are more relaxed the allocation and scheduling decisions are
more straightforward and the more advanced reasoning offered by CP is less
necessary.

(a) 25 Jobs (b) 50 Jobs

Fig. 1. 8 Node - Base Set

It is easy also to see that the hybrid method with the feedback mechanism
always outperforms the feedback-less as we expected: the feedback-less solu-
tion has always inferior quality w.r.t. to those generated by the method with
the interaction - the solution produced by DEC feedBack is the same produced
by DEC noFeedBack. Our focus should be on the extent of the improvement
guaranteed by the feedback mechanism in relation to the longer time required
to reach a solution. For example, Fig.1 (corresponding to the original Eurora
workloads) shows that the feedback method offers clear advantages, in particu-
lar with tighter power constraints (around 10%-15% gain over the feedback-less
one), a fact that could justify its use despite the longer times required to reach a
solution. Conversely Figure 2 reveals that the two hybrid approaches offer very
similar performance if the workload is very intensive, leading us to prefer the
faster DEC noFeedBack in these circumstances in case of the implementation of
the real dispatcher.
If we consider the workload characterized by an unusually high amount of
job-units per job we can see slightly different results: as displayed in Figure 3,
DEC feedBack definitely outperforms the other two methods with tight power
capping values, especially in the case of instances of 25 jobs. When the power
constraints get more relaxed the three approaches offers almost the same results.
536 A. Borghesi et al.

(a) 25 Jobs (b) 50 Jobs

Fig. 2. 8 Node - HighLoad Set

Machine with 32 nodes: In Figures 4 and 5 we can see the results of some of the
experiments done on the machine with 32 nodes (in particular we present the
case size 40 and 100); we present only a subset of the experiments made due to
space limitations, but the results not shown comply with those presented here.
Figure 4 shows again the comparison between the two hybrid approaches
and the heuristic technique in the case of average workload. The pattern here is
slightly different from the 8-nodes case: after an initial phase where the hybrid
methods perform better, PRB offers better results for intermediate levels of
power capping (around 3000W); after that we can see a new gain offered by
the hybrid techniques until we reach a power capping level around 6000W (the
power constraint relaxes), where the three approaches provide more or less
the same results. It is evident again that the method with feedback outperforms
the feedback-less one, especially with the larger instances.

(a) 25 Jobs (b) 50 Jobs

Fig. 3. 8 Node - ManyUnits Set

 Power Capping in High Performance Computing Systems 537

(a) 40 Jobs (b) 100 Jobs

Fig. 4. 32 Node - Base Set

In Figure 5 we can see the results obtained with the computationally more
intensive workload. In this case PRB performs generally better at lower power
capping levels until the power constraint become less tight and the three methods
produce similar outcomes again.

(a) 40 Jobs (b) 100 Jobs

Fig. 5. 32 Node - Highload Set

We have performed a number of additional experiments, not shown here,

and we could summarize that with the machine with 32 nodes with average
workloads the hybrid approaches perform definitely better than the heuristic
technique for most power capping levels except a small range of intermediate
values (up to a 60% reduction of average queue times). On the contrary, if we
force very intensive workloads we obtain better outcomes with PRB, even though
with usually smaller, but still significant, differences (around 10%). We can also
see that with these intensive workloads the difference between the two hybrid
approaches is minimal and this suggest that the basic feedback mechanism needs
to be refined.
538 A. Borghesi et al.

7 Conclusions
In this paper we dealt with the allocation and scheduling problem on Eurora
HCP system subject to the power consumption constraint - power capping. We
presented two approaches to solve this problem: 1) a heuristic algorithm and
2) a novel, hybrid approach which combines a CP model for scheduling and a
heuristic component for the allocation. We compared the two approaches using
the average queue times as an evaluation metrics. Short waiting times correspond
to a higher quality of service for the system users.
We tackled a complex problem due to the limited amount of multiple, hetero-
geneous, resources and the additional constraint introduced by the finite power
budget. In addition to the complexity of the problem (scheduling and allocation
are NP-Hard problems) we also had to consider the real-time nature of the appli-
cation we wanted to develop, thus we had to focus on methods able to quickly
produce good solutions. In this preliminary study we shown that the quality of
the solution found by the different approaches varies with the levels of power
capping considered.
As a long-term goal we plan to further develop the preliminary power capping
feature presented here in order to integrate it within the actual dispatcher we
already developed and implemented on Euora. In order to do that we will need
to develop methods to allow our approach to operate quickly enough to match
the frequency of job arrivals; the heuristic would be already fast enough but the
hybrid approach will require us to research new techniques in order to cope with
the real-sized version of Eurora - and possibly even larger systems.
Acknowledgement. This work was partially supported by the FP7 ERC Advance
project MULTITHERMAN (g.a. 291125). We also want to thank CINECA and
Eurotech for granting us the access to their systems.

References
1. Mudge, N.: In: Culler, P., Druschel, D.E. (eds.) OSDI. USENIX Association (2002).
Operating Systems Review 36(Special Issue), Winter 2002
2. Auweter, A., et al.: A case study of energy aware scheduling on supermuc.
In: Kunkel, J.M., Ludwig, T., Meuer, H.W. (eds.) ISC 2014. LNCS, vol. 8488,
pp. 394–409. Springer, Heidelberg (2014)
3. Banerjee, A., Mukherjee, T., Varsamopoulos, G., Gupta, S.K.S.: Cooling-aware and
thermal-aware workload placement for green hpc data centers. In: Green Comput-
ing Conference, pp. 245–256 (2010)
4. Baptiste, P., Laborie, P., Le Pape, C., Nuijten, W.: Constraint-based scheduling
and planning. Foundations of Artificial Intelligence 2, 761–799 (2006)
5. Baptiste, P., Le Pape, C., Nuijten, W.: Constraint-based scheduling. Kluwer Aca-
demic Publishers (2001)
6. Bartolini, A., Cacciari, M., Cavazzoni, C., Tecchiolli, G., Benini, L.: Unveiling
eurora - thermal and power characterization of the most energy-efficient supercom-
puter in the world. In: Design, Automation Test in Europe Conference Exhibition
(DATE), 2014, March 2014
 Power Capping in High Performance Computing Systems 539

7. Bartolini, A., Borghesi, A., Bridi, T., Lombardi, M., Milano, M.: Proactive work-
load dispatching on the EURORA supercomputer. In: O’Sullivan, B. (ed.) CP 2014.
LNCS, vol. 8656, pp. 765–780. Springer, Heidelberg (2014)
8. Bartolini, A., Cacciari, M., Tilli, A., Benini, L.: Thermal and energy management
of high-performance multicores: Distributed and self-calibrating model-predictive
controller. IEEE Trans. Parallel Distrib. Syst. 24(1), 170–183 (2013)
9. Bergman, K., Borkar, S., Campbell, D., Carlson, W., Dally, W., Denneau, M.,
Franzon, P., Harrod, W., Hiller, J., Karp, S., Keckler, S., Klein, D., Lucas, R.,
Richards, M., Scarpelli, A., Scott, S., Snavely, A., Sterling, T., Williams, R.S.,
Yelick, K., Bergman, K., Borkar, S., Campbell, D., Carlson, W., Dally, W., Den-
neau, M., Franzon, P., Harrod, W., Hiller, J., Keckler, S., Klein, D., Kogge, P.,
Williams, R.S., Yelick, K.: Exascale computing study: Technology challenges in
achieving exascale systems, September 2008
10. Chu, R.C., Simons, R.E., Ellsworth, M.J., Schmidt, R.R., Cozzolino, V.: Review
of cooling technologies for computer products. IEEE Transactions on Device and
Materials Reliability 4(4), 568–585 (2004)
11. Collina, F.: Tecniche di workload dispatching sotto vincoli di potenza. Master’s
thesis, Alma Mater Studiorum Università di Bologna (2014)
12. Conficoni, C., Bartolini, A., Tilli, A., Tecchiolli, G., Benini, L.: Energy-aware cool-
ing for hot-water cooled supercomputers. Proceedings of the 2015 Design. Automa-
tion & Test in Europe Conference & Exhibition, DATE 2015, pp. 1353–1358. EDA
Consortium, San Jose (2015)
13. Dongarra, J.J.: Visit to the national university for defense technology changsha,
china. Technical report, University of Tennessee, June 2013
14. Dongarra, J.J., Meuer, H.W., Strohmaier, E.: 29th top500 Supercomputer Sites.
Technical report, Top500.org, November 1994
15. Fan, X., Weber, W.-D., Barroso, L.A.: Power provisioning for a warehouse-sized
computer. In: ACM SIGARCH Computer Architecture News, vol. 35, pp. 13–23.
ACM (2007)
16. Feng, W.-C., Cameron, K.: The Green500 List: Encouraging Sustainable Super-
computing. IEEE Computer 40(12) (2007)
17. Gandhi, A., Harchol-Balter, M., Das, R., Kephart, J.O., Lefurgy, C.: Power capping
via forced idleness (2009)
18. Haupt, R.: A survey of priority rule-based scheduling. Operations-Research-
Spektrum 11(1), 3–16 (1989)
19. Kim, J.M., Chung, S.W., Seo, S.K.: Looking into heterogeneity: When simple is
faster
20. Kim, J., Ruggiero, M., Atienza, D.: Free cooling-aware dynamic power manage-
ment for green datacenters. In: 2012 International Conference on High Performance
Computing and Simulation (HPCS), pp. 140–146, July 2012
21. Kim, J., Ruggiero, M., Atienza, D., Lederberger, M.: Correlation-aware virtual
machine allocation for energy-efficient datacenters. In: Proceedings of the Confer-
ence on Design. Automation and Test in Europe, DATE 2013, pp. 1345–1350. EDA
Consortium, San Jose (2013)
22. Kontorinis, V., Zhang, L.E., Aksanli, B., Sampson, J., Homayoun, H., Pettis, E.,
Tullsen, D.M., Rosing, T.S.: Managing distributed ups energy for effective power
capping in data centers. In: 2012 39th Annual International Symposium on Com-
puter Architecture (ISCA), pp. 488–499. IEEE (2012)
540 A. Borghesi et al.

23. Kudithipudi, D., Qu, Q., Coskun, A.K.: Thermal management in many core sys-
tems. In: Khan, S.U., Koodziej, J., Li, J., Zomaya, A.Y. (eds.) Evolutionary Based
Solutions for Green Computing. SCI, vol. 432, pp. 161–185. Springer, Heidelberg
(2013)
24. Laborie, P., Rogerie, J.: Reasoning with conditional time-intervals. In: Proc. of
FLAIRS, pp. 555–560 (2008)
25. Lefurgy, C., Wang, X., Ware, M.: Power capping: a prelude to power shifting.
Cluster Computing 11(2), 183–195 (2008)
26. Pape, C.L., Couronné, P., Vergamini, D., Gosselin, V.: Time-versus-capacity com-
promises in project scheduling. In Proc. of the 13th Workshop of the UK Planning
Special Interest Group, pp. 1–13 (1994)
27. Reda, S., Cochran, R., Coskun, A.K.: Adaptive power capping for servers with
multithreaded workloads. IEEE Micro 32(5), 64–75 (2012)
28. Altair PBS Works: Pbs professional12.2 administrator’s guide (2013). http://
resources.altair.com/pbs/documentation/support/PBSProAdminGuide12.2.pdf
29. You, H., Zhang, H.: Comprehensive workload analysis and modeling of a petascale
supercomputer. In: Cirne, W., Desai, N., Frachtenberg, E., Schwiegelshohn, U.
(eds.) JSSPP 2012. Lecture Notes in Computer Science, vol. 7698, pp. 253–271.
Springer, Heidelberg (2013)
A Constraint-Based Approach to the Differential
Harvest Problem

Nicolas Briot1(B) , Christian Bessiere1 , and Philippe Vismara1,2

1
LIRMM, CNRS–University Montpellier, Montpellier, France
{briot,bessiere,vismara}@lirmm.fr
2
MISTEA, Montpellier SupAgro, INRA, 2 Place Viala, 34060 Montpellier, France

Abstract. In this paper, we study the problem of differential har-

vest in precision viticulture. Some recent prototypes of grape harvesting
machines are supplied with two hoppers and are able to sort two types
of grape quality. Given estimated qualities and quantities on the differ-
ent areas of the vineyard, the problem is to optimize the routing of the
grape harvester under several constraints. The main constraints are the
amount of first quality grapes to harvest and the capacity of the hoppers.
We model the differential harvest problem as a constraint optimization
problem. We present preliminary results on real data. We also compare
our constraint model to an integer linear programming approach and
discuss expressiveness and efficiency.

1 Introduction
In precision viticulture, many studies have proposed to define field quality zones
[14]. They demonstrated the technical and economic value of a differential har-
vesting of these different zones. This interest justifies the recent development
of prototypes of conventional grape harvesting machines able to sort two types
of harvest quality, such as the EnoControlTM system prototype (newHolland
Agriculture, PA, USA). These grape harvesting machines have two tanks, called
hoppers, able to differentiate two types of grape quality, named A and B, accord-
ing to the harvested zone.
Optimizing harvest consists on minimizing the working time of grape har-
vester. This time corresponds to travel time and emptying time of the machine.
Ideally, this goal requires that both hoppers of the machine are full at each
emptying. In the case of selective harvesting, the simultaneous filling of the
two hoppers is combinatorial and complex. Indeed, the hopper that contains A
grapes can fill up at a speed different from the hopper that contains B grapes,
depending on the harvested zone. Other issues have to be considered. Top quality
grapes should not be altered (mixed with lower quality grapes) when the har-
vester moves from one quality zone to another. Turning radius of the machine
must also be taken into account.
This problem, called Differential Harvest Problem, has not been studied in
the literature except in a preliminary study we published in [4]. A comparison
with routing problems can be established. For instance, Bochtis et al. [1,2] show

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 541–556, 2015.
DOI: 10.1007/978-3-319-23219-5 38
542 N. Briot et al.

various formulations for many related problems in agriculture, such as precision

spraying, fertilizing in arable farming, etc. In [9], Kilby and Shaw give a Con-
straint Programming (CP) formulation for the Vehicle Routing Problem (VRP)
and detail other exact methods and heuristics to solve the problem. Balancing
Bike Sharing System (BBSS) is an example of real application solved with VRP
formulation and CP approach. Di Gaspero et al. [7] show two novel CP models
for BBSS and a Large Neighborhood Search approach is proposed.
In this paper, we explore how to solve the Differential Harvest Problem with
a CP approach. We consider two models: a Step model we introduced in our
previous work [4] and a new routing model that we call the Precedence model.
We also compare these two models to an Integer Linear Programming approach.
The paper is organized as follows. Section 2 describes the problem of differ-
ential harvest in viticulture. Section 3 proposes the two CP models. We present
preliminary results on real data in Section 4. We compare our CP model to an
Integer Linear Programming approach and discuss expressiveness and efficiency
in Section 5.

2 The Differential Harvest Problem

Precision viticulture is a new farming management approach that focuses on the

intra-field variability in a vineyard. Thanks to high resolution satellite images
and infield sensors, it is possible to predict the variability of quality and yield
within a vine field. The Differential Harvest Problem (DHP) is defined on a vine
field with two qualities of grapes. The first one will be denoted A grapes and
corresponds to high quality grapes. These grapes will be used to produce high
quality wine. The rest of the harvest will be put in the B-grapes class. Note that
high quality grapes (A grapes) can be downgraded to the low quality B. But the
converse is not true.
We consider a harvesting machine with two hoppers. We denote by Cmax the
maximum capacity of a hopper. With such a machine, two categories of grapes
can be separated. Generally, one hopper (called A hopper) receives only A grapes
whereas the other one (called B & A hopper) can contain either B grapes only or
both A grapes and B grapes. It is important not to contaminate the grapes in A
hopper with B grapes. When the hoppers are full, the machine can dump them
into a bin located around the plot before continuing its work. Generally, the bin
is immediately brought to the winery after each dump of the hoppers in order
to preserve grapes quality. A second bin waits in stand-by to avoid delaying the
harvester.
The harvesting machine takes some time between picking grapes and putting
them into hoppers. This time (≈ 10 seconds), required to empty the conveyor
belt, is called latency. When the harvesting machine passes over two zones, the
quality of grapes harvested is not guaranteed during this time because latency
may lead to mixing both qualities of grapes. Hence, when the machine leaves a
zone of A grapes to enter a zone of B grapes or leaves a zone of B grapes to
enter a zone of A grapes, the grapes must be considered B grapes. For the same
 A Constraint-Based Approach to the Differential Harvest Problem 543

row, the type of transitions may vary according to the direction that the row is
harvested. Because of latency, the quantity of A grapes (resp. B grapes) that
are collected within a given row can change with the direction. For instance,
consider a row where the sequence of quality areas is A-B-A-B (see Figure 1). If
the machine harvests the row from left to right, there is only one B-A transition.
If the machine harvests the row in the opposite direction, two transitions B-A
appear. Thus, quantities of A grapes and B grapes harvested will necessarily
depend on the direction in which the rows are harvested.

A grapes B grapes A grapes B grapes

A hopper:
B&A hopper:
Latency

A grapes B grapes A grapes B grapes

A hopper:
B&A hopper:
Latency Latency

Fig. 1. Quantities of grapes depend on the direction of row harvest.

Consider Rmin, the desired minimum volume of A grapes harvested in the

vine field. Rmin is a threshold. In practice, Rmin corresponds to shortfall of a
volume of A grapes according to the objectives of the winery. If the vineyard
contains more than Rmin A grapes, the excess amount can be downgraded to
B-grapes quality. As long as Rmin has not been reached, A grapes are stored
into the A hopper. Once Rmin has been reached and after the hoppers have
been emptied, A grapes and B grapes can be mixed in both hoppers. Regarding
Rmin, there are three possibilities to fill the two hoppers. When the harvesting
machine must differentiate qualities in hoppers and the A hopper is not full, A
grapes are put in the A hopper and B grapes in the B & A hopper (see Figure 2.a).
When A hopper is full, the machine can put A grapes in the B & A hopper (see
Figure 2.b). In such a case, these A grapes are downgraded. Once Rmin has
been reached, the machine can mix grapes in both hoppers (see Figure 2.c).
The vine field, composed of n rows, is modelled by differentiating the two
extremities of each row. A row r ∈ {0, . . . , n − 1} is represented by extremities
2r and 2r + 1. For each row, we denote QA B A
2r→2r+1 and Q2r→2r+1 (resp. Q2r+1→2r
B
and Q2r+1→2r ) the quantities of A grapes and B grapes that will be collected
in row r with orientation 2r → 2r + 1 (resp. 2r + 1 → 2r). These quantities are
computed according to the latency.
544 N. Briot et al.

A grapes B grapes A grapes B grapes A grapes B grapes

A hopper B&A hopper A hopper B&A hopper A hopper B&A hopper

(a) full (b) (c)

Fig. 2. Three possibilities to fill the hoppers. (a) A grapes and B grapes are separated.
(b) When A hopper is full, A grapes and B grapes are mixed in the B&A hopper. (c)
Once Rmin has been reached, A grapes and B grapes are mixed in the two hoppers.

Another important information is the cost of the path between two extremi-
ties of different rows and between row extremities and the bin place. We denote
d(p, q) = d(q, p) ∀p, q ∈ {0, 1, . . . , 2n − 1} ∪ {2n} the time required to go from
an extremity p to an extremity q (where 2n denotes the bin place). This cost
depends on the distances between extremities and the turning radius of the
harvesting machine.
We are ready to define the problem.

Definition 1 (Differential Harvest Problem). Given a vine field described

by a cost path matrix between row extremities (or the bin place) and an estimation
of the quantity of A grapes and B grapes on each row according to the direction,
given a harvesting machine with a hopper capacity of Cmax and a latency delay,
given a threshold of Rmin A grapes to harvest, the Differential Harvest Problem
consists in finding a sequence of extremities (that is, an order and orientation of
the rows) that minimizes the time required to harvest the vine field and ensures
at least Rmin A grapes harvested.

3 Constraint Programming for the DHP

The problem of differential harvest in precision viticulture is a recent problem.

Its specification is not yet stable as it corresponds to a simplification of the real
agronomic problem. In this context, other constraints will certainly appear when
the first solutions will be applied. An example of probable extra constraint is
shown in section 5.2. Constraint programming is well recognized for its flexibility
and ease of maintenance. For these reasons, using constraint programming to
model the DHP seems to be a good approach.
This section is devoted to present two CP models. First, the Step model
consists in choosing which row to harvest step by step. Second, the Precedence
model is devoted to find the best predecessor of each row.
 A Constraint-Based Approach to the Differential Harvest Problem 545

Given a set of variables and a set of possible values for each variable, given a
set of constraints, and given an optimization criterion, constraint optimization
consists in finding an instantiation of variables that satisfies all constraints of
the problem and minimizes the criterion.

3.1 The Step Model

In [4], we presented a preliminary CP model to solve the Differential Harvest
problem. This model was dedicated to measure the gain of an optimized rout-
ing compared to the traditional approach. The traditional routing consists in
systematically taking the second next row on the same side,1 until the capacity
of one hopper is reached. We showed in [4] that using a CP model to optimize
the routing can reduce the harvest time of almost 40% compared to traditional
harvest routing.
The Step model presented in [4] is illustrated in Figure 3. This model is based
on three main kinds of variables that describe:
– the number of the row visited at each step of the route,
– the orientation of the visited row (harvest direction),
– the act of emptying the hoppers into the bin at the end of the visited row
or not.

0 1 2 3 4 5 6 7

Step 0 1 2 3 4 5 6 7
Row 1 4 5 7 0 2 3 6
Orientation 0 1 0 1 0 1 0 1
Emptying into the bin? 0 1 0 0 0 1 0 1

Fig. 3. The Step model.

An Alldifferent constraint ensures that the machine passes exactly once in

each row. Additional variables are used to represent the quantity of grapes har-
vested until each step. They are subject to a set of numeric constraints that
control the capacities of the hoppers.
The first k steps of the route perform differential harvest. Before step k,
B grapes cannot appear in A hoppers, i.e. harvest is separated. In step k, the
1
Because of the turning radius of the machine, turning into the next row is generally
longer than jumping to the second next row.
546 N. Briot et al.

machine goes to the bin and must have harvested at least a total of Rmin
A grapes. After step k, A grapes and B grapes are mixed in both hoppers.
In the first version of this Step model we included k as a variable but experi-
mental results were very bad. The combinatorial aspect of the problem decreases
if we set k to a fixed value. Hence, each instance of the original problem is
replaced by a set of sub-instances with different values of k. According to the
capacity of the hoppers and the value of Rmin, few values of k must be consid-
ered. Since these instances are independent, they can be executed in parallel.
In section 4, we will see that this model is not effective. The next subsection
gives another model for the DHP, based on models for routing problems.

3.2 The Precedence model

The Precedence model is based on variables that represent the previous row of a
row. Our model shares similarities with the model presented in [9] for the Vehicle
Routing Problem.
For a given instance of the Differential Harvest Problem, suppose that we
have an upper bound λ on the number of times the hoppers have to be dumped
into the bin. We note γ the number of dumps used in the differential part (begin-
ning) of the harvest (γ ≤ λ ≤ n) where n is the number of rows in the vine field.
We call R = {0, . . . , n−1} the set of rows. S = R∪{n, . . . , n+γ −1, . . . , n+λ−1}
denotes the set of sites (rows and hopper dumps into the bin).

0 1 2 3 4 5 6 7

8 9
10

3 harvest mixed 8 9 10 bins

row
1 harvest separated link predecessor/successor

Site 0 1 2 3 4 5 6 7 8 9 10
Predecessor 7 10 0 8 1 9 3 5 2 4 6
Orientation 0 0 1 0 1 0 1 1
Mix 0 0 0 1 0 0 1 0 0 0 1

Fig. 4. The Precedence model.

 A Constraint-Based Approach to the Differential Harvest Problem 547

Variables. For each site i ∈ S we create:

– Pi , Si : are integer variables that represent the direct predecessor (Pi ) and
direct successor (Si ) of site i. We have D(Pi ) = D(Si ) = S \ {i};
– Mixi : is a Boolean variable that represents the harvest mode in site i. If
M ixi = 0 the A grapes are separated (differential harvest) otherwise they
are mixed with B grapes. The domain of this variable is: D(M ixi ) = {0, 1};
– Ti : is an integer variable that represents the time to travel from Pi to site
i, including the time to harvest row i. D(Ti ) = N;
– UA B
i , Ui are integer variables that represent the quantity of A grapes and B
grapes harvested up to site i since the last dump into the bin. The domains
are: D(UiA ) = D(UiB ) = {0, . . . , 2 × Cmax}.

For each row r ∈ R we have:

– Orir : is a Boolean variable that represents the orientation for row r (0 is the
direction from odd to even extremities, as described in Section 2);
– uA B
r (resp. ur ): represents the quantity of grapes of quality A (resp. B)
harvested in row r according to the direction of harvest. D(uA B
r ) = D(ur ) =
N;

Constraints. Predecessor variables form a permutation of rows and are subject

to the alldifferent constraint (1).

AllDif f erent(P0 , . . . , Pn+λ−1 ) (1)

Constraint (2) is a channelling constraint between predecessor and successor

variables.

P Si = i SPi = i ∀i ∈ S (2)

Constraints (3) and (4) force the harvest mode to be differential or mixed accord-
ing to the index of the corresponding hopper dump. Constraint (5) is an element
constraint that communicates the harvest mode between successors:

M ixi = 0 ∀i ∈ {n, . . . , n + γ − 1} (3)

M ixi = 1 ∀i ∈ {n + γ, . . . , n + λ − 1} (4)
M ixr = M ixSr ∀r ∈ R (5)

The following constraints give the quantities of grapes according to the orienta-
tion of the row. It can be implemented as a table constraint or as an if . . . then
. . . else one. ∀r ∈ R, we have :

uA A A
r = Orir × Q2r→2r+1 + (1 − Orir ) × Q2r+1→2r (6)
uB B B
r = Orir × Q2r→2r+1 + (1 − Orir ) × Q2r+1→2r (7)

Constraint (8) fixes quantities A and B for all sites representing dumps into
the bin. Constraint (9) computes the quantities at the end of row i by adding
548 N. Briot et al.

the accumulated amount from predecessor Pi and the quantity in row i given by
the precedent constraints.

Uiα = 0 ∀α ∈ {A, B} ∀i ∈ S \ R (8)

Uiα = UPαi + uα
i ∀α ∈ {A, B} ∀i ∈ R (9)

Harvested quantities are limited by the capacity of hoppers. Variable UiA always
have an upper bound of 2Cmax because A grapes can be put in the two hoppers.
When variable M ixi = 0, harvest is not mixed and quantity of B grapes is bound
by Cmax (10). When variable M ixi = 1, A grapes and B grapes are mixed in the
two hoppers. Constraint (11) checks that the total amount of harvested grapes
is smaller than the capacity of the two hoppers:

UiB ≤ (1 + M ixi ) × Cmax ∀i ∈ R (10)

UiA + UiB ≤ 2 × Cmax ∀i ∈ R (11)

Constraint (12) requires to harvest at least Rmin A grapes. Only the A grapes
stored in A hopper must be considered. This quantity corresponds to the part of
UiA which is smaller than the capacity of the A hopper. It concerns the differential
harvest mode only, i.e. dumps from n to n + γ − 1.
n+γ−1

min(UpAi , Cmax) ≥ Rmin (12)
i=n

It is possible to reformulate constraint (12) without the minimum operator. We

add new variables called Aci ∀i ∈ {n, . . . , n + γ − 1} such that:

Aci ≤Cmax that is: D(Aci ) = {0, . . . , Cmax} (12’a)

Aci ≤UPAi ∀i ∈ {n, . . . , n + γ − 1} (12’b)
n+γ−1

Aci ≥Rmin (12’c)
i=n

Constraint (13) forces the exit from row Pi to be on the same side of the vine
field as the entrance of row i. Hence, Pi and i have inverse orientations. This is
the case in practice with traditional routing.

Orii = 1 − OriPi ∀i ∈ R (13)

Next constraints ((14a) and (14b)) require a unique cycle (subtour elimination)
on predecessor variables and successor variables. Figure 5 is an example of fil-
tering for the circuit constraint described in [5,12].

Circuit(1, (P0 , . . . , Pn+λ−1 )) (14a)

Circuit(1, (S0 , . . . , Sn+λ−1 )) (14b)
 A Constraint-Based Approach to the Differential Harvest Problem 549

Fig. 5. An example of circuit constraint propagator on successor variables. Arrows rep-

resent the link (successor) between two variables (circles). (a) is a current instantiation.
(b) filtering prohibits links between variables that represent the end and the start of
a same path (i.e., the value that represents the start of a subpath is removed from
the domain of variables that represent the end of this path). (c) and (d) represent two
possible solutions in this example.

Constraints (15) force Ti to be equal to the time to travel from Pi to site i,

including the time to harvest row i if it is the case. Function d gives the time to
travel between two extremities of rows (or an extremity and the bin) as defined
in Section 2. Note that for two consecutive dumps into the bin (i ∈ R and
Pi ∈ R) the travel time is equal to 0 (15b).

if Pr ∈ R then Tr = d(2Pr + OriPr , 2r + 1 − Orir )

else Tr = d(2n, 2r + 1 − Orir ) ∀r ∈ R (15a)
if Pi ∈ R then Ti = d(2Pi + OriPi , 2n) else Ti = 0 ∀i ∈ S \ R (15b)

Constraint (16) is the objective function that minimizes the travel time of
the grape harvester:
n+λ−1

M inimize Ti (16)
i=0

Symmetry Breaking. The main variables of this model are predecessor vari-
ables. They take their values in a set of rows and dumps into the bin. For the
set of dumps in the differential (or mixed) mode, there is a value symmetry
caused by the possibility to dump indexes. To break this symmetry, we add an
550 N. Briot et al.

ordering constraint (17a) (resp. (17b)) on the variables that correspond to the
same harvest mode.

Pi < Pi+1 ∀i ∈ {n, . . . , n + γ − 2} (17a)

Pi < Pi+1 ∀i ∈ {n + γ, . . . , n + λ − 2} (17b)

4 Experimental Results

In this section, we present some experimental results. We have implemented the

two models in Choco [13] using the variable ordering strategy DomOverWDeg
[3] limited to the Pi variables. All experiments were executed on a Linux server
with an Intel(R) Xeon(R) CPU E5-2697 2.60GHz processor. It has 14 cores that
makes it possible to solve several instances in parallel but the solver was not
configured to exploit the multicore architecture.
Our models were tested on a real data from an experimental vineyard of
INRA Pech-Rouge (Gruissan) located in southern France (see Figure 6). In this
experiment we want to compute the optimal solution. To test the models on
instances of increasing size, we have extracted subsets of continuous rows from
the 24 rows vine field.

Fig. 6. A vine field with two qualities B grapes in red and A grapes in green.

Table 1 reports results with Step and Precedence models for 10, 12 and 14
rows. Each row in the table reports the average time on 12 instances of the
given size generated from contiguous sequences of rows of the vine field. For
each sequence, we experimented with two capacities of the hoppers (1000 and
2000) and two values for the desired threshold Rmin (50% and 70% of the total
amount of A grapes).
 A Constraint-Based Approach to the Differential Harvest Problem 551

For the Step model, each instance of the Differential Harvest Problem is split
into a set of more specific sub-instances of the constraint optimization problem.
Each sub-instance is associated with a fixed value k of number of rows harvested
on differential mode. For 10 (resp. 12) rows we varied k from 4 to 8 (resp. 6 to
10). Each set of sub-instances can be solved in parallel. So, Table 1 shows the
maximum time needed to solve a sub-instance of a given set. The total time (for
all sub-instances) corresponds to a single core architecture. Table 1 also gives
the time for the sub-instance which has the optimal solution. With sequences
of 14 rows, at least one sub-instance of each instance is not solved before the
timeout (7200 sec).

Table 1. Comparison of the Step model and the Precedence model. All CPU times
are averages on 12 instances and are given in seconds. Timeout at 7200 seconds.

Step model Precedence model

# Parallel solving Single processor Time for the sub-instance Time
rows (maximum time) (total time) with the optimal solution
10 262 527 180 8
12 3414 8466 2749 118
14 timeout 8 instances solved

These results show that the Step model is clearly outperformed by the Prece-
dence model. One disadvantage of the Step model is that it has to communicate
the harvested quantities between steps. Hence, as long as the route is undefined,
it is impossible to know if the capacity of hoppers is exceeded and if Rmin is
reached. This hinders propagation. There is another drawback of the Step model.
It assumes a fixed value k for the step on which differential harvest stops (A
grapes and B grapes are mixed from step k to step n). Finding an optimal solu-
tion requires to solve several instances of the problem with different values of k.
This can be done in parallel as the problems are independent. But Table 1 shows
that, even with an oracle predicting the value k of the sub-instance containing
the optimal solution of the original instance, the Step model is outperformed by
the Precedence model.
Table 2 shows results for instances of the Precedence model for sequences
of 12, 14 and 16 rows. The first column shows the number of rows in the vine
field. The second column gives the capacity of the hoppers (Cmax) and the third
column gives the desired threshold Rmin. For each size of problem, we give the
average time to solve it and the average number of nodes. Timeout is 7200 sec.
For 16 rows with a Cmax value of 1000 and for any greater number of rows, all
instances exeeded the timeout.
552 N. Briot et al.

Table 2. Precedence model on 12, 14 and 16 rows.

# Rows Cmax Rmin Time (s) # Nodes

50% 227 663267
1000
70% 170 667112
12
50% 35 153300
2000
70% 41 175017
50% 2/3 instances solved
1000
70% 0/3 instances solved
14
50% 52 181262
2000
70% 21 72468
50% 856 2835199
16 2000
70% 106 318246

Table 3. Impact of different configurations for the Precedence model

Precedence without symmetry with min with minDomain

model breaking constraints function in (12) variable ordering
# Row Time # Nodes Time # Nodes Time # Nodes Time # Nodes
10 8 41811 17 95363 167 382345 101 1066550
12 118 441674 539 4017780 1042 4824446 776 3582405

These results indicate fairly high variability in CPU times. Instances with
a small value of Cmax seem to be more difficult to solve. When the threshold
Rmin is equal to 50% of A grapes, CPU time generally increases.
Table 3 shows results for different configurations of the Precedence model.
All CPU times and node numbers are average on all instances of the same
number of rows. The first two columns correspond to the complete Precedence
model. The next columns give the results for variants of this model: without
symmetry breaking (constraints (17a) and (17b)); with constraint (12) instead
of constraints (12’a), (12’b) and (12’c); and with the default Choco variable
ordering (minDomain) instead of the DomOverWDeg strategy.
Despite a small number of bins, adding constraints to eliminate symmetric
solutions helps a lot to reduce the search effort. It is also the case when con-
straint (12) with the minimum function is replaced by constraints (12’a), (12’b)
and (12’c), or when the default Choco variable ordering is replaced by DomOver-
WDeg.

5 Discussion

Our experimental results show that the Precedence model is orders of magnitude
faster than the Step model. For small instances (up to 16 rows), the Precedence
model solves the Differential Havest Problem in a reasonable amount of time.
For larger instances, the timeout of 2 hours is generally reached. This amount
of time has to be compared to the time needed to collect data and to build the
 A Constraint-Based Approach to the Differential Harvest Problem 553

quality and yield map for the vine field. With additional improvements of the
CP model, we can expect to solve larger real instances.
Our results confirm what other studies have reported on similar problems
like the BBSS problem presented in [7]. In that paper, Di Gaspero et al. show
that a Step formulation is outperformed by Routing (similar to Precedence)
formulation for the optimal solution. In a second experiment, they show that
large neighborhood search (LNS) is a good approach to decrease the time to
solve, though the optimal solution can no longer be guaranteed. It could be
interesting to explore that direction.

5.1 Comparison with an ILP Formulation

Contrary to our expectations, it was not difficult to express our model of the
Differential Harvesting Problem as a set of linear constraints on integer variables.
Such a problem can be solved with Integer Linear Programming (ILP).
So we have designed an ILP model using the two-index vehicle flow formula-
tion of the Capacited Vehicle Routing Problem (CVRP) introduced by Laporte,
Nobert, and Desrochers [10]. There is one Boolean variable xt:i→j , for each pair
i, j of rows extremities and bin on tour t. Each tour corresponds to a single mode
of harvest (differential or not). xt:i→j is equal to 1 if the harvesting machine goes
from i to j on tour t and 0 otherwise. Constraints on the xt:i→j variables ensure
that each extremity is visited exactly once on all routes. Quantities of grapes
harvested in each tour are expressed by a summation over the traversed rows and
similarly for travel cost (objective function). As in CVRP formulation, there are
constraints on capacity of hoppers according to the tour t. Variables and linear
constraints (12’a), (12’b), and (12’c) on threshold Rmin can directly be added
to the ILP formulation.
Unfortunately, the cycle constraint cannot be added so easily. The basic ILP
approach to forbid sub-tours(cycles that do not pass through the bin) is to
post constraints of the form i,j∈S xt:i→j < |S| for any subset S of extremities
of rows. The number of such constraints is exponential. Another formulation
introduced by Miller, Tucker, and Zemlin (M T Z − f ormulation [11]) makes the
number of constraints polynomial but its linear relaxation generally produces a
significantly weaker lower bound compared to the basic model [6]. Hence, a typ-
ical approach consists in adding the sub-tour elimination constraints incremen-
tally, as follows: Step 1: find an optimal solution (without sub-tour constraints
at the beginning). Step 2: If the solution does not contain sub-tours, it is an
optimal solution of the problem; Otherwise, new constraints are added to forbid
all the sub-tours in the current solution. Then, proceed to Step 1. In the worst
case, this algorithm has to solve an exponential number of instances with finally
an exponential number of constraints on sub-tours. Thus, each intermediate ILP
loop gives relaxed solutions that contain sub-tours between rows. It is only in
the last ILP step that the reported (optimal) solution does not contain sub-
tours and can be exploited. As a consequence, if we interrupt the ILP solver at
a given timeout limit, it has not yet produced any feasible solution, not even
554 N. Briot et al.

Table 4. Comparison between CP model and ILP model for 12, 14, 16 and 24 rows,
Cmax = 1000, Rmin = 70% and timeout of 7200s. In bold, optimal solutions.

upper-bound for CP model ILP model

# rows
harvesting time harvesting time CPU time harvesting time CPU time
+∞ 960 556s
12 960 3s
2671 960 496s
+∞ 1120 timeout
14 1112 113s
2436 1116 timeout
+∞ 1260 timeout
16 - timeout
2694 1260 timeout
+∞ 1805 timeout
24 - timeout
2059 1800 timeout

suboptimal ones. This is an advantage of the CP model over the ILP model. Any
of its solutions, at any time in the solving process, can be exploited.
We have implemented the ILP model with the Cplex solver [8]. Table 4 shows
a comparison between the Precedence model and the ILP formulation using a
single core. Tests are performed with and without an initial upper bound. This
bound is an estimation of the harvest time using a manual routing based on
a repetitive pattern. This upper bound does not improve the ILP results so
we give only one value for the ILP model. Preliminary experiments show that
the ILP formulation clearly outperforms the CP model for small instances. But
for hard instances (n ≥ 16 and Cmax = 1000), ILP search fails to obtain a
solution before the timeout of 7200 seconds. On hard instances the CP model can
find suboptimal solutions that significantly improve the manual routing solution
whilst ILP provides no solution at all.

5.2 Complements to the Model

Another advantage of the CP model is its ability to integrate new constraints.

The formulation of the problem presented in this paper is a quite simplified
version of the Differential Harvest Problem and it is dedicated to evolve. For
instance, consider the case where the harvesting machine finishes a tour on an
extremity of a row opposite to the bin. If the row is in the middle of the vineyard,
the shortest path to the bin passes through a row. But it is not possible to go
to the bin by passing through a non-harvested row. This can be expressed by
the following rule: the path to the bin, for an extremity which is on the side
opposite to the bin, must pass through a row that precedes the extremity in the
global route. Such a constraint is very difficult to add to the ILP formulation
but can be easily implemented in the CP approach. The propagate procedure
of the circuit constraint already computes the set of rows that precede the last
step of any tour in order to detect subtours. So it is easy to find in this set what
is the best row to go back to the bin. This can be done without changing the
overall complexity of the propagate procedure.
 A Constraint-Based Approach to the Differential Harvest Problem 555

6 Conclusion
In this paper, we have presented the Differential Harvest Problem in precision
viticulture. We have proposed to use constraint programming to solve it. Two
models have been presented, the Step model and the Precedence model. In the
Step model, variables represent the row that is visited at a given time step and in
which direction the row is traversed. In the Precedence model, variables connect
a row to its predecessor and successor. The experiments we have performed to
assess the behavior of these models show that the Precedence model is orders
of magnitude faster than the Step model. We have also experimentally shown
that an ILP formulation of the Differential Harvest Problem outperforms our
CP approach on easy instances. However, such an ILP formulation requires an
exponential space, and more importantly, fails to produce solutions on hard
instances. All in all, our Precedence model seems to be a good approach. It
allows to solve the problem on real data within reasonable time and it inherits
the flexibility of CP models, that allows the addition of extra user-constraints
in a simple way.

References
1. Bochtis, D.D., Sørensen, C.G.: The vehicle routing problem in field logistics: part
i. Biosystems Engineering 104, 447–457 (2009)
2. Bochtis, D.D., Sørensen, C.G.: The vehicle routing problem in field logistics: part
ii. Biosystems Engineering 105, 180–188 (2010)
3. Boussemart, F., Hemery, F., Lecoutre, C., Sais, L.: Boosting systematic search
by weighting constraints. In: Proceedings of the 16th Eureopean Conference on
Artificial Intelligence (ECAI 2004), Valencia, Spain, pp. 146–150 (2004)
4. Briot, N., Bessiere, C., Tisseyre, B., Vismara, P.: Integration of operational con-
straints to optimize differential harvest in viticulture. In: Proc. 10th European
Conference on Precision Agriculture, July 2015 (to appear)
5. Caseau, Y., Laburthe, F.: Solving small TSPs with constraints. In: Naish, L. (ed.)
Logic Programming, Proceedings of the Fourteenth International Conference on
Logic Programming, July 8–11, 1997, pp. 316–330. MIT Press, Leuven, Belgium
(1997)
6. Desrochers, M., Laporte, G.: Improvements and extensions to the miller-tucker-
zemlin subtour elimination constraints. Operations Research Letters 10(1), 27–36
(1991)
7. Di Gaspero, L., Rendl, A., Urli, T.: Constraint-based approaches for balancing
bike sharing systems. In: Schulte, C. (ed.) CP 2013. LNCS, vol. 8124, pp. 758–773.
Springer, Heidelberg (2013)
8. IBM ILOG. Cplex. http://www-01.ibm.com/software/integration/optimization/
cplex-optimizer/ (accessed April 2015)
9. Kilby, P., Shaw, P.: Vehicle routing. In: Rossi, F., Beek, P.V., Walsh, T. (eds),
Handbook of Constraint Programming, chapter 23, pp. 799–834. Elsevier (2006)
10. Laporte, G., Nobert, Y., Desrochers, M.: Optimal routing under capacity and dis-
tance restrictions. Operations Research 33(5), 1050–1073 (1985)
11. Miller, C.E., Tucker, A.W., Zemlin, R.A.: Integer programming formulation of
traveling salesman problems. Journal of the ACM (JACM) 7(4), 326–329 (1960)
556 N. Briot et al.

12. Pesant, G., Gendreau, M., Potvin, J.-Y., Rousseau, J.-M.: An exact constraint logic
programming algorithm for the traveling salesman problem with time windows.
Transportation Science 32(1), 12–29 (1998)
13. Prud’homme, C., Fages, J.G., Lorca, X.: Choco3 Documentation. TASC, INRIA
Rennes, LINA CNRS UMR 6241, COSLING S.A.S. (2014)
14. Tisseyre, B., Ojeda, H., Taylor, J.: New technologies and methodologies for site-
specific viticulture. Journal International des Sciences de la Vigne et du Vin 41,
63–76 (2007)
 Constrained Minimum Sum of Squares
Clustering by Constraint Programming

Thi-Bich-Hanh Dao(B) , Khanh-Chuong Duong, and Christel Vrain

University of Orléans, INSA Centre Val de Loire LIFO EA 4022,

F-45067 Orléans, France
{thi-bich-hanh.dao,khanh-chuong.duong,christel.vrain}@univ-orleans.fr

Abstract. The Within-Cluster Sum of Squares (WCSS) is the most

used criterion in cluster analysis. Optimizing this criterion is proved to
be NP-Hard and has been studied by different communities. On the
other hand, Constrained Clustering allowing to integrate previous user
knowledge in the clustering process has received much attention this
last decade. As far as we know, there is a single approach that aims at
finding the optimal solution for the WCSS criterion and that integrates
different kinds of user constraints. This method is based on integer lin-
ear programming and column generation. In this paper, we propose a
global optimization constraint for this criterion and develop a filtering
algorithm. It is integrated in our Constraint Programming general and
declarative framework for Constrained Clustering. Experiments on clas-
sic datasets show that our approach outperforms the exact approach
based on integer linear programming and column generation.

1 Introduction
Cluster analysis is a Data Mining task that aims at partitioning a given set of
objects into homogeneous and/or well-separated subsets, called classes or clus-
ters. It is usually formulated as an optimization problem and different optimiza-
tion criteria have been defined [18]. One of the most used criteria is minimizing the
Within-Cluster Sum of Squares (WCSS) Euclidean distances from each object to
the centroid of the cluster to which it belongs. The well-known k-means algorithm
as well as numerous heuristic algorithms optimize it and find a local optimum [27].
Finding a global optimum for this criterion is a NP-Hard problem and even finding
a good lower bound is difficult [1]. The best exact approach for clustering with this
criterion is based on Integer Linear Programming (ILP) and column generation [2].
On the other hand, since this last decade, user-defined constraints have been
integrated to clustering task to make it more accurate, leading to Constrained
Clustering. User constraints usually make the clustering task harder. The exten-
sion to user constraints is done either by adapting classic algorithms to handle
constraints or by modifying distances between objects to take into account con-
straints [4,9,30]. Recently an exact approach has been proposed, which aims
at finding an optimal solution for the WCSS criterion satisfying all the user
constraints [3]. This approach extends the method based on ILP and column

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 557–573, 2015.
DOI: 10.1007/978-3-319-23219-5 39
558 T.-B.-H. Dao et al.

generation [2]. It integrates different kinds of user constraints but only handle
the WCSS criterion.
Recently general and declarative frameworks for Data Mining have attracted
more and more attention from Constraint Programming (CP) and Data Mining
communities [3,11,16,28]. In our previous work [7,8] we have proposed a gen-
eral framework based on CP for Constrained Clustering. Different from classic
algorithms that are designed for one specific criterion or for some kinds of user
constraints, the framework offers in a unified setting a choice among different
optimization criteria and a combination of different kinds of user constraints. It
allows to find a global optimal solution that satisfies all the constraints if one
exists. Classic heuristic algorithms can quickly find a solution and can scale on
very large datasets, however they do not guarantee the satisfaction of all the con-
straints or the quality of the solution. A declarative and exact approach allows
a better understanding of the data, which is essential for small valuable datasets
that may take years to collect. In order to make the CP approach efficient, it is
important to invest in dedicated global constraints for the clustering tasks [8].
In this paper we propose a global optimization constraint wcss to represent
the WCSS criterion. We propose a method based on dynamic programming to
compute a lower bound and develop a filtering algorithm. The filtering algorithm
filters not only the objective variable, but also the decision variables. The con-
straint extends our framework to the WCSS criterion. Experiments on classic
datasets show that our approach based on CP outperforms the state-of-the-art
approach based on ILP and column generation that handles user constraints [3].

Outline. Section 2 gives preliminaries on constrained clustering and reviews

related work on existing approaches. Section 3 presents the framework based
on CP and introduces the constraint wcss. Section 4 presents our contributions:
the computation of a lower bound and the filtering algorithm for the constraint
wcss. Section 5 is devoted to experiments and comparisons of our approach with
existing approaches. Section 6 discusses perspectives and concludes.

2 Preliminaries

2.1 Constrained Clustering

A clustering task aims at partitioning a given set of objects into clusters, in

such a way that the objects in the same cluster are similar, while being different
from the objects belonging to other clusters. These requirements are usually
expressed by an optimization criterion and the clustering task consists in finding
a partition of objects that optimizes it. Let us consider a dataset of n objects
O and a dissimilarity measure d(o, o ) between any two objects o, o ∈ O (d is
defined by the user). A partition Δ of objects into k classes C1 , . . . , Ck is such
that: (1) ∀c ∈ [1, k]1 , Cc = ∅, (2) ∪c Cc = O and (3) ∀c = c , Cc ∩ Cc = ∅. The
optimization criterion may be, among others:
1
For integer value we use [1, k] to denote the set {1, .., k}.
 Constrained Minimum Sum of Squares Clustering 559

– Maximizing the minimal split between clusters, where the minimal split of
a partition Δ is defined by: S(Δ) = minc=c ∈[1,k] mino∈Cc ,o ∈Cc d(o, o ).
– Minimizing the maximal diameter of clusters, which is defined by:
D(Δ) = maxc∈[1,k] maxo,o ∈Cc d(o, o ).
– Minimizing the Within-Cluster Sum of Dissimilarities, which is defined by:
 1 
W CSD(Δ) = d(o, o )
2 
c∈[1,k] o,o ∈Cc

– Minimizing the Within-Cluster Sum of Squares:

 
W CSS(Δ) = d(o, mc )
c∈[1,k] o∈Cc

where for each c ∈ [1, k], mc is the mean (centroid) of the cluster Cc and
d(o, mc ) is defined by the squared Euclidean distance ||o − mc ||2 . When the
dissimilarity between objects is defined by the squared Euclidean distance
d(o, o ) = ||o − o ||2 , we have [12,18]:

 21 o,o ∈C d(o, o )
c
W CSS(Δ) =
|Cc |
c∈[1,k]

All these criteria except the split criterion are NP-Hard (see e.g. [14] for the
diameter criterion, [1] for WCSS, or concerning WCSD, the weighted max-cut
problem, which is NP-Complete, is an instance with k = 2). Most of classic
algorithms use heuristics and search for a local optimum [17]. For instance the k-
means algorithm finds a local optimum for the WCSS criterion or FPF (Furthest
Point First) [14] for the diameter criterion. Different optima may exist, some may
be closer to the one expected by the user. In order to better model the task, user
knowledge is integrated to clustering task. It is usually expressed by constraints,
leading to Constrained Clustering. User constraints can be cluster-level, giving
requirements on clusters, or instance-level, specifying requirements on pairs of
objects. The last kind, introduced in [29], is the most used. An instance-level
constraint on two objects can be a must-link or a cannot-link constraint, which
states that the objects must be or cannot be in the same cluster, respectively.
Different kinds of cluster-level constraints exist. The minimal (maximal)
capacity constraint requires that each cluster must have at least (at most, resp.)
a given α (β, resp.) objects: ∀c ∈ [1, k], |Cc | ≥ α (resp. ∀c ∈ [1, k], |Cc | ≤ β).
A diameter constraint sets an upper bound γ on the cluster diameter: ∀c ∈
[1, k], ∀o, o ∈ Cc , d(o, o ) ≤ γ. A split constraint, named δ-constraint in [9], sets
a lower bound δ on the split: ∀c = c ∈ [1, k], ∀o ∈ Cc , o ∈ Cc , d(o, o ) ≥ δ. A
density constraint, which extends -constraints in [9], requires that each object
o must have in its neighborhood of radius  at least m objects belonging to the
same cluster as itself: ∀c ∈ [1, k], ∀o ∈ Cc , ∃o1 , .., om ∈ Cc , oi = o ∧ d(o, oi ) ≤ .
User-constraints can make the clustering task easier (e.g. must-link con-
straints) but usually they make the task harder, for instance the split criterion
that is polynomial becomes NP-Hard with cannot-link constraints [10].
560 T.-B.-H. Dao et al.

2.2 Related Work

We are interested in constrained clustering with the WCSS criterion. The clus-
tering task with this criterion is NP-Hard [1]. The popular k-means algorithm
finds a local optimum for this criterion without user constraints. This algo-
rithm starts with a random partition and repeats two steps until a stable state:
(1) compute the centroid of the clusters, (2) reassign the objects, where each
one is assigned to the cluster whose centroid is the closest to the object. The
found solution thus depends on the initial random partition. Numerous heuris-
tic algorithms have been proposed for this criterion, see e.g. [27] for a survey.
Exact methods are much less numerous than heuristic. Some of them are based
on branch-and-bound [6,20] or on dynamic programming [19,23]. An algorithm
based on Integer Linear Programming (ILP) and column generation is proposed
in [22]. This algorithm is improved in [2] and to our knowledge, it is the most
efficient exact approach. It can handle datasets up to 2300 objects with some
heuristics [2], but it does not handle user constraints.
The COP-kmeans algorithm [30] extends the k-means algorithm to handle
must-link and cannot-link constraints and tries to satisfy all of them. This algo-
rithm changes k-means in step (2): it tries to assign each object to the cluster
whose centroid is the closest, without violating the user constraints. If all the pos-
sible assignments violate a user constraint, the algorithm stops. This is a greedy
and fast algorithm without backtrack, but it can fail to find a solution that
satisfies all the user constraints, even when such a solution exists [10]. Another
family of approaches tries to satisfy only most of the user constraints. The CVQE
(Constrained Vector Quantization Error) algorithm [9] or an improved version
LCVQE [24] extend the k-means algorithm to must-link and cannot-link con-
straints by modifying the objective function.
Recently, a general framework, which minimizes the WCSS criterion and
which can handle different kinds of user constraints has been developed [3]. This
framework extends the approach [2] based on ILP and column generation. It
allows to find a global optimum and which satisfies all the user constraints.
Due to the choice of variables in the linear program, the framework is however
specified for this very criterion and does not handle other optimization criteria.
In our previous work, we have developed a general framework based on Con-
straint Programming for Constrained Clustering [7,8]. This framework offers
a choice among different optimization criteria (diameter, split or WCSD) and
integrates all popular kinds of user constraints. Moreover, the capacity of han-
dling different optimization criteria and different kinds of user constraints allows
the framework to be used in bi-criterion constrained clustering tasks [8]. In this
paper we extend this framework to integrate the WCSS criterion.

3 Constraint Programming Model

We are given a collection of n points and a dissimilarity measure between pairs
of points i, j, denoted by d(i, j). Without loss of generality, let us suppose that
points are indexed and named by their index (1 represents the first point).
 Constrained Minimum Sum of Squares Clustering 561

In [8] we have presented a CP model for constrained clustering, which inte-

grates the diameter, split and WCSD criteria. We present an extension of this
model to the WCSS criterion.
In this model the number of clusters k is not set, but only bounded by
kmin and kmax (kmin ≤ k ≤ kmax ). The values kmin and kmax are set by the
user. If the user needs a known number k of clusters, all he has to do is to set
kmin = kmax = k. The clusters are identified by their index, which is a value
from 1 to k for a partition of k clusters. In order to represent the assignment of
points into clusters, we use integer value variables G1 , . . . , Gn , each one having as
domain the set of integers [1, kmax ]. An assignment Gi = c expresses that point
i is assigned to the cluster number c. Let G be [G1 , . . . , Gn ]. To represent the
WCSS criterion, we introduce a float value variable V , with Dom(V ) = [0, ∞).

3.1 Constraints
Any complete assignment of the variables in G defines a partition of points
into clusters. In order to break symmetries between the partitions, we put the
constraint precede(G, [1, . . . , kmax ]) [21], which states that G1 = 1 (the first point
is in cluster 1) and ∀c ∈ [2, kmax ], if there exists Gi = c with i ∈ [2, n], then there
must exists j < i such that Gj = c − 1. Since the domain of the variables Gi is
[1, kmax ], there are at most kmax clusters. The fact that there are at least kmin
clusters means that all the values from 1 to kmin must appear in G. We only
need to require that kmin must appear in G, since with the use of the constraint
precede, if kmin is taken, then kmin −1, kmin −2, . . . , 1 are also taken. This means
#{i | Gi = kmin } ≥ 1 and can be expressed by the constraint: atleast(1, G, kmin ).
Instance-level user constraints can be easily integrated within this model. A
must-link constraint on two points i, j is expressed by Gi = Gj and a cannot-link
constraint by Gi = Gj . All popular cluster-level constraints are also integrated.
For instance, a minimal capacity α constraint is expressed by the fact that each
point must be in a cluster with at least α points including itself. Therefore,
for each i ∈ [1, n], the value taken by Gi must appear at least α times in G,
i.e. #{j | Gj = Gi } ≥ α. For each i ∈ [1, n], we put: atleast(α, G, Gi ). This
requirement allows to infer an upper bound on the number of clusters. Indeed,
Gi ≤ n/α, for all i ∈ [1, n]. For other kinds of user constraints, such as maximal
capacity, diameter or density constraints, we refer the reader to [8].
In order to express that V is the within-cluster sum of squares of the partition
formed by the assignment of variables in G, we develop a global optimization
constraint wcss(G, V, d). The filtering algorithm for this constraint is presented
in Section 4. The value of V is to be minimized.

3.2 Search Strategy

The branching is on the variables in G. A mixed strategy is used with a branch-
and-bound mechanism. A greedy search is used to find the first solution: at each
branching, a variable Gi and a value c ∈ Dom(Gi ) such that the assignment
Gi = c increases V the least are chosen. The value of V at the first found
562 T.-B.-H. Dao et al.

solution gives an upper bound of the domain of V . After finding the first solution,
the search strategy changes. In the new strategy, at each branching, for each
unassigned variable Gi , for each value c ∈ Dom(Gi ) we compute the value aic ,
which is the added amount to V if point i is assigned to cluster c. For each
unassigned variable Gi , let ai = minc∈Dom(Gi ) aic . The value ai thus represents
the minimal added amount to V when point i is assigned to a cluster. Since each
point must be assigned to a cluster, at each branching, the variable Gi with the
greatest value ai is chosen, and for this variable, the value c with the smallest
value aic is chosen. Two branches are then created, corresponding to two cases
where Gi = c and Gi = c. This strategy tends to detect failure more quickly or
in case of success, to find a solution with the value of V as small as possible.

4 Filtering Algorithm for WCSS

We consider that the objects are in an Euclidean space and the dissimilarity
measure is defined by the squared Euclidean distance.  The sum of dissimilarities
of a cluster Cc is defined by W CSD(Cc ) = 12 o,o ∈Cc d(o, o ). The sum of
squares of Cc is defined by W CSS(Cc ) = |C1c | W CSD(Cc ). The WCSS of a

partition Δ = {C1 , . . . , Ck } is W CSS(Δ) = c∈[1,k] W CSS(Cc ).
We introduce a new constraint wcss(G, V, d) and develop a filtering algorithm.
This constraint maintains the relation that the float value variable V is the sum
of squares of the clusters formed by the assignment of the decision variables of
G, given a dissimilarity measure d. Given a partial assignment of some variables
in G, we develop a lower bound computation for V and an algorithm to filter the
domains of the variables. Since the variable V represents the objective function,
this constraint is a global optimization constraint [13,26]. The filtering algorithm
filters not only the domain of the objective variable V , but also the domain of
decision variables in G.
A partial assignment of variables of G represents a case where some points
have been already  assigned to a cluster and there are unassigned points. Let
k = max{c | c ∈ i Dom(Gi )}. The value k is the greatest cluster index among
those remaining in all the domains Dom(Gi ) and thus it is the greatest possible
number of clusters in the partition. Let C be the set of clusters C1 , . . . , Ck .
Some of these clusters can be empty, they correspond to indexes that remain
in some non-singleton domains Dom(Gi ) but not in a singleton domain. For
each cluster Cc , let nc be the number of points already assigned to Cc (nc ≥ 0)
and let S1 (Cc ) be the sum of dissimilarities of all the assigned points in Cc :
S1 (Cc ) = 12 i,j∈Cc d(i, j). Let U be the set of unassigned points and let q = |U |.

4.1 Lower Bound Computation

We compute a lower bound of V considering all the possibilities for assigning all
the points in U into the clusters C1 , . . . , Ck . This is done in two steps:
 Constrained Minimum Sum of Squares Clustering 563

1. For each m ∈ [0, q] and c ∈ [1, k], we compute a lower bound V (Cc , m) of
W CSS(Cc ) considering all possible assignments of m points of U into Cc .
2. For each m ∈ [0, q] and c ∈ [2, k], we compute a lower bound V (C1 . . . Cc , m)
of W CSS({C1 , .., Cc }) considering all the possibilities for assigning any m
points of U into the clusters C1 , . . . , Cc .

Existing branch-and-bound approaches [6,20] are also based on the computation

of a lower bound. However, these algorithms only make use of dissimilarities
between the unassigned points. In our lower bound computation, we exploit not
only dissimilarities between the unassigned points, but also the dissimilarities
between unassigned points and assigned points. The computation is achieved by
Algorithm 1. The two steps are detailed below.

Algorithm 1. Lower bound()

input : a partial assignment of G, a set U = {i | Gi unassigned}, q = |U |
output: a lower bound of the sum of squares V
1 foreach x ∈ U do
2 for c ← 1 to k do
3 if c ∈ Dom(Gx ) then s2 [x, c] ← ∞ ;
4 else s2 [x, c] ← 0;
5 foreach v ∈ [1, n] such that |Dom(Gv )| = 1 do
6 if val(Gv ) ∈ Dom(Gx ) then s2 [x, val(Gv )] = s2 [x, val(Gv )] + d(x, v);
7 sort u ∈ U in the increasing order of d(x, u)
8 s3 [x, 0] ← 0
9 for m ← 1 to q do
10 s3 [x, m] ← s3 [x, m − 1] + d(x, um )/2

11 for c ← 1 to k do
12 for m ← 0 to q do
13 foreach x ∈ U do
14 s[x] = s2 (x, c) + s3 (x, m)
15 sort the arrays increasingly
16 S2 (Cc , m) ← m i=1 s[i]
17 if nc + m = 0 then V (Cc , m) ← 0 ;
18 else
19 V (Cc , m) ← (S1 (Cc ) + S2 (Cc , m))/(nc + m)

20 for c ← 2 to k do
21 V (C1 ..Cc , 0) ← V (C1 ..Cc−1 , 0) + V (Cc , 0)
22 for m ← 1 to q do
23 V (C1 ..Cc , m) ← mini∈[0,m] (V (C1 ..Cc−1 , i) + V (Cc , m − i))

24 return V (C1 ..Ck , q)

564 T.-B.-H. Dao et al.

Assignment of any m points of U into a cluster Cc . If we choose a subset U  ⊆ U

with |U  | = m and we assign the points of U  into the cluster Cc , the sum of
squares of Cc after the assignment will be:

S1 (Cc ) + S2 (Cc , U  )
V (Cc , U  ) =
nc + m
Here S1 (Cc ) is the sum of dissimilarities of the points already assigned in Cc
and S2 (Cc , U  ) is the sum of dissimilarities related to points in U  . The value of
S1 (Cc ) is known. If the set U  is known the value S2 (Cc , U  ) can be computed
exactly by:
 1 
S2 (Cc , U  ) = d(u, v) + d(u, v)

2 
u∈U ,v∈Cc u,v∈U
 
But S2 (Cc , U ) remains unknown while U is not defined. However, for any subset
U  of size m, we can compute a lower bound S2 (Cc , m) as follows. Each point
x ∈ U , in case of assignment to the cluster Cc together with other m − 1 points
of U , will contribute an amount s(x, c, m) = s2 (x, c) + s3 (x, m), where:
– s2 (x, c) represents the sum of dissimilarities between x and the points already
in the cluster Cc . If c ∈ Dom(Gx ) then s2 (x, c) = +∞, since x cannot be
assigned to Cc . Otherwise s2 (x, c) = v∈Cc d(x, v). This value s2 (x, c) is 0
if the cluster Cc is empty. It is computed by lines 2–6 in Algorithm 1.
– s3 (x, m) represents a half of the sum of dissimilarities d(x, z), for all the
m − 1 other points z. These points z can be any points in U , however, if we
order all the points u ∈ U in an increasing order on d(x, u) and we denote
by ui the i-th point in this order, we have a lower bound for s3 (x, m) (lines
m−1
7–10 in Algorithm 1): s3 (x, m) ≥ 12 i=1 d(x, ui ).
A lower bound S2 (Cc , m) is thus the sum of the m smallest contributions
s(x, c, m) (represented by s[x] in Algorithm 1, for fixed values c and m), for
all points x ∈ U . The lower bound V (Cc , m) is 0 if nc + m = 0 or otherwise is
computed by:
S1 (Cc ) + S2 (Cc , m)
V (Cc , m) = (1)
nc + m
This is computed for all c ∈ [1, k] and m ∈ [0, q] (lines 11–19 in Algorithm 1).
For an example, let us consider the partial assignment given in Figure 1 (A)
where k = 3 and some points have been assigned into 3 clusters C1 (square),
C2 (triangle) and C3 (circle). The set of unassigned points U is {3, 4, 8, 9}. A
lower bound V (C3 , 2) for the sum of squares of C3 in case of assignment of any
2 points of U into C3 is computed by Formula (1), with nc = 3 and m = 2.
In this formula, S1 (C3 ) = d(10, 11) + d(11, 12) + d(10, 12) and S2 (C3 , 2) is the
sum of the 2 smallest contributions to C3 among those of all the unassigned
points. They are the contributions of points 4 and 9. Figure 1 (B) presents
the dissimilarities used in the computation of S2 (C3 , 2). For the contribution of
point 4, we make use of one (= m − 1) smallest dissimilarity from point 4 to the
other unassigned points, which is d(4, 3). Idem for point 9, where d(9, 8) is used.
 Constrained Minimum Sum of Squares Clustering 565

5 5
7 7
6 6
1 1

4 4
1/2
3 3
2 2

10 10
8
8 11 11
1/2

(A) 9 12 (B) 9 12

Fig. 1. Example: (A) partial assignment, (B) dissimilarities used in S2 (C3 , 2)

Let us note that the contribution of each point is computed separately, in order
to avoid combinatory cases. Therefore d(4, 9) is not used, even though points 4
and 9 are assumed to be assigned to C3 .

Assignment of any m points of U into c clusters C1 , .., Cc . Any assignment of

m points to c clusters is expressed by an assignment of some i points to the first
c − 1 clusters and the remaining m − i points to the last cluster. Reasoning only
on the number of points to be assigned, we always have the following relation:

V (C1 ..Cc , m) ≥ min (V (C1 ..Cc−1 , i) + V (Cc , m − i))

i∈[0,m]

A lower bound V (C1 ..Cc , m) therefore can be defined by:

V (C1 ..Cc , m) = min (V (C1 ..Cc−1 , i) + V (Cc , m − i)) (2)

i∈[0,m]

This is computed by a dynamic program for all c ∈ [2, k] and m ∈ [0, q] (lines
20–23 in Algorithm 1). Let us notice that with (2), for all c ∈ [1, k] and m ∈ [0, q]:

V (C1 ..Cc , m) = min (V (C1 , m1 ) + · · · + V (Cc , mc )) (3)

m1 +..+mc =m

Let us reconsider the example given in Figure 1. The value V (C1 ..C3 , 4) com-
puted by (2) corresponds to the case V (C1 ..C2 , 1) + V (C3 , 3), i.e when one point
is assigned to the clusters C1 , C2 and 3 points are assigned to C3 . The value
V (C1 ..C2 , 1) corresponds to the case V (C1 , 0) + V (C2 , 1). The value V (C2 , 1)
corresponds to the case where point 4 is assigned to cluster C2 and V (C3 , 3) to
the case where points 4, 8 and 9 are assigned to cluster C3 . We note that in this
lower bound, point 4 is considered twice and point 3 is not considered.
Concerning the complexity, the complexity of the first loop (lines 1–10) is
O(q(k + n + q log q + q)) = O(q 2 log q + qn). The complexity of the second
loop (lines 11–19) is O(kq(q + q log q)) = O(kq 2 log q) and the complexity of
the last loop (lines 20–23) is O(kq 2 ). The complexity of Algorithm 1 is then
O(kq 2 log q + qn). Let us notice that in clustering tasks, the number of clusters
k is usually constant or much smaller than n.
566 T.-B.-H. Dao et al.

4.2 Filtering Algorithm

The filtering algorithm for the constraint wcss(G, V, d) is presented in Algo-

rithm 2, given a partial assignment of variables in G. The value V (C1 ..Ck , q) in
Algorithm 1 represents a lower bound for the sum of squares V , for all possible
assignments of all the points in U into the clusters C1 , . . . , Ck . Let [V.lb, V.ub)
be the actual domain of V , where V.lb is the lower bound, which can be initially
0, and V.ub is the upper bound, which can be either +∞ or the value of V in the
previous solution. The upper bound is strict since in a branch-and-bound search
the next solution must be strictly better than the previous solution. The lower
bound V.lb is then set to max(V.lb, V (C1 ..Ck , q)). We present below the filtering
of unassigned decision variables in G.
For each value c ∈ [1, k], for each unassigned variable Gi , if c ∈ Dom(Gi ),
with the assumption of assigning point i into the cluster Cc , we compute a new
lower bound of V . Let Cc be the cluster Cc ∪ {i} and let C  = {Cl | l = c} ∪ {Cc }.
A new lower bound V  of V is the value V (C  , q − 1), since there still remain
q − 1 points of U \{i} to be assigned to the k clusters. According to (2):

V (C  , q − 1) = min (V (C  \{Cc }, m) + V (Cc , q − 1 − m))

m∈[0,q−1]

For all m ∈ [0, q −1], we revise the lower bounds V (C  \{Cc }, m) and V (Cc , m) by
exploiting informations constructed by Algorithm 1. The revision will be detailed
in the remainder of this subsection. The new lower bound V  is computed by
line 8 of Algorithm 2. Therefore, since Dom(V ) = [V.lb, V.ub), if V  ≥ V.ub, the
variable Gi cannot take the value c. The value c is then removed from Dom(Gi )
(lines 9–10). The complexity of Algorithm 2 is the complexity of computing
the lower bound O(kq 2 log q + qn) plus the complexity of the loop (lines 2–10)
O(kq 2 ). The overall complexity therefore is O(kq 2 log q + qn).

Algorithm 2. Filtering of wcss(G, V, d)

input : a partial assignment of G, a set U = {i | Gi unassigned}, q = |U |
1 V.lb ← max(V.lb, Lower bound())
2 for c ← 1 to k do
3 for m ← 0 to q − 1 do
4 V (C  \{Cc }, m) ← maxm ∈[0,q−m] (V (C, m + m ) − V (Cc , m ))
5 foreach i ∈ U such that c ∈ Dom(Gi ) do
6 for m ← 0 to q − 1 do
7 V (Cc , m) ← ((nc + m)V (Cc , m) + s2 (i, c) + s3 (i, m))/(nc + m + 1)
8 V  ← minm∈[0,q−1] (V (C  \{Cc }, m) + V (Cc , q − 1 − m))
9 if V  ≥ V.ub then
10 remove c from Dom(Gi )
 Constrained Minimum Sum of Squares Clustering 567

Computing V (Cc , m). Let us recall that Cc is the cluster Cc augmented by point
i, and V (Cc , m) is the lower bound of the sum of squares of Cc after adding any
m points of U \{i} into Cc . According to (1):

S1 (Cc ) + S2 (Cc , m)
V (Cc , m) =
nc + 1 + m

We have S1 (Cc ) = S1 (Cc ) + s2 (i, c). The value of S2 (Cc , m) can be revised from
S2 (Cc , m) by:
S2 (Cc , m) = S2 (Cc , m) + s3 (i, m)
According to (1), we have (line 7 of Algorithm 2):

(nc + m)V (Cc , m) + s2 (i, c) + s3 (i, m)

V (Cc , m) =
nc + m + 1

Computing V (C  \{Cc }, m). This value represents a lower bound of the sum of
squares for any assignment of m points in U \{i} into the clusters different from
Cc . According to (3), for all q  ∈ [m, q]:

V (C, q  ) = min (V (C\{Cc }, m) + V (Cc , m ))

m+m =q 

so for all q  ∈ [m, q] and with m + m = q  , we have:

V (C, m + m ) ≤ V (C\{Cc }, m) + V (Cc , m )

which corresponds to:

V (C\{Cc }, m) ≥ V (C, m + m ) − V (Cc , m )

Since m ≤ q  ≤ q and m + m = q  , we have 0 ≤ m ≤ q − m. We then have:

V (C\{Cc }, m) ≥ max (V (C, m + m ) − V (Cc , m ))

m ∈[0,q−m]

We also have:
V (C  \{Cc }, m) ≥ V (C\{Cc }, m)
since C  \{Cc } and C\{Cc } denote the same set of clusters, V (C  \{Cc }, m) is
computed for any m points of U \{i}, while V (C\{Cc }, m) is computed for any
m points of U . Therefore we can exploit the columns computed by the dynamic
program in Algorithm 1 to revise a new lower bound (line 4 in Algorithm 2):

V (C  \{Cc }, m) = max (V (C, m + m ) − V (Cc , m ))

m ∈[0,q−m]
568 T.-B.-H. Dao et al.

5 Experiments
Our model is implemented with Gecode library version 4.2.72 , which supports
float and integer variables. Experiments have been performed on a 3.0 GHz
Core i7 Intel computer with 8 Gb memory under Ubuntu. All our programs are
available at http://cp4clustering.com. We have considered the datasets Iris,
Soybean and Wine from the UCI Machine Learning Repository3 . The number
of objects and the number of classes are respectively 150 and 3 for Iris, 47 and
4 for Soybean and 178 and 3 for Wine dataset. We compare our model with
the approach proposed in [3], based on Integer Linear Programming and column
generation and optimizing WCSS criterion with user constraints. Our approach is
also compared to COP-kmeans [30] that extends k-means algorithm to integrate
user constraints and to Repetitive Branch-and-Bound Algorithm (RBBA) [5],
without user constraints, since this algorithm is not able to handle them.
In MiningZinc, a modeling language for constraint-based mining [15], it is
shown that clustering with the WCSS criterion can be modeled4 . The model can
be translated to different backend solvers including Gecode. However, because of
the intrinsic difficulty of the WCSS criterion, this example model cannot handle
14 points randomly selected from Iris dataset within 30 min, whereas our model
takes 0.01s to solve them.

5.1 Optimizing WCSS in Presence of User Constraints

The most widespread constraints in clustering are must-link or cannot-link con-

straints, since they can be derived from partial previous knowledge (e.g. cluster
labels known for a subset of objects). Therefore we choose these two kinds of
constraints in the experiments. To generate user constraints, pairs of objects are
randomly drawn and either a must-link or a cannot-link constraint is created
depending on whether the objects belong to the same class or not. The process
is repeated until the desired number for each kind of constraints is reached. For
each number of constraints, five different constraint sets are generated for the
tests. In each test, we compute the WCSS value, the Rand index of the solution
compared to the ground truth partition and the total run-time. The Rand index
[25] measures the similarity between two partitions, P and P ∗ . It is defined by
RI = (a + b)/(a + b + c + d), where a and b are the number of pairs of points
for which P and P ∗ are in agreement (a is the number of pairs of points that
are in the same class in P and in P ∗ , b is the number of pairs of points that are
in different classes in P and in P ∗ ), c and d are the number of pairs of points
for which P and P ∗ disagree (same class in P but different classes in P ∗ and
vice versa). This index varies between 0 and 1 and the better the partitions
are in agreement, the closer to 1. Since experiments are performed on 5 sets of
constraints, the mean value μ and the standard deviation σ are computed for
2
http://www.gecode.org
3
http://archive.ics.uci.edu/ml
4
http://inductiveconstraints.eu/miningzinc/examples/kmeans.mzn
 Constrained Minimum Sum of Squares Clustering 569

run-time and RI. A timeout is set to 30 min. A minus sign (-) in the tables
means that the timeout has been reached without completing the search. Since
the ground truth partition is used to generate user constraints, experiments are
done with k equal to the ground truth number of classes for each dataset.
Table 1 gives results for our model (CP) and for the approach based on ILP
and column generation (ILP) [3] for the Iris dataset with different numbers #c
of must-link constraints. For both the execution time and the Rand index, the
mean value of the five tests and the coefficient of variation (σ/μ) are reported.
Since the two approaches are exact, they find partitions having the same WCSS
value. It can be noticed that must-link constraints help to improve quality of the
solution as well as to reduce the execution time for this dataset. Our approach
can find a global optimal without user constraints, whereas ILP approach needs
at least 100 must-link constraints to be able to prove the optimality. With more
than 100 must-link constraints, our approach always takes less time to complete
the search. Our approach is also more efficient to handle must-link and cannot-
link constraints. Table 2 (left) reports mean execution time in seconds for the
five tests and the coefficient of variation σ/μ. In each case, the same number
#c of must-link and cannot-link constraints are added. We can see that when
#c ≤ 75, our approach can complete the search within the timeout and in the
other cases, it performs better than ILP. Concerning the Wine dataset, the two
approaches cannot prove the optimality of the solution in less than 30 min, when
there are less than 150 must-link constraints, as shown in Table 2 (right).
Our CP approach makes better use of cannot-link constraints, as shown in
Table 3. This table reports the mean time in seconds and the percentage of tests
for which each system completes the search within the timeout. The execution
time varies a lot, depending on the constraints. If we consider the Iris database,
Table 3 (left) shows that our model is able to find an optimal solution and to
prove it for roughly 60 % cases, wheres ILP can solve no cases. If we consider the
Wine dataset, Table 3 (right) shows that when 100 must-link and 100 cannot-link
constraints are added, CP can solve all the cases, whereas ILP cannot solve them.
When 125 must-link constraints and 125 cannot-link constraints are added, both
approaches can solve all the cases, but our approach is less time-consuming.
Experiments with the Soybean dataset lead to the same observations. With
a number of must-link constraints varying from 10 to 80, the mean run-times for

Table 1. Time (in seconds) and RI for Iris dataset with #c must-link constraints

#c CP ILP RI
μ σ/μ μ σ/μ μ σ/μ
0 888.99 0.83 % - - 0.879 0%
50 332.06 78.96 % - - 0.940 1.66 %
100 7.09 40.18 % 62 74.24 % 0.978 1.68 %
150 0.31 36.39 % 0.45 44.55 % 0.989 0.66 %
200 0.07 24.83 % 0.11 48.03 % 0.992 0.66 %
250 0.05 10.63 % 0.06 35.56 % 0.996 0.70 %
300 0.04 9.01 % 0.04 19.04 % 0.998 0.35 %
570 T.-B.-H. Dao et al.

Table 2. Time in seconds for Iris dataset with #c must-link and #c cannot-link
constraints (left) and Wine dataset with #c must-link constraints (right)

#c CP ILP
μ σ/μ μ σ/μ
#c CP ILP
25 969.33 51.98 % - -
150 6.84 12.98
50 43.85 46.67 % - -
200 0.11 0.32
75 4.97 150 % - -
250 0.08 0.11
100 0.41 49.8 % 107 72.35 %
300 0.08 0.06
125 0.09 52.07 % 4.4 95.85 %
150 0.06 22.6 % 0.8 50 %

Table 3. Iris dataset with #c cannot-link constraints (left) and Wine dataset with #c
must-link and #c cannot-link constraints (right)

#c CP ILP
μ solved μ solved
#c CP ILP
50 1146.86 20 % - 0%
μ solved μ solved
100 719.53 80 % - 0%
100 10.32 100 % - 0%
150 404.77 60 % - 0%
125 0.35 100 % 497.6 100 %
200 1130.33 40 % - 0%
150 0.12 100 % 13.98 100 %
250 172.81 60 % - 0%
300 743.64 60 % - 0%

both CP and ILP approaches decrease from 0.3 s to 0.01 s. However, with differ-
ent numbers of cannot-link constraints, CP always outperforms ILP approach.
For instance, the mean time is 5.19 s (CP) vs. 278.60 s (ILP) with 20 cannot-link
constraints, or 2.5 s (CP) vs. 126 s (ILP) with 80 cannot-link constraints.

5.2 Comparisons with COP-kmeans and RBBA

RBBA [5] is based on a repetitive branch-and-bound strategy and finds an exact
solution for WCSS. This algorithm takes 0.53 s to find and prove the optimal
solution for the Iris dataset and 35.56 s for the Wine dataset. But it does not
handle user constraints. Concerning the quality of the lower bound, when most
of the points are unassigned, the lower bound of RBBA is better than ours.
However when more points are assigned to the clusters, our lower bound can be
better than the lower bound of RBBA.
On the other hand, COP-kmeans algorithm [30] extends k-means algorithm
to must-link and cannot-link constraints. This algorithm is based on a greedy
strategy to find a solution that satisfies all the constraints. When there are only
must-link constraints, COP-kmeans always finds a partition satisfying all the
constraints, which is a local optimum of WCSS. Nevertheless, when considering
also cannot-link constraints, the algorithm may fail to find a solution satisfying
all the constraints, even when such a solution exists.
 Constrained Minimum Sum of Squares Clustering 571

We perform the same tests, but for each set of constraints, COP-kmeans is
run 1000 times and we report the number of times COP-kmeans has been able to
find a partition. Figure 2 (left) shows the percentage of successes when cannot-
link constraints are added. With the two datasets Iris and Wine, COP-kmeans
fails to find a partition when 150 constraints are added. Our CP model always
find a solution satisfying all the constraints. For Iris dataset, our model succeeds
in proving the optimality for roughly 60 % cases (Table 3 left).
Figure 2 (right) gives the results when #c must-link and #c cannot-link
constraints are added. For Wine dataset, COP-kmeans always fail to find a
partition when #c = 75, 125 or 150. Our CP approach finds a solution satisfying
all the constraints in all the cases. It completes the search in all the cases for
Iris dataset, as shown in Table 2 (left), and in all the cases where #c ≥ 100 for
Wine dataset, as shown in Table 3 (right).

Fig. 2. COP-kmeans with cannot-link (left), with #c must-link and #c cannot-link

constraints (right)

6 Conclusion
In this paper we address the well-known WCSS criterion in cluster analysis.
We develop a new global optimization constraint wcss and present a filtering
algorithm, which filters not only the domain of the objective variable but also
those of decision variables. This constraint integrated to our CP framework [7,
8] extends it to model constrained minimum sum of squares clustering tasks.
Experiments on classic datasets show that our framework outperforms the state-
of-the-art best exact approach, which is based on Integer Linear Programming
and column generation [3].
Working on search strategies and on constraint propagation enables to
improve substantially the efficiency of our CP model. We continue studying
these aspects to make the framework able to deal with larger datasets. We are
working on exploiting user constraints inside the filtering algorithm, either by
using connected components or by modifying the dissimilarities according to the
user constraints. We exploit the flexibility of the CP framework to offer a choice
between exact or approximate solutions, by studying the use of approximate
search strategies, such as local search methods.
572 T.-B.-H. Dao et al.

References
1. Aloise, D., Deshpande, A., Hansen, P., Popat, P.: NP-hardness of Euclidean Sum-
of-squares Clustering. Machine Learning 75(2), 245–248 (2009)
2. Aloise, D., Hansen, P., Liberti, L.: An improved column generation algorithm
for minimum sum-of-squares clustering. Mathematical Programming 131(1–2),
195–220 (2012)
3. Babaki, B., Guns, T., Nijssen, S.: Constrained clustering using column generation.
In: Proceedings of the 11th International Conference on Integration of AI and OR
Techniques in Constraint Programming for Combinatorial Optimization Problems,
pp. 438–454 (2014)
4. Bilenko, M., Basu, S., Mooney, R.J.: Integrating constraints and metric learning
in semi-supervised clustering. In: Proceedings of the 21st International Conference
on Machine Learning, pp. 11–18 (2004)
5. Brusco, M., Stahl, S.: Branch-and-Bound Applications in Combinatorial Data
Analysis (Statistics and Computing). Springer, 1 edn. (2005)
6. Brusco, M.J.: An enhanced branch-and-bound algorithm for a partitioning prob-
lem. British Journal of Mathematical and Statistical Psychology 56(1), 83–92
(2003)
7. Dao, T.B.H., Duong, K.C., Vrain, C.: A Declarative Framework for Constrained
Clustering. In: Proceedings of the European Conference on Machine Learning and
Principles and Practice of Knowledge Discovery in Databases, pp. 419–434 (2013)
8. Dao, T.B.H., Duong, K.C., Vrain, C.: Constrained clustering by constraint pro-
gramming. Artificial Intelligence (2015). doi:10.1016/j.artint.2015.05.006
9. Davidson, I., Ravi, S.S.: Clustering with Constraints: Feasibility Issues and the
k-Means Algorithm. In: Proceedings of the 5th SIAM International Conference on
Data Mining, pp. 138–149 (2005)
10. Davidson, I., Ravi, S.S.: The Complexity of Non-hierarchical Clustering with
Instance and Cluster Level Constraints. Data Mining Knowledge Discovery 14(1),
25–61 (2007)
11. De Raedt, L., Guns, T., Nijssen, S.: Constraint Programming for Data Mining and
Machine Learning. In: Proc. of the 24th AAAI Conference on Artificial Intelligence
(2010)
12. Edwards, A.W.F., Cavalli-Sforza, L.L.: A method for cluster analysis. Biometrics
21(2), 362–375 (1965)
13. Focacci, F., Lodi, A., Milano, M.: Cost-based domain filtering. In: Proceedings of
the 5th International Conference on Principles and Practice of Constraint Pro-
gramming, pp. 189–203 (1999)
14. Gonzalez, T.: Clustering to minimize the maximum intercluster distance. Theoret-
ical Computer Science 38, 293–306 (1985)
15. Guns, T., Dries, A., Tack, G., Nijssen, S., De Raedt, L.: Miningzinc: A modeling
language for constraint-based mining. In: IJCAI (2013)
16. Guns, T., Nijssen, S., De Raedt, L.: k-Pattern set mining under constraints. IEEE
Transactions on Knowledge and Data Engineering 25(2), 402–418 (2013)
17. Han, J., Kamber, M., Pei, J.: Data Mining: Concepts and Techniques, 3rd edn.
Morgan Kaufmann Publishers Inc., San Francisco, CA, USA (2011)
18. Hansen, P., Jaumard, B.: Cluster analysis and mathematical programming. Math-
ematical Programming 79(1–3), 191–215 (1997)
19. Jensen, R.E.: A dynamic programming algorithm for cluster analysis. Journal of
the Operations Research Society of America 7, 1034–1057 (1969)
 Constrained Minimum Sum of Squares Clustering 573

20. Koontz, W.L.G., Narendra, P.M., Fukunaga, K.: A branch and bound clustering
algorithm. IEEE Trans. Comput. 24(9), 908–915 (1975)
21. Law, Y.C., Lee, J.H.M.: Global constraints for integer and set value precedence. In:
Wallace, M. (ed.) Proceedings of the 10th International Conference on Principles
and Practice of Constraint Programming, pp. 362–376 (2004)
22. du Merle, O., Hansen, P., Jaumard, B., Mladenovic, N.: An interior point algorithm
for minimum sum-of-squares clustering. SIAM Journal on Scientific Computing
21(4), 1485–1505 (1999)
23. B.J. van Os, J.M.: Improving Dynamic Programming Strategies for Partitioning.
Journal of Classification (2004)
24. Pelleg, D., Baras, D.: K -Means with Large and Noisy Constraint Sets. In: Kok,
J.N., Koronacki, J., Lopez de Mantaras, R., Matwin, S., Mladenič, D., Skowron,
A. (eds.) ECML 2007. LNCS (LNAI), vol. 4701, pp. 674–682. Springer, Heidelberg
(2007)
25. Rand, W.M.: Objective Criteria for the Evaluation of Clustering Methods. Journal
of the American Statistical Association 66(336), 846–850 (1971)
26. Régin, J.C.: Arc consistency for global cardinality constraints with costs. In: Pro-
ceedings of the 5th International Conference on Principles and Practice of Con-
straint Programming, pp. 390–404 (1999)
27. Steinley, D.: k-means clustering: A half-century synthesis. British Journal of Math-
ematical and Statistical Psychology 59(1), 1–34 (2006)
28. Ugarte Rojas, W., Boizumault, P., Loudni, S., Crémilleux, B., Lepailleur, A.: Min-
ing (Soft-) Skypatterns Using Dynamic CSP. In: Simonis, H. (ed.) CPAIOR 2014.
LNCS, vol. 8451, pp. 71–87. Springer, Heidelberg (2014)
29. Wagstaff, K., Cardie, C.: Clustering with instance-level constraints. In: Proceedings
of the 17th International Conference on Machine Learning, pp. 1103–1110 (2000)
30. Wagstaff, K., Cardie, C., Rogers, S., Schrödl, S.: Constrained K-means Clustering
with Background Knowledge. In: Proceedings of the 18th International Conference
on Machine Learning, pp. 577–584 (2001)
 A Constraint Programming Approach
for Non-preemptive Evacuation Scheduling

Caroline Even1 , Andreas Schutt1,2(B) , and Pascal Van Hentenryck1,3

1
NICTA, Optimisation Research Group, Melbourne, Australia
{caroline.even,andreas.schutt,pascal.hentenryck}@nicta.com.au
http://org.nicta.com.au
2
The University of Melbourne, Parkville, Victoria3010 , Australia
3
Australian National University, Canberra, Australia

Abstract. Large-scale controlled evacuations require emergency ser-

vices to select evacuation routes, decide departure times, and mobilize
resources to issue orders, all under strict time constraints. Existing algo-
rithms almost always allow for preemptive evacuation schedules, which
are less desirable in practice. This paper proposes, for the first time,
a constraint-based scheduling model that optimizes the evacuation flow
rate (number of vehicles sent at regular time intervals) and evacuation
phasing of widely populated areas, while ensuring a non-preemptive evac-
uation for each residential zone. Two optimization objectives are consid-
ered: (1) to maximize the number of evacuees reaching safety and (2)
to minimize the overall duration of the evacuation. Preliminary results
on a set of real-world instances show that the approach can produce,
within a few seconds, a non-preemptive evacuation schedule which is
either optimal or at most 6% away of the optimal preemptive solution.

Keywords: Constraint-based evacuation scheduling · Non-preemptive

scheduling · Phased evacuation · Simultaneous evacuation · Actionable
plan · Real-world operational constraints · Network flow problem

1 Introduction
Evacuation planning is a critical part of the preparation and response to nat-
ural and man-made disasters. Evacuation planning assists evacuation agencies
in mitigating the negative effects of a disaster, such as loss or harm to life, by
providing them guidelines and operational evacuation procedures so that they
can make informed decisions about whether, how and when to evacuate resi-
dents. In the case of controlled evacuations, evacuation agencies instruct each
endangered resident to follow a specific evacuation route at a given departure
time. To communicate this information in a timely fashion, evacuation plan-
ners must design plans which take into account operational constraints arising
NICTA is funded by the Australian Government through the Department of Commu-
nications and the Australian Research Council through the ICT Centre of Excellence
Program.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 574–591, 2015.
DOI: 10.1007/978-3-319-23219-5 40
 A Constraint Programming Approach 575

in actual evacuations. In particular, two critical challenges are the deployment

of enough resources to give precise and timely evacuation instructions to the
endangered population and the compliance of the endangered population to the
evacuation orders. In practice, the control of an evacuation is achieved through
a mobilization process, during which mobilized resources are sent to each resi-
dential area in order to give instructions to endangered people. The number of
mobilized resources determines the overall rate at which evacuees leave. Finally,
to maximize the chances of compliance and success of a controlled evacuation,
the evacuation and mobilization plans must be easy to deploy for evacuation
agencies and should not leave, to the evacuees, uncontrolled alternative routes
that would affect the evacuation negatively.
Surprisingly, local authorities still primarly rely on expert knowledge and sim-
ple heuristics to design and execute evacuation plans, and rarely integrate human
behavioral models in the process. This is partly explained by the limited avail-
ability of approaches producing evacuation plans that follow the current practice.
Apart from a few exceptions [2,5,6,9,11,13] existing evacuation approaches rely
on free-flow models which assume that evacuees can be dynamically routed in
the transportation network [3,10,14]. These free-flow models however violate a
desirable operational constraint in actual evacuation plans, i.e., the fact that all
evacuees in a given residential zone should preferably follow the same evacuation
route.
Recently, a handful of studies considered evacuation plans where each resi-
dential area is assigned a single evacuation path. These studies define both a set
of evacuation routes and a departure schedule. Huibregtse et al. [9] propose a
two-stage algorithm that first generates a set of evacuation routes and feasible
departure times, and then assigns a route and time to each evacuated area using
an ant colony optimization algorithm. In subsequent work, the authors studied
the robustness of the produced solution [7], and strategies to improve the com-
pliance of evacuees [8]. Pillac et al. [13] first introduced the Conflict-based Path
Generation (CPG) approach which was extended to contraflows by Even et al.
[5]. CPG features a master problem which uses paths for each residential node
to schedule the evacuation and a pricing problem which heuristically generates
new paths addressing the conflicts in the evacuation schedule.
These evacuation algorithms however do not guarantee that evacuees will
follow instructions. If the evacuation plan contains forks in the road, evacuees
may decide to change their evacuation routes as the evacuation progresses. This
issue is addressed in [1,6] which propose evacuation plans without forks. The
resulting evacuation plan can be thought of as a forest of evacuation trees where
each tree is rooted at a safe node (representing, say, an evacuation center) and
with residential areas at the leaves. By closing roads or controlling intersections,
these evacuation trees ensure the compliance of the evacuees and avoid conges-
tions induced by drivers slowing down at a fork. Even et al. [6] produce such
convergent evacuation plans by decomposing the evacuation problem in a tree-
design problem and an evacuation scheduling problem. Andreas and Smith [1]
576 C. Even et al.

developed a Benders decomposition algorithm that selects convergent evacuation

routes that are robust to a set of disaster scenarios.
All of the approaches reviewed above allow preemption: The evacuation of a
residential area can be interrupted and restarted arbitrarily. This is not desir-
able in practice, since such schedules will confuse both evacuees and emergency
services, and will be hard to enforce. Disallowing preemption makes the evacua-
tion plans harder to solve for these mathematical programming based approaches
opposing to a constraint programming one. Non-preemptive schedules have been
considered in [4,12] in rather different ways. In [4], a phased evacuation plan
evacuates each area separately, guaranteeing that no vehicles from different
areas travel on a same path at the same time. By definition, phased evacua-
tion does not merge evacuation flows, which is motivated by empirical evidence
that such merging can reduce the road network capacities. The algorithm in [12]
is a column-generation approach for simultaneous evacuation, i.e., evacuations
where multiple paths can share the same road segment at the same time. Each
column represents the combination of a path, a departure time, and a response
curve capturing the behavioral response of evacuees for each evacuation area.
Here the flow rate of each evacuation area is restricted to pre-existing response
curves, and columns are generated individually for each evacuation area. This
column-generation approach requires a discretization of the evacuation horizon.
This paper proposes, for the first time, a constraint programming approach
to generate non-preemptive evacuation schedules. It takes as input a set of evac-
uation routes, which are either chosen by emergency services or computed by an
external algorithm. The constraint-based scheduling model associates a task with
each residential area, uses decision variables for modeling the number of evac-
uees (i.e., the flow ), the number of vehicles to be evacuated per time unit (i.e.,
the flow rate), and the starting time of the area evacuation; It also uses cumu-
lative constraints to model the road capacities. In addition, the paper presents
a decomposition scheme and dominance relationships that decrease the compu-
tational complexity by exploiting the problem structure. Contrary to [12], the
constraint programming model uses a decision variable for the flow rate of each

Fig. 1. Departure times and flows of 7 residential areas of the HN80 instance with the
preemptive algorithm FSP from [6].
 A Constraint Programming Approach 577

evacuation area (instead of a fixed set of values) and avoids discretizing time.
In contrast to [4], the constraint programming model allows for simultaneous
evacuation while satisfying practice-related constraints.
The constraint programming model was applied on a real-life evacuation
case study for the Hawkesbury-Nepean region in New South Wales, Australia.
This region is a massive flood plain protected from a catchment area (the blue
mountains) by the Warragamba dam. A major spill would create damages that
may reach billions of dollars and require the evacuation of about 80,000 peo-
ple. Preliminary experimental results indicate that the constraint programming
model can be used to generate non-preemptive schedules that are almost always
within 5% of the optimal preemptive schedules generated in prior work. These
results hold both for maximizing the number of evacuees for a given time horizon
and for minimizing the clearance time (i.e., the earliest time when everyone is
evacuated). These results are particularly interesting, given that the optimal pre-
emptive solutions produce evacuation plans which are far from practical. Indeed,
Fig. 1 shows the repartition of departure times for seven residential areas in the
original HN80 instance using a preemptive schedule produced by the algorithm
in [6]. Observe how departure times are widely distributed within the scheduling
horizon, indicating that the plan makes heavy use of preemption and is virtually
impossible to implement in practice. Finally, experimental results on the phased
version of the algorithm indicate that phased evacuations are much less time
effective, and should not be the preferred method for short-notice or no-notice
evacuation.
The rest of this paper is organized as follows. Section 2 defines the prob-
lem. Section 3 presents the constraint programming model, including the core
model, the decomposition scheme, the dominance relationships, and the search
procedure. Section 4 presents the experimental results. Section 5 concludes the
paper.

2 Problem Description

The Evacuation Planning Problem (EPP) was introduced by the authors in [11].
It is defined on a directed graph G = (N = E ∪ T ∪ S, A), where E, T , and
S are the set of evacuation, transit, and safe nodes respectively, and A is the
set of edges. The EPP is designed to respond to a disaster scenario, such as a
flood, which may determine a time at which some edges become unavailable.
Each evacuation node k ∈ E is characterized by a number of evacuees dk , while
each arc e is associated with a triple (te , ue , be ), where te is the travel time, ue
is the capacity, and be is the time at which the arc becomes unavailable. We
denote by e.tail (resp. e.head) the tail (resp. head) of an edge e. The problem
is defined over a scheduling horizon H, which depends on the disaster forecast
and the time to mobilize resources. The objective is either (1) to maximize the
total number of evacuees reaching a safe node (for a fixed horizon) or (2) to
minimize the time at which the last evacuee reaches a safe zone (for a variable
horizon). In the following, we assume that the evacuation is carried out using
578 C. Even et al.

private vehicles, but the proposed approach could be adapted to other contexts,
such as building evacuation.
This paper extends the EPP to the non-preemptive (simultaneous) evacua-
tion planning problem (NEPP) and the non-preemptive phased evacuation plan-
ning problem (NPEPP). Both are designed to assist evacuation planners with
the scheduling of fully controlled evacuations. Given a set of evacuation paths,
the NEPP decides the start time, flow, and flow rate at which vehicles are evac-
uating each individual evacuation node, ensuring that the evacuation operates
without interruption. The NEPP allows several evacuation nodes to use the same
road segments at the same time. In contrast, the NPEPP guarantees that no two
evacuation nodes use the same road segment at the same time. The practical
interest of the NPEPP is to evacuate designated priority areas quickly and effi-
ciently, by limiting the risk of any delay caused by slowdown or traffic accidents
which may result from merging traffic.
Formally, an evacuation plan associates with each evacuation area k ∈ E
exactly one evacuation path  pk which is used to route all residents in k to a
same safe node. Let Ωpk = k∈E pk the set of evacuation paths for all evacuations
nodes in E. The characteristics of a path pk are as follows. We denote by Apk
(resp. Npk ) the set of edges (resp. nodes) of pk and by E(e) the set of evacuation
areas whose path contains edge e, i.e., e ∈ Apk . The travel time tk,n between
the evacuation area k and a node n ∈ Npk is equal to the sum of the path edges
travel times separating k from n and tpk is the total travel time between the start
and end of pk . The path capacity upk is the minimum edge capacity of pk . The
last possible departure time lastdep(pk ) along path pk , i.e., the latest time at
which a vehicle can depart on pk without being blocked, is easily derived from
all tk,n (n ∈ Npk ) and the time be at which each path edge e ∈ Apk becomes
unavailable. If none of the path edges e ∈ Apk are cut by the disaster then
lastdep(pk ) = ∞; otherwise lastdep(pk ) = mine∈Apk (be − tk,e.head ). Note
that the latest path departure time only depends on the path characteristics
and not on H.

3 The Constraint Programming Model

The NEPP and NPEPP are two optimization problems whose objective is either
to maximize the number of evacuees or to minimize the overall evacuation clear-
ance time. The key contribution of this paper is to model them as constraint-
based scheduling problems and to use CP to solve them. This modeling avoids
time dicretization and makes it possible to design non-preemptive plans with
variable flow rates. This section presents the constraint-based scheduling models,
including their decision variables, their domains, and the constraints common
to both problems. This section then presents the constraint-based scheduling
models for NEPP and NPEPP.
 A Constraint Programming Approach 579

3.1 Decision Variables

The models associate with each evacuation area k ∈ E the following decision
variables: the total flow of vehicles evacuated flowk (i.e., the number of vehicles
evacuated from area k), the flow rate λflowk representing the number of vehicles
departing per unit of time, the evacuation start time startk (i.e., the time at
which the first vehicle is evacuated from area k), the evacuation end time endk ,
and the total evacuation duration time durk . The last three decision variables
are encapsulated into a task variable taskk which links the evacuation start
time, the evacuation end time, and the evacuation duration and ensures that
startk + durk = endk .
The decision variables range over natural numbers. The flow and flow rates
can only be non-negative and integral since a number of vehicles is a whole
entity. The models use a time step of one minute for flow rates and task variables
which, from an operational standpoint, is a very fine level of granularity: Any
time step of finer granularity would only be too complex to handle in practice.
The domains of the decision variables are defined as follows. The flow variable is
at most equal to the evacuation demand: flowk ∈ [0, dk ] where [a, b] = {v ∈ N |
a ≤ v ≤ b}. The flow-rate variable has an upper bound which is the minimum
of the evacuation demand and the path capacity rounded down to the nearest
integer, i.e., λflowk ∈ [1, min(dk , upk )]. The upper bounds for the evacuation
start time and evacuation end time are the smallest of the scheduling horizon
minus the path travel time, which is rounded up to the nearest integer, and the
latest path departure time, i.e., startk ∈ [0, min(H − tpk , lastdep(pk ))].
The evacuation of an area k can last at most dk minutes assuming the flow rate
is set to one vehicle per minute: durk ∈ [0, dk ]. Note that the lower bound for
duration is zero in order to capture the possibility of not evacuating the area.

3.2 Constraints
The NEPP requires to schedule the flow of evacuees coming from each evacuation
area k on their respective path pk such that, at any instant t, the flow sent on all
paths through the network does not exceed the network edges capacities. These
flow constraints can be expressed in terms of cumulative constraints. Consider
an edge e and the set E(e) of evacuation areas whose evacuation paths use e. For
each evacuation area k ∈ E(e), the model introduces a new task taskek which is
a view over task taskk satisfying:

startek = startk + tk,e.tail , durek = durk , endek = endk + tk,e.tail .

This new task variable accounts for the number of vehicles from evacuation area
k traveling on edge e at any time during the scheduling horizon. Note that tk,e.tail
is computed as the sum of the travel times on each edge, each rounded up to the
next integer for consistency with the domains of the decision variables. While this
approximation may slightly overestimates travel times, it also counterbalances
possible slowdowns in real-life traffic, which are not taken into account in this
model.
580 C. Even et al.

The constraint-based scheduling model for the NEPP introduces the following
cumulative constraint for edge e:

cumulative({(taskek , λflowk ) | k ∈ E(e)}, ue ).

The constraint-based scheduling model for the NPEPP introduces a disjunctive

constraint for edge e instead:

disjunctive({taskek | k ∈ E(e)}). (1)

3.3 The Constraint-Based Scheduling Models

We are now in a position to present a constraint-based scheduling model for

NEPP-MF:

max objective = flowk (2)
k∈E
s.t. k = durk × λflowk
flowub ∀k ∈ E (3)
flowk = min(flowub
k , dk ) ∀k ∈ E (4)
e
cumulative({(taskk , λflowk ) | k ∈ E(e)}, ue ) ∀e ∈ A (5)

The objective (2) maximizes the number of evacuated vehicles. Constraints (3)
and (4) link the flow, flow rate, and evacuation duration together, by ensuring
that the total flow for each area k is the minimum of the evacuation demand
and the flow rate multiplied by the evacuation duration. They use an auxiliary
variable flowub
k denoting an upper bound on the number of vehicles evacuated
from area k. Constraints (5) impose the capacity constraints.
The model NEPP-SAT is the satisfaction problem version of NEPP-MF
where the objective (2) has been removed and the constraint

flowk = dk ∀k ∈ E (6)

has been added to ensure that every vehicle is evacuated.

To minimize clearance time, i.e., to find the minimal scheduling horizon such
that all vehicles are evacuated, it suffices to add the objective function to NEPP-
SAT

min objective = max(endk + tpk ) (7)

k∈E

and to relax the start and end time domains to [0, Hub ] where Hub is an upper
bound on the horizon required to evacuate all vehicles to a shelter. The resulting
constraint-based scheduling model is denoted by NEPP-CT.
A constraint programming formulation NPEPP-MF of the non-preemptive
phased evacuation planning problem can be obtained from NEPP-MF by replac-
ing (5) with (1), which prevents the flows from two distinct origins to travel on
the same edge at the same time. NPEPP-SAT, which is the satisfaction problem
 A Constraint Programming Approach 581

version of NPEPP-MF, is obtained by removing the objective (2) and adding

the constraint (6). NPEPP-CT, which minimizes the evacuation clearance time,
is obtained from NPEPP-SAT by adding the objective (7). Note that since the
flow-rate bounds ensure that edge capacities are always respected in NPEPP, the
flow-rate variable can be directly set to its upper bound to maximize evacuation
efficiency. Hence, the following constraints are added to the NPEPP model:

λflowk = min(dk , upk ) ∀k ∈ E. (8)

3.4 Problem Decomposition

This section shows how the NEPP and NPEPP can be decomposed by under-
standing which paths compete for edges and/or how they relate to each other.
In the following we introduce the path dependency relationship.

Definition 1. Two paths px and py are directly dependent, which is denoted

by px py , if and only if they share at least a common edge, i.e., Apx ∩ Apy = ∅.

Definition 2. Two paths px and pz are indirectly dependent if and only if

¬(px pz ) and there exists a sequence of directly dependent paths py1 , . . . , pyn
such that px py1 , pyn pz and py1 py2 , . . . , pyn−1 pyn .

Definition 3. Two paths px and py are dependent, which is denoted by px py ,

if they are either directly dependent or indirectly dependent. Conversely, paths px
and py are independent, which is denoted by px ⊥py , if they are neither directly
dependent nor indirectly dependent.

Obviously, the path dependency  forms an equivalence relationship, i.e.,  is

reflexive, symmetric, and transitive.
The key idea behind the decomposition is to partition the evacuation areas
E into Υ = {D0 , . . . , Dn } in such a way that any two paths, respectively from
the set of evacuation areas Di and Dj (0 ≤ i < j ≤ n), are independent. As a

(a) Directly dependent paths (b) Indirectly dependent paths

Fig. 2. Illustrating independent nodes and paths.

582 C. Even et al.

result, it is possible to solve the overall model by solving each subproblem Di

(0 ≤ i ≤ n) independently and concurrently. Figure 2a illustrates two sets of
evacuation nodes D0 = {x, y} and D1 = {z} where paths px and py are directly
dependent and there are no indirectly dependent paths. Figure 2b illustrates a
single set of evacuation nodes D = {x, y, z} where the set of paths {px , py } and
{py , pz } are directly dependent, while the set of paths {px ,pz } are indirectly
dependent. We now formalize these concepts.
Definition 4. Let ΩpDki denote the paths of the set of nodes Di . Two sets D0
and D1 of evacuation areas are independent if and only if any two paths from
ΩpDk0 and ΩpDk1 respectively are independent. They are dependent otherwise.
Theorem 1. Let Υ = {D0 , . . . , Dn } be a partition of E such that Di and Dj
(0 ≤ i < j ≤ n) are independent. Then the NEPP and NPEPP can be solved by
concatenating the solutions of their subproblems Di (0 ≤ i ≤ n).
The partition Υ = {D0 , . . . , Dn } can be generated by an algorithm computing
the strongly connected components of a graph. Let GΩpk be the directed graph
u
consisting of the edges and vertices of all paths pk ∈ Ωpk and let GΩ pk
be
its undirected counterpart, i.e., the graph obtained after ignoring the direction
u
of all edges in GΩpk . The strongly connected components of GΩ pk
define the
partition Υ .

3.5 Dominance Relationships

This section shows how to exploit dominance relationships to simplify the
constraint-based scheduling models. The key idea is to recognize that the capac-
ity constraints of some edges are always guaranteed to be satisfied after intro-
ducing constraints on other particular edges.
Definition 5. Let A be a set of edges and e, e ∈ A, e = e . Edge e dominates
e , denoted by e > e , if and only if
– For simultaneous evacuation, the capacity of e is less than or equal to the
capacity of e : ue ≤ ue ;
– The set of paths using e is a subset of the set of paths using e: E(e ) ⊆ E(e);
– For non-convergent evacuation paths, the travel times for evacuation paths
in E(e ) between e and e are the same.
Note that two edges may be dominating each other. For this reason and with-
out loss of generality, this paper breaks ties arbitrarily (e.g., by selecting the
edge closer to a safe node as the dominating edge). Note also that the capacity
condition is ignored for phased evacuation.
Theorem 2. Let A> the set of dominating edges in A. We can safely substitute
A> for A in (5) in NEPP-MF such that the cumulative constraints are only
stated for dominating edges. Similar results hold for NEPP-CT/SAT, and for
the disjunctive constraints in NPEPP-MF/CT/SAT.
 A Constraint Programming Approach 583

3.6 Additional Constraints to a Real-World Evacuation Scheduling

Problem

The flexibility of the constraint-based evacuation scheduling approach allows to

easily include many constraints appearing in real-world evacuation scheduling.
For example, each flow rate variable domain may be restricted to a subset of
values only, in order to account for the number of door-knocking resources avail-
able to schedule the evacuation [12]. Other real-world evacuation constraints
may restrict departure times in order to wait for a more accurate prediction of
an upcoming disaster or to restrict evacuation end times to ensure that the last
vehicle has left a certain amount of time before the disaster strikes the evacuation
area.

3.7 Complexity of Phased Evacuations with Convergent Paths

When using convergent paths, phased evacuations may be solved in polynomial
time.

Theorem 3. Model NPEPP-MF can be solved in polynomial time for conver-

gent paths if all evacuation paths share the same latest completion time at the
safe node.

Proof (Sketch). Using the decomposition method and the dominance criterion,
each subproblem with at least two evacuation paths includes exactly one domi-
nating edge e which is part of every evacuation path. An optimal schedule can
then be obtained in two steps. The first step builds a preemptive schedule by a
sweep over the time starting from the minimal earliest start time of the tasks
on e and ending before the shared completion time. For each point in time,
it schedules one eligible task (if existing) with the largest flow rate (ties are
broken arbitrarily) where a task is eligible if the point in time is not after its
earliest start time on e and it has not been fully scheduled before that time.
Note if a task is almost fully scheduled except the last evacuation batch then
this step considers the actual flow rate of the last batch instead, which may be
smaller than the task flow rate. This preemptive schedule is optimal as for each
point in time, the unscheduled eligible tasks do not have an (actual) greater flow
rate than the scheduled ones. The second step converts this optimal preemptive
schedule to a non-preemptive one by postponing tasks interrupting others until
after the interrupted tasks are completed. This transformation does not change
the flows and hence the non-preemptive schedule has the same objective value
as the preemptive schedule. 


3.8 The Search Procedure

The search procedure considers each unassigned task in turn and assigns its
underlying variables. A task taskk is unassigned if the domain of any of its
variables {startk , durk , endk , flowk , λflowk } has more than one value. The
584 C. Even et al.

search procedure selects an unassigned task taskk and then branches on all its
underlying variables until they all are assigned. Depending on the considered
problem, the models use different heuristics to (1) find the next unassigned task
and to (2) select the next task variable to branch on and the value to assign.
For NPEPP, the flow rate is directly set to the maximum value. The search
strategy is determined by the problem objective as follows. If the objective max-
imizes the number of evacuees for a given scheduling horizon H, the search is
divided into two steps. The first step selects a task with an unfixed duration and
the highest remaining actual flow rate. If the lower bound on duration of the
task is at least two time units less than its maximal duration then a minimal
duration of the maximal duration minus 1 is imposed and a maximal duration
of the maximal duration minus 2 on backtracking. Otherwise the search assigns
duration in decreasing order starting with the largest value in its domain. The
second step selects tasks according to their earliest start time and assigns a start
time in increasing order.1 If the objective is to minimize the horizon H such that
all vehicles are evacuated, then the search selects the next unassigned task with
earliest start time by increasing order among all dominating edges, selecting the
one with maximal flow rate to break ties, and to label the start time in increasing
order.
For NEPP, the different search heuristics are as follows. For the choice (1),
the strategy (1A) randomly selects an unassigned task, and performs geometric
restarts when the number of backtracks equals twice the number of variables in
the model, using a growth factor of 1.5. The strategy (1B) consists in selecting
the next unassigned task in decreasing order of evacuation demand for the dom-
inating edge with the greatest number of tasks. For the choice (2), the strategy
(2A) first labels the flow rate in increasing order, then the task start time also
in increasing order and, finally, the flow in decreasing order. The strategy (2B)
first labels the flow rate in decreasing order, then the flow in decreasing order
again and, finally, the start time in increasing order.

4 Experimental Results
This section reports experiments on a set of instances used in [6]. These instances
are derived from a real-world case study: the evacuation of the Hawkesbury-
Nepean (HN) floodplain. The HN evacuation graph contains 80 evacuated
nodes, 5 safe nodes, 184 transit nodes, 580 edges and 38343 vehicles to evac-
uate. The experimental results also consider a class of instances HN80-Ix
using the HN evacuation graph but a number of evacuees scaled by a factor
x ∈ {1.1, 1.2, 1.4, 1.7, 2.0, 2.5, 3.0} to model population growth. For simplicity,
the experiments did not consider a flood scenario and assume that network edges
are always available within the scheduling horizon H. It is easy to generalize the
results to various flood scenarios.

1
Note that the search procedure does not assume convergent paths or restrictions on
the latest arrival times at the safe node.
 A Constraint Programming Approach 585

Table 1. The strongly connected components associated with each HN80-Ix instance.

Instance #vehicles #scc scc details

HN80 38343 5 {22,9048}, {17,10169}, {14,6490}, {22,9534}, {5,3102}
HN80-I1.1 42183 4 {1,751}, {2,1281}, {40,23656}, {37,16495}
HN80-I1.2 46009 3 {2,1398}, {35,18057}, {43,26554}
HN80-I1.4 53677 5 {2,1631}, {28,16737}, {27,19225}, {4,3824}, {19,12260}
HN80-I1.7 65187 4 {22,16992}, {2,1980}, {42,33240}, {14,12975}
HN80-I2.0 76686 4 {15,13974}, {38,40612}, {2,2330}, {25,19770}
HN80-I2.5 95879 5 {32,36260}, {6,11214}, {16,17983}, {6,9324}, {20,21098}
HN80-I3.0 115029 5 {5,11574}, {12,14184}, {19,23403}, {7,13068}, {29,39651}

For each evacuation instance, a set of convergent evacuation paths was

obtained from the TDFS approach [6]. The TDFS model is a MIP which is
highly scalable as it aggregates edge capacities and abstracts time. The paths
were obtained for the maximization of the number of evacuees within a schedul-
ing horizon of 10 hours, allowing preemptive evacuation scheduling. Thus, for
each instance, the set of evacuation paths can be thought of as forming a forest
where each evacuation tree is rooted at a safe node and each leaf is an evacuated
node. In this particular case, each evacuation tree is a strongly connected compo-
nent. It is important to emphasize that paths are not necessarily the same for the
different HN instances, nor necessarily optimal for non-preemptive scheduling,
which explains some non-monotonic behavior in the results. Table 1 reports the
evacuation paths, the number of vehicles to evacuate (#vehicles), the number
of strongly connected components (#scc), the number of evacuated nodes, the
number of vehicles per scc (scc details) for each HN80-Ix instance. Each strongly
connected component is represented by a pair {x, y} where x is the number of
evacuated nodes and y the number of vehicles to evacuate.
The experimental results compare the flow scheduling results obtained with
the NEPP and NPEPP approaches and the flow scheduling problem (FSP) for-
mulation presented in [6]. The FSP is solved using a LP and it relaxes the
non-preemptive constraints. Indeed, the flow leaving an evacuated node may
be interrupted and restarted subsequently at any time t ∈ H, possibly multiple
times. Moreover, the flow rates in the FSP algorithm are not necessarily constant
over time, giving substantial scheduling flexibility to the FSP but making it very
difficult to implement in practice. Once again, the FSP comes in two versions.
The objective of the core FSP is to maximize the number of vehicles reaching
safety, while the objective of FSP-CT is to minimize the evacuation clearance
time. In order to compare the FSP algorithm and the constraint programming
approaches of this paper fairly, the FSP is solved with a time discretization of 1
minute. The experiments for the NEPP and NPEPP models use different search
586 C. Even et al.

Table 2. Percentage of Vehicles Evacuated with FSP, NEPP-MF/SAT, NPEPP-

MF/SAT.

FSP NEPP-MF/SAT NPEPP-MF/SAT

Instance CPU (s) Perc. Evac. CPU (s) Perc. Evac. Search CPU (s) Perc. Evac.

HN80 0.9 100.0% 0.2 100.0% {1B, 2B} 3.4 96.9%

(SAT)
HN80-I1.1 1.1 100.0% 1538.9 99.2% {1A, 2B} 1295.9 58.4%
HN80-I1.2 1.0 100.0% 0.4 100.0% {1B, 2B} 1444.4s 57.7%
(SAT)
HN80-I1.4 1.3 100.0% 1347.5 99.3% {1A, 2B} 307.0 73.0%
HN80-I1.7 1.8 100.0% 1374.9 97.8% {1A, 2A} 0.3 59.0%
HN80-I2.0 2.0 97.9% 1770.1 93.1% {1A, 2B} 5.9 52.8%
HN80-I2.5 1.8 82.2% 1664.1 79.2% {1A, 2B} 0.1 51.5%
HN80-I3.0 1.4 69.2% 887.2 67.5% {1A, 2B} 0.1 43.1%

heuristics and each experimental run was given a maximal runtime of 1800 sec-
onds per strongly connected component. The results were obtained on 64-bit
machines with 3.1GHz AMD 6-Core Opteron 4334 and 64Gb of RAM and the
scheduling algorithms were implemented using the programming language JAVA
8 and the constraint solver Choco 3.3.0, except for NPEPP-MF where the search
was implemented in ObjectiveCP.

Maximizing the Flow of Evacuees. Table 2 compares, for each HN80-Ix instance
and a 10-hour scheduling horizon, the percentage of vehicles evacuated (Perc.
Evac.) and the solving time in seconds (CPU (s)) with FSP, NEPP-MF/SAT
and NPEPP-MF/SAT. All solutions found with FSP are optimal and are thus
an upper bound on the number of vehicles that can be evacuated with NEPP
and NPEPP. Prior to solving NEPP-MF (resp. NPEPP-MF), the algorithm
attempts to solve NEPP-SAT (resp. NPEPP-SAT) with a 60s time limit and,
when this is successful, the annotation (SAT) is added next to the percentage
of vehicles evacuated. As we make use of decomposition and parallel computing,
the reported CPU for NEPP/NPEPP is the latest of the time at which the best
solution is found among all strongly connected components. The table reports
the best results across the heuristics, i.e., the run where the most vehicles are
evacuated ; for the random strategy, the best result is reported across 10 runs
(note that the standard deviation for the objective value ranges between 0.4%
and 1.1% only across all instances). The search strategy for the best run is shown
in column (Search) as a combination {TaskVar, VarOrder} where TaskVar is the
heuristic for choosing the next task variable and VarOrder is the heuristic for
labeling the task variables.
The results highlight the fact that the constraint-based simultaneous schedul-
ing model finds very high-quality results. On the first five instances, with popu-
lation growth up to 70%, the solutions of NEPP-MF are within 2.2% of the
 A Constraint Programming Approach 587

Fig. 3. Quality of solutions over Time for NEPP-MF.

preemptive bound. This is also the case for the largest instance. In the worst case,
the constraint-based scheduling model is about 4.9% away from the preemptive
bound. It is thus reasonable to conclude that the constraint-based algorithms
may be of significant value to emergency services as they produce realistic plans
for large-scale controlled evacuations. For NPEPP-MF, the solver found optimal
solutions and proved optimality for all instances, except HN80-I1.1 and HN80-
I1.2 for which the best found solution was within 0.1% of the optimal one.2 The
results indicate that a phased evacuation is much less effective in practice and
decreases the number of evacuees reaching safety by up to 40% in many instances.
Unless phased evacuations allow an evacuation of all endangered people, they are
unlikely to be applied in practice, even if they guarantee the absence of traffic
merging.
Figure 3 shows how the quality of solutions improves over time for all HN-
Ix instances which are not completely evacuated, for a particular run. For all
instances, a high-quality solution is found within 10 seconds, which makes the
algorithm applicable to a wide variety of situations. When practical, giving the
algorithm more time may still produce significant benefits: For instance, on HN-
I1.7 the percentage of vehicles increases from 93.0% to 97.6% when the algorithm
is given 1800s. Such improvements are significant in practice since they may be
the difference between life and death.

Profile of the Evacuation Schedules. To demontrate the benefits of NEPP, it

is useful to look at the evacuation profiles produced by the various algorithms.
Recall that Fig. 1 displays a repartition of departure times for seven evacuated
nodes in the original HN80 instance in the optimal solution produced by the FSP
solver. The key observation is that, for several residential areas, the departure
times are widely distributed within the scheduling horizon, indicating that the
FSP solution makes heavy use of preemption. In the FSP solution, the number
of vehicles departing at each time step is often equal to the path capacity. But
2
In our experiments, the problem NPEPP satisfies the condition for Theorem 3. Thus,
these instances can be solved almost instantly using the algorithm outlined in the
proof of Theorem 3.
588 C. Even et al.

Fig. 4. Departure times and flows of 7 evacuated nodes for the NEPP solver.

Table 3. Evacuation clearance time (CT) with FSP-CT, NEPP-CT, and NPEPP-CT.

FSP-CT NEPP-CT NPEPP-CT

Instance CPU (s) CT (min) CPU (s) CT (min) Search CPU (s) CT (min)

HN80 4.3 398 1370.9 409 {1B, 2A} 0.2 680

HN80-I1.1 7.5 582 280.3 616 {1A, 2A} 0.3 1716
HN80-I1.2 5.0 577 6.2 590 {1A, 2B} 0.3 1866
HN80-I1.4 6.1 587 1386.0 614 {1A, 2A} 0.4 1226
HN80-I1.7 7.3 583 1298.6 610 {1A, 2A} 0.3 2307
HN80-I2.0 4.0 625 1713.7 657 {1A, 2B} 0.4 2909
HN80-I2.5 8.4 1092 110.6 1133 {1A, 2B} 0.4 1884
HN80-I3.0 9.4 1232 212.6 1235 {1A, 2B} 0.3 2467

there are also some suprising combinations {evacuated node, time step}, such
as {3, 50}, {3, 84} and {3, 85} where the flow rate is respectively 22, 3, and 12
for evacuation area 3. In summary, the FSP solution is unlikely to be the basis
of a controlled evacuation: It is just too difficult to enforce such a complicated
schedule. Figure 4 shows a repartition of departure times for the same nodes
in the original HN80 instance using the NEPP. The evacuation profile for the
departure times is extremely simple and can easily be the basis of a controlled
evacuation. Its simplicity contrasts with the complexity of the FSP solution and
demonstrates the value of the constraint programming approach promoted in
this paper.

Minimizing the Clearance Time. Table 3 compares, for each HN80-Ix instance,
the minimal clearance time in minutes (CT (min)) found with FSP-CT, NEPP-
CT and NPEPP-CT. All solutions found with FSP-CT and NPEPP-CT are
optimal for the given set of paths. Once again, solutions found by NEPP-CT are
of high-quality and reasonably close to the preemptive lower bound produced
by FSP-CT. In the worst case, the results are within 5.1% of the preemptive
lower bound. The clearance times of the phased evacuations, which are optimal,
are significantly larger than for the NEPP. Note again that paths are differ-
ent between instances and are not necessarily optimal with respect to different
scheduling horizons, which explain inconsistencies such as the horizon found for
HN80-I1.4 being shorter than the horizon found for HN80-I1.2 with NPEPP-CT.
 A Constraint Programming Approach 589

Table 4. Vehicles Evacuated with NEPP-MF with Flow Rates in {2, 6, 10, 15, 20}.

Instance CPU (s) Perc. Evac. Search

HN80 0.4 100.0% (SAT) {1B, 2B}
HN80-I1.1 1538.9 99.2% {1A, 2B}
HN80-I1.2 0.9 100.0% (SAT) {1B, 2B}
HN80-I1.4 986.0 99.5% {1A, 2B}
HN80-I1.7 1289.5 97.1% {1A, 2A}
HN80-I2.0 1614.3 91.0% {1A, 2B}
HN80-I2.5 1784.9 77.0% {1A, 2B}
HN80-I3.0 1558.7 65.6% {1A, 2B}

Table 5. Comparison of FSP and NEPP problem sizes.

FSP-10 FSP-15 NEPP-MF

Instance #cols #rows #cols #rows #vars #ctrs
HN80 44651 145880 68651 218780 1958 2288

The Impact of the Flow Rates. The constraint-based scheduling models have the
flow rates as decision variables, which increases the flexibility of the solutions.
Table 4 studies the benefits of this flexibility and compares the general results
with the case where the flow rates must be selected from a specific set, here
{2, 6, 10, 15, 20}. This is similar to the approach proposed in [12], which uses a
fixed set of response curves and their associated mobilization resources. Note that
the column-generation algorithm in [12] does not produce convergent plans and
discretizes time. The results seem to indicate that flexible flow rates sometimes
bring benefits, especially for the larger instances where the benefits can reach
3.0% ; nonetheless the possible loss when using fixed rates is not substantial and
may capture some practical situations.

Comparison of Model Sizes. One of the benefits of the constraint-based schedul-

ing models is that they do not discretize time and hence are highly scalable
in memory requirements. This is important for large-scale evacuations which
may be scheduled over multiple days. Table 5 compares the FSP problem size
for a scheduling horizon of 10 hours (FSP-10) and 15 hours (FSP-15) with the
NEPP-MF problem size for the HN80 instance, when using 1 minute time steps.
It reports the number of columns (#cols) and the number of rows (#rows) of
the FSP model, as well as the number of variables (#vars) and the number of
constraints (#ctrs) of the NEPP model. As can be seen, the number of vari-
ables and constraints grow quickly for the FSP model and are about 2 orders of
magnitude larger than those in the NEPP-MF model which is time-independent.
590 C. Even et al.

5 Conclusion
This paper proposes, for the first time, several constraint-based models for con-
trolled evacuations that produce practical and actionable evacuation schedules.
These models address several limitations of existing methods, by ensuring non-
preemptive scheduling and satisfying operational evacuation constraints over
mobilization resources. The algorithms are scalable, involve no time discretiza-
tion, and are capable of accommodating side constraints for specific disaster sce-
narios or operational evacuation modes. Moreover, the models have no restriction
on the input set of evacuation paths, which can be convergent or not. Preliminary
experiments show that high-quality solutions, with an objective value close to an
optimal value of an optimal preemptive solution, can be obtained within a few
seconds, and improve over time. Future work will focus on improving the prop-
agation strength of the cumulative constraint for variable durations, flows, and
flow rates, and on generalizing the algorithm for the joint evacuation planning
and scheduling.

References
1. Andreas, A.K., Smith, J.C.: Decomposition algorithms for the design of a nonsi-
multaneous capacitated evacuation tree network. Networks 53(2), 91–103 (2009)
2. Bish, D.R., Sherali, H.D.: Aggregate-level demand management in evacuation plan-
ning. European Journal of Operational Research 224(1), 79–92 (2013)
3. Bretschneider, S., Kimms, A.: Pattern-based evacuation planning for urban areas.
European Journal of Operational Research 216(1), 57–69 (2012)
4. Cepolina, E.M.: Phased evacuation: An optimisation model which takes into
account the capacity drop phenomenon in pedestrian flows. Fire Safety Journal
44, 532–544 (2008)
5. Even, C., Pillac, V., Van Hentenryck, P.: Nicta evacuation planner: Actionable
evacuation plans with contraflows. In: Proceedings of the 20th European Confer-
ence on Artificial Intelligence (ECAI 2014). Frontiers in Artificial Intelligence and
Applications, vol. 263, pp. 1143–1148. IOS Press, Amsterdam (2014)
6. Even, C., Pillac, V., Van Hentenryck, P.: Convergent plans for large-scale evac-
uations. In: Proceedings of the 29th AAAI Conference on Artificial Intelligence
(AAAI 2015) (in press 2015)
7. Huibregtse, O.L., Bliemer, M.C., Hoogendoorn, S.P.: Analysis of near-optimal evac-
uation instructions. Procedia Engineering 3, 189–203 (2010), 1st Conference on
Evacuation Modeling and Management
8. Huibregtse, O.L., Hegyi, A., Hoogendoorn, S.: Blocking roads to increase the evac-
uation efficiency. Journal of Advanced Transportation 46(3), 282–289 (2012)
9. Huibregtse, O.L., Hoogendoorn, S.P., Hegyi, A., Bliemer, M.C.J.: A method to
optimize evacuation instructions. OR Spectrum 33(3), 595–627 (2011)
10. Lim, G.J., Zangeneh, S., Baharnemati, M.R., Assavapokee, T.: A capacitated net-
work flow optimization approach for short notice evacuation planning. European
Journal of Operational Research 223(1), 234–245 (2012)
 A Constraint Programming Approach 591

11. Pillac, V., Even, C., Van Hentenryck, P.: A conflict-based path-generation heuristic
for evacuation planning. Tech. Rep. VRL-7393, NICTA (2013), arXiv:1309.2693,
submitted for publication
12. Pillac, V., Cebrian, M., Van Hentenryck, P.: A column-generation approach for
joint mobilization and evacuation planning. In: International Conference on Inte-
gration of Artificial Intelligence and Operations Research Techniques in Constraint
Programming for Combinatorial Optimization Problems (CPAIOR), Barcelona
(May 2015)
13. Pillac, V., Van Hentenryck, P., Even, C.: A Path-Generation Matheuristic for Large
Scale Evacuation Planning. In: Blesa, M.J., Blum, C., Voß, S. (eds.) HM 2014.
LNCS, vol. 8457, pp. 71–84. Springer, Heidelberg (2014)
14. Richter, K.F., Shi, M., Gan, H.S., Winter, S.: Decentralized evacuation manage-
ment. Transportation Research Part C: Emerging Technologies 31, 1–17 (2013)
Solving Segment Routing Problems with Hybrid
Constraint Programming Techniques

Renaud Hartert(B) , Pierre Schaus, Stefano Vissicchio, and Olivier Bonaventure

UCLouvain, ICTEAM, Place Sainte Barbe 2, 1348 Louvain-la-Neuve, Belgium

{renaud.hartert,pierre.schaus,stefano.vissicchio,
olivier.bonaventure}@uclouvain.be

Abstract. Segment routing is an emerging network technology that

exploits the existence of several paths between a source and a desti-
nation to spread the traffic in a simple and elegant way. The major
commercial network vendors already support segment routing, and sev-
eral Internet actors are ready to use segment routing in their network.
Unfortunately, by changing the way paths are computed, segment rout-
ing poses new optimization problems which cannot be addressed with
previous research contributions. In this paper, we propose a new hybrid
constraint programming framework to solve traffic engineering problems
in segment routing. We introduce a new representation of path variables
which can be seen as a lightweight relaxation of usual representations.
We show how to define and implement fast propagators on these new
variables while reducing the memory impact of classical traffic engineer-
ing models. The efficiency of our approach is confirmed by experiments
on real and artificial networks of big Internet actors.

Keywords: Traffic engineering · Segment routing · Constraint program-

ming · Large neighborhood search

1 Introduction

During the last decades, the Internet has quickly evolved from a small network
mainly used to exchange emails to a large scale critical infrastructure responsible
of significant services including social networks, video streaming, and cloud com-
puting. Simultaneously, Internet Service Providers have faced increasing require-
ments in terms of quality of service to provide to their end-users, e.g., low delays
and high bandwidth. For this reason, controlling the paths followed by traffic has
become an increasingly critical challenge for network operators [22] – especially
those managing large networks. Traffic Engineering – a field at the intersection
of networking, mathematics, and operational research – aims at optimizing net-
work traffic distribution. Among its main objectives, avoiding link overload is
one of the most important as it leads to drop of network reliability, e.g., loss
of packets and increasing delays [1]. New traffic engineering objectives recently
emerged [29]. For instance, a network operator may want specific demands to

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 592–608, 2015.
DOI: 10.1007/978-3-319-23219-5 41
 Solving Segment Routing Problems 593

get through different sequences of network services, e.g., suspect traffic through
a battery of firewalls and high-priority traffic via load-balancer and on low-delay
paths [13].
Segment Routing (SR) [12] has been recently proposed to cope with those
challenges. It is an emerging network architecture that provides enhanced packet
forwarding capabilities while keeping a low configuration impact on networks.
Segment Routing is both an evolution of MPLS (MultiProtocol Label Switch-
ing) [23] and of IPv6 [31]. Many actors of the network industry support segment
routing and several internet service providers will implement segment routing to
manage their networks [12,13,14,28]. All-in-one, segment routing seems to be a
promising technology to solve traffic engineering problems.
The basic idea of Segment Routing is to prepend packets with a stack of
labels, called segments, contained in a segment routing header. A segment rep-
resents an instruction. In this work, we focus on node segments that can be used
to define paths in a weighted graph that represents the network topology. A
node segment contains the unique label of the next router to reach. When such
a router is reached, the current node segment is popped and the packet is sent
to the router referenced by the next segment and so on. Note that segment rout-
ing exploits all the equal cost shortest-paths to reach a given destination. Such
equal cost shortest-paths – called Equal-Cost Multi-Paths (ECMP) paths – are
extremely frequent in network architectures and it is not rare to see as much as
128 different shortest-paths between a source and a destination within a single
network [13]. Fig. 1 illustrates the use of two node segments to define a segment
routing path from router s to router t. Note that the use of segments provides
extreme flexibility in the path selection for different demands.

1) a b 2) a b 3) a b

s c d t s c d t s c d t

SR header SR header

d b packet to t b packet to t packet to t

Fig. 1. A segment routing header with two segments prepended to a packet. First, the
packet is sent to router d using the ECMP path from s to d (assuming unary link
costs). Then, the packet is sent to the next label b following the ECMP path from d to
b. Finally, all the segments have been processed and the original packet is sent to its
destination t using the ECMP path from b to t.

Unfortunately, optimization approaches proposed in the past years do not

consider the enhanced forwarding capabilities of segment routing [12]. Indeed,
they all incur two major limitations. First, they typically focus on the basic
problem of avoiding network congestion while not considering specific additional
requirements [13,29]. Second, all past work assumes network technologies differ-
ent from segment routing. For example, one of the most successful approaches
594 R. Hartert et al.

also used in commercial traffic engineering tools is a tabu search algorithm [18]
proposed by Fortz and Thorup in [15]. However, this method is based on the
assumption that the network paths are computed as shortest paths in the net-
work topology, hence it simply cannot be used to optimize segment routing
networks. Similar considerations also apply to optimization frameworks used for
more flexible protocols, like RSVP-TE [11,20].
In this work, we focus on the traffic placement problem using segment rout-
ing. Precisely, this problem consists in finding an SR-path for each demand such
that the usage of each link is minimized while respecting a set of side constraints
on the demands and the network (§2). We target large networks, like those of
Internet Service Providers, hence both performance and scalability of the opti-
mization techniques are crucial requirements, in addition to good quality of the
solution. We note that classical constraint programming encoding of the prob-
lems generally lead to expensive memory consumption that cannot be sustained
for large networks (several hundreds of nodes) (§3.1). We thus propose a data
structure to encode the domain of our demands that is dedicated to this problem
(§3.2 and §3.3). This encoding has the advantage of reducing the memory con-
sumption of classical constraint programming from O(n4 ) to O(n3 ). We describe
specific and light propagators for constraints defined on top of this new represen-
tation (§4). Our results are finally evaluated on synthetic and real topologies (§5
and §6). They highlight that constraint programming and large neighborhood
search is a winning combination for segment routing traffic engineering.

2 The General Segment Routing Problem

Let us introduce some notations and definitions. A network is a strongly con-
nected directed graph that consists of a set of nodes N (i.e. the routers) and
a set of edges E (i.e. the links). An edge e ∈ E can be represented as the pair
(u, v) where u ∈ N is the source of the edge and v ∈ N is its destination. The
capacity of an edge is denoted capa(e) ∈ N. For every demand d in the set D,
we have an origin src(d) ∈ N, a destination dest(d) ∈ N, and a bandwidth
requirement bw(d) ∈ N. We now introduce the notions of forwarding graph and
segment routing path.
Definition 1 (Forwarding Graph). A forwarding graph describes a flow
between a pair of nodes in the network. Formally, a forwarding graph F G(s, t)
is a non-empty directed acyclic graph rooted in s ∈ N that converges towards
t ∈ N and such that s = t.
Definition 2 (Flow Function). A forwarding graph F G(s, t) is associated
with a flow function flow(s,t) (e, b) → N that returns the amount of the band-
width b ∈ N received at node s that is forwarded to node t through edge e ∈ E
by the forwarding graph. Particularly, flow functions respect the equal spreading
mechanism of ECMP paths. That is, each node of the forwarding graph F G(s, t)
splits its incoming traffic equally on all its outgoing edge contained in F G(s, t).
Flow functions thus respect the flow conservation constraints.
Fig. 2 illustrates the flow function of some ECMP paths.
 Solving Segment Routing Problems 595

ECMP from s to a ECMP from s to d ECMP from s to t

2
a b a b a b
8 4 4 4 2 2

s c d t s c d t s c d t
4 4 4 4 6

Fig. 2. Three different Equal-Cost Multi-Paths used to transfer 8 units of bandwidth

on a network with unary link costs. Each router splits its incoming traffic equally on
its outgoing edges. In this context, the value of flow(s,a) ((a, d), 8) is 0, the value of
flow(s,d) ((a, d), 8) is 4, and the value of flow(s,t) ((a, d), 8) is 2.

Definition 3 (Segment Routing Path). A Segment Routing path (SR-path)

from s ∈ N to t ∈ N is a non-empty sequence of forwarding graphs

F G(s, v1 ), F G(v1 , v2 ), . . . , F G(vi , vi+1 ), . . . , F G(vk , vk−1 ), F G(vk , t)

denoted (s, v1 , . . . , vk , t) and such that the destination of a forwarding graph is the
the source of its successor in the sequence. Also, the source of the first forwarding
graph and the destination of the last forwarding graph respectively correspond to
the source and the destination of the SR-path.

Segment routing paths are illustrated in Fig. 3.

a b a b a b

s c d t s c d t s c d t

FG(s, t) FG(s, a), FG(a, t) FG(s, d), FG(d, b), FG(b, t)

Fig. 3. Three different SR-paths based on the forwarding graphs of Fig. 2. An SR-path
is represented by the sequence of nodes (top) corresponding to the extremities of its
forwarding graphs (bottom).

We can now formalize the problem we want to solve. Let FG be a set of

forwarding graphs on a network such that there is at most one forwarding graph
for each pair of nodes (u, v) ∈ N × N with u = v. Let SR(d) be the SR-path
of demand d ∈ D using the forwarding graph in FG. The General Segment
Routing Problem (GSRP) consists in finding a valid SR-path for each demand
d such that the capacity of each edge is not exceeded
 
∀e ∈ E : flow(i,j) (e, bw(d)) ≤ capa(e) (1)
d∈D F G(i,j)∈SR(d)

and such that a set of constraints on the SR-paths and the network is respected.
596 R. Hartert et al.

Proposition 1. The general segment routing problem is N P-Hard.

Proof. The Partition problem [6] is an N P-complete problem that consists of n
numbers c1 , . . . , cn ∈ N. The question is whether there is a set A ⊆ {1, . . . , n}
such that  
ci = cj
i∈A j∈A

where A is the set of elements not contained in A. This problem can easily be
reduced to the instance of the GSRP depicted in Fig. 4 with all edge capacities
n
fixed to i=1 ci /2. First, consider n demands d1 , . . . , dn from node s to node t
such that bw(di ) = ci . Then, consider that forwarding graphs are defined such
that there are only two possible SR-paths from node s to node t (see Fig. 4).
Finding a valid SR-path for each demand amounts to find a solution to the
Partition problem, i.e., demands having (s, A, t) as SR-path are part of the set
A while the remaining demands are part of the set A.


A A
s t s t

Fig. 4. The N P-complete Partition problem can be encoded as an instance of the

GSRP on this network. Forwarding graphs are defined such that there are only two
possible SR-paths from node s to node t.

Practical considerations. Due to hardware limitations, segment routing headers

usually contain no more than k ∈ [4, 16] segments which limits the number of
forwarding graphs to be contained in an SR-path to k. Furthermore, an SR-path
should not include a loop, i.e., packets should never pass twice on the same link
in the same direction.

3 Segment Routing Path Variables

This section is dedicated to the ad-hoc data structure we use to model the deci-
sion variables of the GSRP. Specialized domain representations are not new in con-
straint programming [34] and several data structures have already been proposed
to represent abstractions such as intervals, sets, and graphs [8,10,16,17].

3.1 Shortcomings of Classical Path Representations

Observe that an SR-path from node s to node t can be seen as a simple path
from s to t in a complete graph on N where each link (u, v) corresponds to the
 Solving Segment Routing Problems 597

forwarding graph F G(u, v). With this consideration in mind, let us motivate the
need of an alternative representation by reviewing classical ways to model path
variables, i.e., path-based, link-based, and node-based representations [39].
In path-based representations, a single variable is associated to each path. The
domain of this variable thus contains all the possible paths to be assigned
to the variable. This representation is usual in column generation frame-
works [2]. In the context of the GSRP, complete path-based representations
have an impracticable memory cost of Θ(|D||N|k ) where k is the maximal
length of the SR-paths.
Link-based representations are surely the most common way to model path
variables [25,42]. The idea is to associate a boolean variable x(d, e) to each
demand d ∈ D and edge e ∈ E. The boolean variable is set to true if edge e is
part of the path, false otherwise. Hence, modeling SR-paths with link-based
representations requires Θ(|D||N|2 ) boolean variables.
Basically, a node-based representation models a path as a sequence of visited
nodes [33]. This could be easily modeled using a discrete variable for each
node that represents the successor of this node. The memory impact of such
representation is Θ(|D||N|2 ).
As mentioned above, memory and computational costs are of practical impor-
tance as we want to solve the GSRP on networks containing several hundreds of
nodes with tens of thousands of demands. In this context, even link-based and
node-based representations suffer from important shortcomings in both aspects.

3.2 Segment Routing Path Variables

Given the impossibility of using classic representations, we propose a new type
of structured domain which we call SR-path variable. An SR-path variable repre-
sents a sequence of visited nodes1 from the source of the path to its destination.
The domain representation of this variable contains (see Fig. 5):
1. A prefix from the source to the destination that represents the sequence of
already visited nodes ;
2. The set of possible nodes, called candidates, to append to the prefix of already
visited nodes.
Basically, the domain of an SR-path variable is the set of all the possible exten-
sions of the already visited nodes followed directly by a node contained in the
set of candidates and finishing at its destination node. An SR-path variable can
thus be seen as a relaxation of classic node-representations. An SR-path variable
is assigned when the last visited node is the path destination. It is considered as
invalid if it is unassigned and if its set of candidates is empty.
The particularity of SR-path variables is that the filtering of the domain is
limited to the direct successor of the last visited node, i.e., the set of candidates
1
Recall that an SR-path is a sequence of forwarding graphs denoted by a sequence of
nodes (see Definition 3).
598 R. Hartert et al.

(see Fig. 5). Hence, no assumption can be made on the part of the sequence
that follows the prefix of already visited nodes and the set of candidates. As a
side effect, visiting a new node c automatically generates a new set of candidates
as assumptions on the successor of c were impossible until now. It is then the
responsibility of constraints to reduce this new set of candidates.

visited a FG (d, a)
a b
b FG (d, b)
s d candidates
c FG (d, c)
s c d t
t FG (d, t)
FG (s, d), . . . FG (s, d), . . .

Fig. 5. A partial SR-path (left) and its corresponding SR-path variable (right). The
SR-path variable maintains the partial sequence of visited nodes from the source s
to the destination t and a set of candidates. Candidates represent the possible direct
successors of the last visited node.

An SR-path variable S supports the following operations:

– visited(S): returns the sequence of visited nodes.
– candidates(S): returns the set of candidates.
– visit(S, c): appends c to the sequence of visited nodes.
– remove(S, c): removes c from the current set of candidates.
– position(S, c): returns the position of c in the sequence of visited nodes.
– isAssigned(S): returns true iff the path has reached its destination.
– length(S): returns an integer variable representing the length of the path.
– src(S), dest(S), last(S): returns the source node, the destination node,
and the last visited node respectively.

3.3 Implementation
We propose to use array-based sparse-sets [5] to implement the internal struc-
ture of SR-path variables. Our data structure is similar to the one proposed
to implement set variables in [8]. The sparse-set-based data structure relies on
two arrays of |N| elements, nodes and map, and two integers V and R. The V
first nodes in nodes correspond to the sequence of visited nodes. The nodes at
positions [V, . . . , R[ in nodes form the set of candidates. The map array maps
each node to its position in nodes. Fig. 6 illustrates this data structure and its
corresponding partial SR-path.
The sparse-set-based implementation of SR-path variables offers several
advantages. First, it is linear in the number of nodes since it only relies on
two arrays and two integers. Also, it implicitly enforces the AllDifferent con-
straint on the sequence of visited nodes. Finally, it allows optimal computational
complexity for all the operations presented in Table 1. Many of these operations
 Solving Segment Routing Problems 599

visited candidates

nodes s d c a b t visited a
0 1 2 3 4 5
b candidates
V R s d
c
map 0 3 4 2 1 5 t
s a b c d t

Fig. 6. An incremental sequence variable implemented in a sparse set. The sequence

of visited nodes is s, d the set of candidates is {a, b, c}. Node t is not a valid candidate.

Table 1. A sparse-set-based implementation offers optimal time-complexities for sev-

eral operations.

Iterate on the sequence of visited nodes Θ(|visited|)

Iterate on the set of candidates Θ(|candidates|)
Iterate on the set of removed candidates Θ(|removed|)
Returns the last visited node in the sequence O(1)
Test if a node has been visited, removed, or is still a candidate O(1)
Returns the position of a visited node in the sequence O(1)
Remove a candidate O(1)
Visit a candidate O(1)

can be implemented trivially by comparing node positions to the value of V

and R. However, the visit and remove operations require more sophisticated
manipulations of the data structure.

Visit a new node. Each time a node is visited, it swaps its position in nodes
with the node at position V. Then, the value of V is incremented to append
the visited node to the sequence of visited nodes. Finally, the value of R is set
to |N| to restore the set of candidates to all the non-visited nodes. The visit
operation is illustrated in Fig. 7. Observe that the sequence of visited nodes
keeps its chronological order.

Remove a candidate. The remove operation is performed in a similar way as the

visit operation. First, the removed node swaps its positions in nodes with the
node at position R. Then, the value of R is decremented to exclude the removed
node from the set of candidates. Fig. 8 illustrates this operation.
Backtracking is achieved in O(1). Indeed, we only need to trail the value of
V to restore the previous sequence of visited nodes. Unfortunately, this efficient
backtracking mechanism cannot restore the previous set of candidates. Never-
theless, this problem could be addressed by one of the following ways:

– The set of candidates could be recomputed on backtrack ;

– Changes in the set of candidates could be trailed during search ;
– Search could be restricted to visit all valid candidates with n-ary branching.
600 R. Hartert et al.

visited candidates visited candidates

nodes s d c a b t nodes s d b a c t
0 1 2 3 4 5 0 1 2 3 4 5
V R V R

map 0 3 4 2 1 5 map 0 3 2 4 1 5
s a b c d t s a b c d t

Fig. 7. State of the internal structure before (left) and after (right) visiting node b.
First, node b exchanges its position with the node at position V in nodes, i.e., c. Then,
V is incremented and R is set to |N|.

visited candidates visited candidates

nodes s d c a b t nodes s d b a c t
0 1 2 3 4 5 0 1 2 3 4 5
V R V R

map 0 3 4 2 1 5 map 0 3 2 4 1 5
s a b c d t s a b c d t

Fig. 8. State of the internal structure before (left) and after (right) removing node c.
First, node c exchanges its position with the node at position R in nodes, i.e., b. Then,
R is decremented.

We chose to apply the third solution as visiting a new node automatically restores
the set of candidates to all the non-visited nodes (by updating the value of R
to |N|). This search mechanism thus allows us to keep backtracking in constant
time since previous sets of candidates do not need to be recovered.

4 Constraints on SR-Path Variables

We model the GSRP by associating an SR-path variable to each demand in D.

These variables are the decisions variables of the problem. Also, each link of the
network e ∈ E is associated with an integer variable load(e) that represents
the total load of this link, i.e., the total amount of traffic routed through e. We
now present some constraints developed on top of the SR-path variables. These
constraints are designed to meet requirements of network operators. Constraints
on SR-path variables are awaken if one of both following events occurs in a
variable of their scope:
– visitEvent(S, a, b) : tells the constraints that node b has been visited just
after node a in the SR-path variable S ;
– removeEvent(S, a, b) : tells the constraints that node b is not a valid candi-
date to be visited after node a in the SR-path variable S.
 Solving Segment Routing Problems 601

We implemented a propagation algorithm specialized for these events in an AC5-

like framework [41].

4.1 The Channeling Constraint

The role of the Channeling constraint is to ensure consistency between an SR-

path variable S and the load of each link in the network. The Channeling
constraint maintains the following property:

∀c ∈ candidates(S), ∀e ∈ F G(last(S), c) :
(2)
load(e) + flow(last(S),c) (e, bw(d)) ≤ capa(e)

The filtering of the Channeling constraint is enforced using two filtering pro-
cedures. The first filtering procedure is triggered each time a new candidate is
visited by an SR-path variable to adjust the load variable of all the links tra-
versed by the visited forwarding graph. Then, the second filtering procedure is
called to reduce the set of candidates by removing forwarding graphs which can-
not accommodate demand d due to the insufficient remaining bandwidth of their
links. This second filtering procedure is also called each time the load variable
of a link is updated.
The Channeling constraint relies intensively on forwarding graphs and flow
functions. In our implementation, we chose to precompute these flow functions
for fast propagation of the Channeling constraint. Such data structures may
have an important impact of O(|N|2 |E|) in memory. Fortunately, this cost can
be reduced to O(|N||E|) by exploiting the shortest-paths DAG2 returned by
Dijkstra’s algorithm.3 Indeed, we can test the presence of an edge (u, v) in
a shortest-path from s to t using the transitive property of shortest-paths as
follows:
dist(s, u) + cost(u, v) + dist(v, t) = dist(s, t). (3)
This property allows one to use shortest-paths DAGs to recompute flow functions
and forwarding graphs in O(|E|) when required.

4.2 The Length Constraint

The Length constraint limits the number of forwarding graphs contained in an

SR-path variable S:
|visited(S)| = length(S). (4)
This constraint aims to respect real network hardware limitations in terms of
maximum number of segments to be appended to packets [12]. We designed a
simple filtering procedure for the Length constraint by comparing the number

2
Shortest-paths DAGs can be easily computed by maintaining several shortest-paths
to a given node in Dijkstra’s algorithm.
3
Recall that we make the assumption that forwarding graphs correspond to ECMP
paths.
602 R. Hartert et al.

of visited nodes to the bounds of the length variable. Let min(length(S)) and
max(length(S)) respectively be the lower bound and the upper bound of the
length variable of S. Changes on the bounds of this variable trigger both following
filtering procedures :
– If |visited(S)| < min(length(S)) − 1, we know that the destination of the
SR-path is not a valid candidate as visiting it next will result in a path that
is smaller than the lower bound of its length variable ;
– If |visited(S)| = max(length(S)) − 1, the destination of the SR-path must
be visited next to respect the upper bound of its length variable.
The Length constraint is also awakened when a new node is visited. In this
case, the filtering procedure first checks if the visited node is the destination
of the path. If it is the case, the length variable is assigned to |visited(S)|.
Otherwise, the lower bound of the length variable is updated to be strictly greater
than |visited(S)| as the path destination still has to be visited.

4.3 The ServiceChaining Constraint

Many operators have lately shown interest for service chaining, i.e., the ability
to force a demand to pass through a sequence of services [29]. The aim of the
ServiceChaining constraint is to force an SR-path variable S to traverse a par-
ticular sequence of nodes described by a sequence of services. Each service is
traversed by visiting one node in a given set, i.e., the set of nodes providing the
corresponding service in the network. Let services = (service1 , . . . , servicek )
denote the sequences of services to be traversed. The filtering rule of the
ServiceChaining constraint enforces the sequence of visited nodes to contain a
non-necessarily contiguous subsequence of services:

visited(S) = . . . , s1 , . . . , s2 , . . . , sk , . . .
(5)
∀i ∈ {1, . . . , k} : si ∈ servicei

The filtering of the ServiceChaining constraint uses similar procedures as those

used by the Length constraint.

4.4 The DAG Constraint

The DAG constraint prevents cycles in the network. A propagator for this con-
straint has already been proposed in [10]. The idea behind this propagator is to
forbid the addition of any link that will result in a cycle in the transitive closure
of the set of edges visited by an SR-path variables. In the context of the GSRP,
the filtering of the DAG constraint can be strengthened by taking in consideration
that an SR-path is a sequence of forwarding graphs which are acyclic by defini-
tion. First, we know that the source (resp. destination) of an SR-path variable
S has no incoming (resp. outgoing) traffic:

F G(u, v) ∈ S : src(S) ∈ nodes(F G(u, v)) (6)

 Solving Segment Routing Problems 603

F G(u, v) ∈ S : dest(S) ∈ nodes(F G(u, v)) (7)

Additional filtering is achieved using the definition of SR-paths and the following
proposition.
Proposition 2. Let c be a node traversed by two forwarding graphs such that c
is not one of the extremities of those forwarding graphs. Then, there is no acyclic
SR-path that visits both forwarding graph.
Proof. Let c be a node traversed by F G(i, j) and F G(u, v) such that c ∈
{i, j, u, v} and that F G(i, j) and F G(u, v) are part of the same SR-path. As
SR-paths are defined as a sequence of forwarding graphs, we know that F G(i, j)
is traversed before F G(u, v) or that F G(u, v) is traversed before F G(i, j). Let us
consider that F G(i, j) is traversed before F G(u, v). In this case, we know that
there is a path from node j to node u. According to the definition of forwarding
graphs, we also know that there is a path from c to j and from u to c. Therefore,
there is a path from c to c which is a cycle. The remaining part of the proof is
done symmetrically.

The DAG constraint thus enforce the following redundant property:
∀{F G(u, v), F G(i, j)} ⊆ S :
(8)
nodes(F G(u, v)) ∩ nodes(F G(i, j)) ⊆ {u, v}.
The filtering procedures of the DAG constraint are triggered each time an SR-path
variable visits a new node. Note that the set of nodes traversed by the visited
forwarding graph of an SR-path variable – necessary to implement these filtering
procedures – can be maintained incrementally with a memory cost of O(|D||N|)
or it can be recomputed when required with a time complexity of O(|N|).

4.5 The MaxCost Constraint

Often, service level agreements imply that some network demands must be
routed on paths with specifics characteristics, e.g., low delays. Such requirements
can easily be enforced using the MaxCost constraint that ensures that the total
cost of an SR-path does not exceed a maximum cost C. To achieve this, we asso-
ciate a positive cost to each forwarding graph F G(u, v). Let cost(F G(u, v)) ∈ N
denote this cost and minCost(s, t) denote the minimum cost of reaching node t
from node s with an SR-path of unlimited length. Such minimum costs could
be easily precomputed using shortest-paths algorithms and only require a space
complexity of Θ(|N|2 ) to be stored. The filtering rule of the MaxCost constraint
(based on the transitive property of shortest-paths) enforces:
∀c ∈ candidates(S) :
(9)
cost(visited(S)) + cost(last(S), c) + minCost(c, dest(S)) ≤ C
where cost(visited(S)) is the total cost of already visited forwarding graphs.
Substantial additional filtering could be added by specializing the minCost
function to also consider the length variable of an SR-path variable. This would
require to pre-compute the all pair shortest-distance for all the k possible lengths
of the path with, for instance, labeling algorithms [3,9].
604 R. Hartert et al.

5 Hybrid Optimization
Finding a first feasible solution of the GSRP can be a difficult task. In this
context, it is often more efficient to relax the capacity constraint of each link
and to minimize the load of the maximum loaded links until respecting the
original capacity constraints.
Frameworks such as Large Neighborhood Search (LNS) [38] have been
shown to be very efficient to optimize large scale industrial problems in many
domains [4,21,26,27,32,37]. The idea behind LNS is to iteratively improve a best-
so-far solution by relaxing some part of this solution. In the context of GSRP, this
could be easily achieved by removing demands that seem inefficiently placed and
to replace them in the network to improve the objective function. The selected
set of removed demands defines the neighborhood to be explored by a branch-
and-bound constraint programming search.
Instead of dealing with all the load variables with a unique aggregation func-
tion (e.g. the maximum load), we designed a specific hybrid optimization scheme
to make the optimization more aggressive. At each iteration, we force a randomly
chosen most loaded link to strictly decrease its load. The load of the remaining
links are allowed to be degraded under the following conditions:
– The load of the most loaded links cannot increase;
– The load of the remaining links are allowed to increase but must remain under
the maximum load, i.e., the set of the most loaded links cannot increase;
– The load of the non-saturated links must remain under their link capacity
to not increase the number of saturated links.
This optimization framework can be seen as particular instantiation of the
variable-objective large neighborhood search framework [36].
Neighborhoods to be explored by constraint programming are generated
using the following procedure [38]:
1. Let Dmax be the set of demands routed through the most loaded links;
2. While fewer than k demands have been selected:
(a) Let A be the array of the not yet selected demands in Dmax sorted by
non-increasing bandwidth;
(b) Generate a random number r ∈ [0, 1[;
(c) Select the rα |A| demand in array A.
The parameter α ∈ [1, ∞[ is used to control the diversification of the selection
procedure. If α = 1 all demands have the same chance to be selected. If α = ∞
the k demands with the highest bandwidth are always selected. The sub-problem
generated by removing the selected demands is then explored by constraint pro-
gramming with a CSPF-like (Constrained Shortest Path First) heuristic [7]:
1. Select the unassigned demand with the largest bandwidth;
2. Try to extend the SR-path variable of this demand as follows:
(a) If the demand destination is a valid candidate, visit the destination to
assign the SR-path variable;
 Solving Segment Routing Problems 605

(b) Otherwise, visit all the candidates sorted by non-decreasing order of their
impact on the load of the most loaded links.
3. If the SR-path variable has no valid candidate, backtrack and reconsider the
previous visit.

We call this heuristic an assign-first heuristic because it tends to assign SR-path

variables as soon as possible.

6 Experiments and Results

We performed many experiments on real and synthetic topologies and demands.

Real topologies are meant to assess the efficiency of our approach on real practical
situations. Conversely, synthetic topologies were used to measure the behavior
of our approach on complex networks. The data set is summarized in Table 2.
Eight of these topologies are large real-world ISP networks and from the publicly
available Rocket Fuel topologies [40]. Synthetic topologies were generated using
the Delaunay’s triangulation algorithm implemented in IGen [30]. Notice that
these synthetic topologies are highly connected from a networking point of view,
i.e., they have an average node degree of 5.72.
Whenever available – that is for topology RealF, RealG, and RealH – we
used the real sets of demands measured by operators. In the other cases, we
generated demands with similar characteristics as the ones in our real data sets
as described in [35]. Finally, we proportionally scaled up all the demands until
a maximum link usage of approximatively 120% was reached.4 This allowed us
to consider critical situations in our analysis.
We measured the efficiency of our approach by minimizing the maximum link
load on each network presented in Table 2. Our experiments were performed on
top of the open-source OscaR Solver [24] with a MacBook Pro configured with
2.6 GHz Intel CPU and 16GB of memory, using a 64-Bit HotSpotTM JVM 1.8.

Table 2. Topologies and results.

Topology |N| |E| |D| Before Relaxation Max load

RealA 79 294 6160 120% 68% 72%
RealB 87 322 7527 142% 72% 82%
RealC 104 302 10695 130% 86% 86%
RealD 141 418 19740 140% 96% 96%
RealE 156 718 23682 121% 76% 78%
RealF 161 656 25486 117% 65% 74%
RealG 198 744 29301 162% 37% 74%
RealH 315 1944 96057 124% 76% 76%
SynthA 50 276 2449 103% 69% 71%
SynthB 100 572 9817 100% 59% 75%
SynthC 200 1044 37881 135% 53% 84%

4
In a solution with all demands routed through their ECMP path.
606 R. Hartert et al.

For each instance, we computed a linear programming lower bound on the

maximum load by solving the linear multi-commodity flow problem on the same
topology and demands with Gurobi 6.0 [19] (see column “Relaxation” of Table 2).
We also show the initial maximum load of the network if no segment is used to
redirect demands through the network (see column “Before” of Table 2). The
results provided by our optimization framework after 60 seconds of computations
are presented in the last column of Table 2. Note that the linear relaxation cannot
be reached in real networks because it assumes that flows can be arbitrarily split,
a function that does not exist in current routers. In practice, even if uneven
splitting became possible in hardware, it would be impossible from an operational
viewpoint to maintain different split ratio for each flow on each router.
As we see, our approach is able to find close to optimal solutions at the
exception of instance RealG. Moreover, the solutions found by our approach are
optimal for instances RealC, RealD, and RealH. This is due to the fact that
these instances contain links that must be traversed by many demands. Fig. 9
illustrates the changes in link loads to highlight such bottleneck links.

258% 130% 124%

90% 90%
86%
76%

90%
75%
20% 22%
34%
15% 1% 0%
SynthB RealC RealH

Fig. 9. Load of the 40 most loaded edges before (left) and after (right) optimization
on a synthetic topology and two real ones. The load of many links have to be reduced
to improve the maximum load on the artificial topologies while only a few have to be
considered on the real ones. Real topologies contain more bottleneck links.

7 Conclusion

This paper presented an hybrid framework to solve segment routing problems

with constraint programming and large neighborhood search. First, we intro-
duced and formalized this new problem encountered by network operators. Then,
we analyzed the shortcomings of classical constraint programming models to
solve this problem on large networks. We thus proposed a new domain structure
for path variable that we called SR-path variable. The particularity of this struc-
ture is that it sacrifices tight domain representation for low memory cost and
fast domain operations. We explained how to implement common requirements
of network operators in terms of constraint on this new variable. The efficiency
of our approach was confirmed on large real-world and synthetic topologies.
 Solving Segment Routing Problems 607

References
1. Awduche, D., Chiu, A., Elwalid, A., Widjaja, I., Xiao, X.: Overview and principles
of internet traffic engineering-rfc3272. IETF (2002)
2. Barnhart, C., Hane, C.A., Vance, P.H.: Using branch-and-price-and-cut to solve
origin-destination integer multicommodity flow problems. Operations Research
48(2), 318–326 (2000)
3. Beasley, J.E., Christofides, N.: An algorithm for the resource constrained shortest
path problem. Networks 19(4), 379–394 (1989)
4. Bent, R., Hentenryck, P.V.: A two-stage hybrid algorithm for pickup and delivery
vehicle routing problems with time windows. Computers & Operations Research
33(4), 875–893 (2006)
5. Briggs, P., Torczon, L.: An efficient representation for sparse sets. ACM Letters on
Programming Languages and Systems (LOPLAS) 2(1–4), 59–69 (1993)
6. Uslar, M., Specht, M., Rohjans, S., Trefke, J., Gonzalez, J.M.V.: Introduction. In:
Uslar, M., Specht, M., Rohjans, S., Trefke, J., Vasquez Gonzalez, J.M. (eds.) The
Common Information Model CIM. POWSYS, vol. 2, pp. 3–48. Springer, Heidelberg
(2012)
7. Davie, B., Rekhter, Y.: MPLS: technology and applications. Morgan Kaufmann
Publishers Inc. (2000)
8. le Clément de Saint-Marcq, V., Schaus, P., Solnon, C., Lecoutre, C.: Sparse-sets
for domain implementation. In: CP workshop on Techniques for Implementing
Constraint programming Systems (TRICS), pp. 1–10 (2013)
9. Desrochers, M., Soumis, F.: A generalized permanent labeling algorithm for the
shortest path problem with time windows. INFOR Information Systems and Oper-
ational Research (1988)
10. Dooms, G., Deville, Y., Dupont, P.E.: CP(Graph): Introducing a Graph Computa-
tion Domain in Constraint Programming. In: van Beek, P. (ed.) CP 2005. LNCS,
vol. 3709, pp. 211–225. Springer, Heidelberg (2005)
11. Figueiredo, G.B., da Fonseca, N.L.S., Monteiro, J.A.S.: A minimum interference
routing algorithm. In: ICC, pp. 1942–1947 (2004)
12. Filsfils, C., et al.: Segment Routing Architecture. Internet draft, draft-filsfils-spring-
segment-routing-00, work in progress (2014)
13. Filsfils, C., et al.: Segment Routing Use Cases. Internet draft, draft-filsfils-spring-
segment-routing-use-cases-00, work in progress (2014)
14. Filsfils, C., et al.: Segment Routing with MPLS data plane. Internet draft, draft-
filsfils-spring-segment-routing-mpls-01, work in progress (2014)
15. Fortz, B., Thorup, M.: Internet traffic engineering by optimizing OSPF weights.
In: Proc. INFOCOM (March 2000)
16. Frei, C., Faltings, B.V.: Resource Allocation in Networks Using Abstraction and
Constraint Satisfaction Techniques. In: Jaffar, J. (ed.) CP 1999. LNCS, vol. 1713,
pp. 204–218. Springer, Heidelberg (1999)
17. Gervet, C.: Conjunto: Constraint logic programming with finite set domains. ILPS
94, 339–358 (1994)
18. Glover, F.: Tabu search: A tutorial. Interfaces 20(4), 74–94 (1990)
19. Inc., Gurobi Optimization. Gurobi optimizer reference manual (2015)
20. Kodialam, M., Lakshman, T.V.: Minimum interference routing with applications to
mpls traffic engineering. In: Proceedings of the Nineteenth Annual Joint Conference
of the IEEE Computer and Communications Societies, INFOCOM 2000, vol. 2,
pp. 884–893. IEEE (2000)
608 R. Hartert et al.

21. Mairy, J.B., Deville, Y., Van Hentenryck, P.: Reinforced adaptive large neighbor-
hood search. In: The Seventeenth International Conference on Principles and Prac-
tice of Constraint Programming (CP 2011), p. 55 (2011)
22. Nucci, A., Papagiannaki, K.: Design, Measurement and Management of Large-
Scale IP Networks - Bridging the Gap between Theory and Practice. Cambridge
University Press (2008)
23. RFC3031 ŒTF. Multiprotocol label switching architecture (2001)
24. OscaR Team. OscaR: Scala in OR (2012). https://bitbucket.org/oscarlib/oscar
25. Ouaja, W., Richards, B.: A hybrid multicommodity routing algorithm for traffic
engineering. Networks 43(3), 125–140 (2004)
26. Pacino, D., Van Hentenryck, P.: Large neighborhood search and adaptive ran-
domized decompositions for flexible jobshop scheduling. In: Proceedings of the
Twenty-Second International Joint Conference on Artificial Intelligence, vol. 3,
pp. 1997–2002. AAAI Press (2011)
27. Shaw, P., Furnon, V.: Propagation Guided Large Neighborhood Search. In: Wal-
lace, M. (ed.) CP 2004. LNCS, vol. 3258, pp. 468–481. Springer, Heidelberg (2004)
28. Previdi, S., et al.: IPv6 Segment Routing Header (SRH). Internet draft, draft-
previdi-6man-segment-routing-header-00, work in progress (2014)
29. Quinn, P., Nadeau, T.: Service function chaining problem statement. draft-ietf-sfc-
problem-statement-07 (work in progress) (2014)
30. Quoitin, B., Van den Schrieck, V., Francois, P., Bonaventure, O.: IGen: Generation
of router-level Internet topologies through network design heuristics. In: ITC (2009)
31. Deering, S.: RFC2460 and R Hinden. Internet protocol
32. Ropke, S., Pisinger, D.: An adaptive large neighborhood search heuristic for the
pickup and delivery problem with time windows. Transportation Science 40(4),
455–472 (2006)
33. Giralt, L.R., Creemers, T., Tourouta, E., Colomer, J.R., et al.: A global constraint
model for integrated routeing and scheduling on a transmission network (2001)
34. Rossi, F., Van Beek, P., Walsh, T.: Handbook of constraint programming. Elsevier
Science (2006)
35. Roughan, M.: Simplifying the synthesis of internet traffic matrices. SIGCOMM
Comput. Commun. Rev. 35(5), 93–96 (2005)
36. Schaus, P.: Variable objective large neighborhood search (Submitted to CP13,
2013)
37. Schaus, P., Hartert, R.: Multi-Objective Large Neighborhood Search. In: Schulte,
C. (ed.) CP 2013. LNCS, vol. 8124, pp. 611–627. Springer, Heidelberg (2013)
38. Shaw, P.: Using Constraint Programming and Local Search Methods to Solve
Vehicle Routing Problems. In: Maher, M.J., Puget, J.-F. (eds.) CP 1998. LNCS,
vol. 1520, pp. 417–431. Springer, Heidelberg (1998)
39. Simonis, H.: Constraint applications in networks. Handbook of Constraint
Programming 2, 875–903 (2006)
40. Spring, N., Mahajan, R., Wetherall, D., Anderson, T.: Measuring isp topologies
with rocketfuel. IEEE/ACM Trans. Netw., 12(1) (2004)
41. Van Hentenryck, P., Deville, Y., Teng, C.-M.: A generic arc-consistency algorithm
and its specializations. Artificial Intelligence 57(2), 291–321 (1992)
42. Xia, Q., Simonis, H.: Primary/Secondary Path Generation Problem: Reformula-
tion, Solutions and Comparisons. In: Lorenz, P., Dini, P. (eds.) ICN 2005. LNCS,
vol. 3420, pp. 611–619. Springer, Heidelberg (2005)
 Modeling Universal Instruction Selection

Gabriel Hjort Blindell1(B) , Roberto Castañeda Lozano1,2 , Mats Carlsson2 ,

and Christian Schulte1,2
1
SCALE, School of ICT, KTH Royal Institute of Technology, Stockholm, Sweden
{ghb,cschulte}@kth.se
2
SCALE, Swedish Institute of Computer Science, Kista, Sweden
{rcas,matsc}@sics.se

Abstract. Instruction selection implements a program under compi-

lation by selecting processor instructions and has tremendous impact
on the performance of the code generated by a compiler. This paper
introduces a graph-based universal representation that unifies data and
control flow for both programs and processor instructions. The repre-
sentation is the essential prerequisite for a constraint model for instruc-
tion selection introduced in this paper. The model is demonstrated to
be expressive in that it supports many processor features that are out of
reach of state-of-the-art approaches, such as advanced branching instruc-
tions, multiple register banks, and SIMD instructions. The resulting
model can be solved for small to medium size input programs and sophis-
ticated processor instructions and is competitive with LLVM in code
quality. Model and representation are significant due to their expres-
siveness and their potential to be combined with models for other code
generation tasks.

1 Introduction

Instruction selection implements an input program under compilation by select-

ing instructions from a given processor. It is a crucial part of code generation in
a compiler and has been actively researched for over four decades (see [23] for a
recent survey). It is typically decomposed into identifying the applicable instruc-
tions and selecting a combination of applicable instructions to meet the semantics
of the input program. Combinations differ in efficiency and hence selecting one
is naturally an optimization problem. Finding an efficient combination is crucial
as efficiency might differ by up to two orders of magnitude [35].
Common approaches use graph-based techniques that operate on the data-
flow graph of a program (nodes represent operations, edges describe data flow).
However, state-of-the-art approaches are restricted to trees or DAGs to avoid
NP-hard methods for general graphs. This restriction is severe: ad-hoc routines
are needed for handling control flow; many instructions of modern processors,
such as DSPs (digital signal processors), cannot be handled; and the scope of
instruction selection is typically local to tiny parts of the input program and
hence by design forsakes crucial optimization opportunities.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 609–626, 2015.
DOI: 10.1007/978-3-319-23219-5 42
610 G. Hjort Blindell et al.

This paper introduces a universal representation based on general graphs. It

is universal as it simultaneously captures both data and control flow for both
programs and processor instructions. By that it overcomes the restrictions of
current approaches. In particular, the universal representation enables a simple
treatment of global code motion, which lets the selection of instructions to be
global for an entire function. The representation is compatible with state-of-the-
art compilers; the paper uses LLVM [26] as compiler infrastructure.
However, the very reason of the proposed approach is that instruction selec-
tion can be expressed as a constraint model. Due to the expressiveness of the
universal representation, the introduced model accurately reflects the interac-
tion between control and data flow of both programs and processor instructions.
The paper presents the model in detail and discusses how it supports processor
features that are out of reach of state-of-the-art approaches, such as advanced
branching instructions, multiple register banks, and SIMD instructions.
The paper shows that the described approach is feasible. The resulting model
can be solved for small to medium size input programs and challenging SIMD
processor instructions and is competitive with LLVM in code quality.
Model and representation are significant for two reasons. First, they are con-
siderably more powerful than the state of the art for instruction selection and can
capture common features in modern processors. Crucially, instruction selection
with a universal representation is only feasible with an approach as expressive
as a constraint model. Second, the paper’s approach will be essential to inte-
grate instruction selection with register allocation and instruction scheduling as
the other code generation tasks that we have explored in previous work [10,11].
It is only the combination of all three interdependent tasks that will enable
generating optimal code for modern processors.

Paper outline. Section 2 introduces graph-based instruction selection and Sect. 3

introduces the representations that enable universal instruction selection. The
corresponding constraint model is introduced in Sect. 4. Section 5 experimentally
evaluates the paper’s approach followed by a discussion of related work in Sect. 6.
Section 7 concludes the paper.

2 Graph-Based Instruction Selection

The most common approach for instruction selection is to apply graph-based
methods. As is common, the unit of compilation is a single program function,
which consists of a set of basic blocks. A basic block, or just block, is a sequence
of computations (like an addition or memory load) and typically ends with a
control procedure (like a jump or function return). Each block has a single entry
point and a single exit point for execution. A program function has exactly one
block as entry point, called entry block.
For a program function a data-flow graph called program graph is constructed,
where each node represents a computation and each edge indicates that one
computation uses the value produced by another computation. As a data-flow
 Modeling Universal Instruction Selection 611

graph does not incorporate control-flow information, a single program function

typically results in several program graphs (at least one per block). A local
instruction selector selects instructions for program graphs of a single block,
whereas a global instruction selector does so for an entire function.
For each instruction of a given processor data-flow graphs called pattern
graphs are also constructed. The set of all pattern graphs for a processor consti-
tute a pattern set. Thus, the problem of identifying the applicable instructions
is reduced to finding all instances where a pattern graph from the pattern set is
subgraph isomorphic to the program graph. Such an instance is called a match.
A set of matches covers a program graph if each node in the program graph
appears in exactly one match. By assigning a cost to each match that corre-
sponds to the cost of the selected instruction, the problem of finding the best
combination of instructions is thus reduced to finding a cover with least cost. It
is well known that the subgraph isomorphism problem and the graph covering
problem are NP-complete in general [8,20], but can be solved optimally in linear
time if the program and pattern graphs are trees [2].
Program trees are typically not expressive enough and hence modern compil-
ers commonly use pattern trees and program DAGs, which are then covered using
greedy heuristics. This, however, suffers from several limitations. First, pattern
trees significantly limit the range of instructions that can be handled. For exam-
ple, ad-hoc routines are required to handle even simple branch instructions as
pattern trees cannot include control procedures. Second, program DAGs exclude
handling of sophisticated instructions of modern processors that span multiple
blocks. Third, more efficient code can be generated if certain computations can
be moved across blocks, provided the program semantics is kept.

i = 0;
(i < N) {
c = A [ i ] + B [ i ];
( MAX < c )
c = MAX ;
C[i] = c;
i ++;
}
(a) C code

Fig. 1. An example program that computes the saturated sums of two arrays, where
A, B, and C are integer arrays of equal lengths and stored in memory, and N and MAX
are integer constants representing the array length and the upper limit, respectively.
An int value is assumed to be 4 bytes.
612 G. Hjort Blindell et al.

A Motivating Example. Assume the C program shown in Fig. 1a, which

computes the saturated sums of two arrays. Saturation arithmetic “clamps”
the computed value such that it stays within a specified range and does not
wrap around in case of overflow, a property commonly desired in many digital
signal processing applications. For simplicity the program in Fig. 1a only clamps
on the upper bound, but the following argumentation can easily be extended
to programs that clamp both bounds. Most compilers do not operate directly
on source code but on the internal representation (IR) of a program, shown
in Fig. 1b, which can be viewed as a high-level assembly language. Programs
written in this form are typically portrayed as a control-flow graph where each
node represents a block and each edge represents a potential jump from one
block to another. Most modern compilers, however, employ a slightly different
representation that will be discussed in Sect. 3.
Assume further that this program will be executed on a processor whose
instruction set includes the following instructions: satadd computes the saturated
sum of two integer values; repeat iteratively executes an instruction sequence a
given number of times; and add4 can compute up to four ordinary integer sums
simultaneously (a so-called vector or SIMD (Single-Input Multiple-Data) instruc-
tion). Clearly, this program would benefit from selecting the satadd instruction
to compute the value c and from selecting the repeat instruction to implement
the control of the loop consisting of blocks bb2 through bb5. What might be less
obvious, however, is the opportunity to select the add4 instruction to compute
values t2 and t3 together with t4 and i . Since these additions all reside
inside the loop and are independent from one another, they can be computed
in parallel provided they can be performed in the same block. This notion of
moving computations across blocks is referred to as global code motion.
But taking advantage of these instructions is difficult. First, describing the
saturated sum involves computations and control procedures that span several
blocks. However, the state of the art in instruction selection is limited to local
instruction selection or cannot handle control procedures. Second, to maximize
the utility of the add4 instruction the additions for producing values t4 and i
must be moved from bb5 to bb3. But it is not known how to perform global code
motion in conjunction with instruction selection, and hence all existing represen-
tations that describe entire program functions inhibit moves by pre-placing each
computation to a specific block. Third, for many processors the registers used by
vector instructions are different from those of other instructions. Consequently,
selecting the add4 instruction might necessitate further instructions for copying
data between registers. These additional instructions can negate the gain of using
the add4 instruction or, in the worst case, even degrade the quality of the gener-
ated code. Making judicious use of such instructions therefore requires that the
instruction selector is aware of this overhead. Fourth, since the program graph
must be covered exactly, the computation of the value i cannot be implemented
by both the add4 and repeat instructions. Instruction selection must therefore
evaluate which of the two will result in the most efficient code, which depends
on their relative costs and the restrictions imposed by the processor.
 Modeling Universal Instruction Selection 613

Fig. 2. The C program from Fig. 1a in SSA form.

3 Representations for Universal Instruction Selection

This section introduces representations for both programs and instructions and
how they can be used to express covering and global code motion.

Program Representation. The key idea is to combine both data flow and
control flow in the very same representation. We start by modifying control-flow
graphs such that the nodes representing blocks no longer contain any compu-
tations. We refer to these as block nodes. In addition, the control procedures
previously found within the blocks now appear as separate control nodes, where
each control node has exactly one inbound control-flow edge indicating to which
block the procedure belongs. Consequently, the control-flow edges that previ-
ously originated from the block nodes now originate from the control nodes. An
example will be provided shortly.
To capture the data flow of entire program functions as a data-flow graph
we use the SSA (Static Single Assignment) graph constructed from programs in
SSA form. SSA is a state-of-the-art program representation where each program
variable must be defined only once [14]. When the definition depends on the
control flow, SSA uses so-called ϕ-functions to disambiguate such cases by taking
a value and the block from where it originates as arguments. Fig. 2 shows the
C program from Fig. 1a in SSA form, and the corresponding control-flow and
SSA graphs are shown in Fig. 3a and Fig. 3b, respectively. Originally the SSA
graph only consists of nodes representing computations – called computation
nodes – but we extend it such that each value is represented by a value node.
Also note that copying of values is not represented as separate computation
nodes in the SSA graph. Therefore, in Fig. 3b the program variables i1 and c2
have been replaced by 0 and MAX , respectively.
To unify the control-flow graph and the SSA graph, we first add data-flow
edges from value nodes to the control nodes that make use of the corresponding
values. For the control-flow graph shown in Fig. 3a this entails the c.br nodes,
which represent conditional jumps. Note that the SSA graph does not indicate
in which block a given computation should be performed. Although we want
to keep such pre-placements to a minimum, having no pre-placements permits
moves that violate the program semantics. For example, in Fig. 1b the assignment
c = MAX must be performed in block bb4 as its execution depends on whether
614 G. Hjort Blindell et al.

bb1 0
ϕ
br
N i2 4

bb2 > ∗

A t1 B
c.br end
F + +
T
bb3 t2 t3

ld ld
c.br bb4
T
a b
F
br + MAX

MAX c1 C 1
bb5
< ϕ + +

br c3 t4 i3
st

(a) Control-flow graph (b) SSA graph

Fig. 3. The program graph constructed from the program shown in Fig. 2. Thick-lined
diamonds, boxes, and arrows represent control nodes, block nodes, and control-flow
edges, respectively. Thin-lined circles, boxes, and arrows represent computation nodes,
value nodes, and data-flow edges, respectively. Dotted lines represent definition edges.

MAX < c holds. Fortunately, these cases can be detected whenever a ϕ-function
appears in the program. The SSA-based program in Fig. 2, for example, has in
block bb2 a statement i2 = ϕ(i1 :bb1, i3 :bb5). Thus, the program variable i2
is assigned either the value i1 or the value i3 depending on whether the jump
to bb2 was made from block bb1 or block bb5. These conditions can be ensured
to hold in the generated code by requiring that (i) due to the arguments to the
ϕ-function, the values i1 and i3 must be computed in blocks bb1 and bb5,
respectively; and (ii) due to the location of the ϕ-function in the program, the
value i2 must be assigned its value in block bb2. We encode these constraints
into the program graph by introducing definition edges to signify that a certain
value must be produced in a specific block. Hence, in case of the example above
three definition edges are added: one from value node 0 to block node bb1, one
from value node i3 to block node bb5, and another from value node i2 to block
node bb2. Such edges are also added for the statement c3 = ϕ(c1 :bb3, c2 :bb4),
which results in the program graph shown in Fig. 3.
 Modeling Universal Instruction Selection 615

Instruction Representation. The procedure entry

for constructing pattern graphs is almost iden- +
tical to constructing program graphs. The only T
c.br if
exception is that the control-flow graph of a pat- v x1 v
tern graph could be empty, which is the case for F < ϕ
br
instructions whose result does not depend on
any control flow. For example, Fig. 4 shows the x2
pattern graph of satadd (introduced in Sect. 2), end
which has a substantial control-flow graph. In
comparison, the pattern graph of a regular add Fig. 4. satadd’s pattern graph.
instruction would only comprise an SSA graph,
consisting of one computation node and three value nodes. Like with the pro-
gram graph, it is assumed that all pattern graphs with a non-empty control-flow
part have exactly one block node representing the instruction’s entry point.
Since the program graph now consists of several kinds of nodes, we need to
refine the notion of coverage. A program graph is covered by a set of matches if
each operation in the program graph appears in exactly one match from the set,
where an operation is either a computation or a control node. Likewise, a match
covers the operations in the program graph corresponding to the operations in
the pattern graph. Consequently, matches are allowed to partially overlap on the
block and value nodes in the program graph. This property is deliberate as it
enables several useful features, which will be seen shortly.
But not all covers of a given program graph yield valid solutions to the
global code motion problem. For example, assume a cover of Fig. 3 where a
match corresponding to the add4 instruction has been selected to cover the
computation nodes that produce values t2 , t3 , and i3 . Because t2 and t3
are data-dependent on value i2 , which must be produced in block bb2 due to
a definition edge, these values cannot be produced earlier than in bb2. Likewise,
because value c1 must be produced in block bb3 and is data-dependent on t2
and t3 , these values cannot be produced later than in bb3. At the same time,
i3 must be produced in block bb5. Hence no single instruction that computes
t2 , t3 , and i3 can be placed in a block such that all conditions imposed by
the program graph are fulfilled.
We use the above observation to formalize the global code motion problem
as follows. If a datum refers to a particular value node in the program graph, a
match m defines respectively uses a datum d if there exists an inbound respec-
tively outbound data-flow edge to d in the pattern graph of m. Hence a datum
can be both defined and used by the same match. We also refer to the data used
but not defined by a match as its input data and to the data defined but not
used as its output data. Next, a block b in the program graph dominates another
block b if every control-flow path from the program function’s entry block to
b goes through b. By definition, a block always dominates itself. Using these
notions, we define a placement of selected matches to blocks to be a solution to
the global code motion problem if each datum d in the program graph is defined
in some block b such that b dominates every block wherein d is used. Note that
616 G. Hjort Blindell et al.

processor pattern set

transformer
instructions matches CP output
matcher modeler solver
input model program
transformer
program program graph

Fig. 5. Overview of our approach.

this definition excludes control procedures because moving such operations to

another block rarely preserves the semantics of the program.
Some instructions impact both the covering and global code motion prob-
lems simultaneously. Assume for example a match in Fig. 3 of the pattern graph
from Fig. 4, which corresponds to the satadd instruction. A match spans across
the blocks in the program graph corresponding to the blocks appearing in the
pattern graph. Hence, the match above spans across blocks bb3, bb4, and bb5. Of
these, we note that the control procedures involving bb4 are all covered by this
match. Consequently, the computations performed within this block must all
be implemented by the satadd instruction. The match therefore consumes bb4,
meaning no other matches can be placed in this block. Consequently, a universal
instruction selector must take both the covering problem and the global code
motion problem into account when selecting the matches.

4 A Constraint Model for Universal Instruction Selection

This section introduces a constraint model for universal instruction selection.
An overview of the approach is shown in Fig. 5.
Match Identification. For a given program graph G and pattern set P , the
matches are identified by finding all instances M where a pattern graph in P
is subgraph isomorphic to G. Hence, separating identification from selection of
matches is an optimality-preserving decomposition. We use an implementation
of VF2 [13] to solve the subgraph isomorphism problem.
Depending on the pattern graph, certain matches may lead to cyclic data
dependencies and must therefore be excluded from the set of identified matches.
An example is a match of add4 in Fig. 3 defining t2 or t3 together with c1 .
Such matches can be detected by finding all connected components in the match
and checking whether a path exists between two components in the program
graph.
Global Instruction Selection. The set of variables sel(m) ∈ {0, 1} models
whether a match m ∈ M is selected, where M denotes the set of identified
matches. As explained in Sect. 3, each operation o ∈ O, where O denotes the
set of operations in the program graph, must be covered by exactly one of the
selected matches. Hence, if covers(m) ⊆ O denotes the set of operations covered
by match m, then the condition can be expressed as:

sel(m) = 1 ∀o ∈ O (1)
m∈M s.t.
o∈covers(m)
 Modeling Universal Instruction Selection 617

Global Code Motion. The set of variables place(m) ∈ B ∪ {bnull } models in

which block a match m is placed. B denotes the set of blocks in the program
graph, and bnull denotes an additional block (not part of the program graph) in
which non-selected matches are placed. In other words:
sel(m) ⇔ place(m) = bnull ∀m ∈ M (2)
where sel(m) abbreviates sel(m) = 1.
Control procedures cannot be placed in another block than originally indi-
cated in the program graph. Let entry(m) ∈ B denote the entry block of match m
if the pattern graph of m has such a node, otherwise entry(m) ∈ B. This condi-
tion can then be expressed as:
sel(m) ⇒ place(m) = entry(m) ∀m ∈ M, entry(m) ∈ B (3)
The set of variables def (d) ∈ B models in which block a datum d ∈ D is
defined, where D denotes the set of data in the program graph. As explained in
Sect. 3, each datum d must be defined in some block b such that b dominates
every block wherein d is used. In addition, the conditions imposed by the defi-
nition edges must be maintained. Let dom(b) ⊆ B denote the set of blocks that
dominate block b, where it is assumed that dom(bnull ) = B. Also let uses(m) ⊆ D
denote the set of data used by match m, and let DE denotes the set of definition
edges in the program graph. The conditions above can then be expressed as:
def (d) ∈ dom(place(m)) ∀m ∈ M, ∀d ∈ uses(m) (4)
def (d) = b ∀ {d, b} ∈ DE (5)
The def (·) variables must be connected to the place(·) variables. Intuitively,
if a selected match m is placed in block b and defines a datum d, then d should
also be defined in b. However, a direct encoding of this condition leads to an
over-constrained model. Assume again a match in Fig. 3 of the pattern graph
from Fig. 4, which thus defines the values c1 and c3 . Due to the definition
edges incurred by the ϕ-node, c1 and c3 must be defined in blocks bb3 and bb5,
respectively. But if def (d) = place(m) is enforced for every datum d defined by
a match m, then the match above will never become eligible for selection because
c1 and c3 must be defined in different blocks whereas a match can only be
placed in a single block. In such cases it is sufficient to require that a datum is
defined in any of the blocks spanned by the match. Let spans(m) ⊆ B denote
the set of blocks spanned by match m and defines(m) ⊆ D denote the set of
data defined by m. Then the condition can be relaxed by assigning def (d) to
place(m) when spans(m) = ∅, otherwise to any of the blocks in spans(m). Both
these conditions can be combined into a single constraint:
sel(m) ⇒ def (d) ∈ {place(m)} ∪ spans(m) ∀m ∈ M, ∀d ∈ defines(m) (6)
Finally, matches cannot be placed in blocks that are consumed by some
selected match. If consumes(m) ⊆ B denotes this set for a match m, then this
condition can be expressed as:
sel(m) ⇒ place(m ) = b ∀m, m ∈ M, ∀b ∈ consumes(m) (7)
618 G. Hjort Blindell et al.

Data Copying. Instructions typically impose requirements on the values that

they operate on, for example that its input and output data must be located
in particular registers. Combinations of such instructions may therefore require
additional copy instructions to be selected in order to fulfill these requirements.
The set of variables loc(d) ∈ L ∪ {lnull } models in which location a datum d
is available. L denotes the set of possible locations, and lnull denotes the location
of a value computed by an instruction that can only be accessed by this very
instruction. For example, the address computed by a memory load instruction
with a sophisticated addressing mode cannot be reused by other instructions.
Thus, if stores(m, d) ⊂ L denotes the locations for a datum d permitted by a
match m – where an empty set means no restrictions are imposed – then such
conditions can be imposed as:

sel(m) ⇒ d ∈ stores(m, d) ∀m ∈ M, ∀d ∈ D s.t. stores(m, d) = ∅ (8)

This alone, however, may cause no solutions to be found for processors where
there is little overlap between the locations – like in the case of multiple register
banks – that can be used by the instructions. This problem is addressed using
a method called copy extension. Prior to identifying the matches, the program
graph is expanded such that every data usage is preceded by a special operation
called a copy, represented by copy nodes. This expansion is done
by inserting a copy node and value node along each data-flow edge
that represents a use of data, as shown in the figure to the right.
The definition edges are also moved such that they remain on the v

data adjacent to the ϕ-nodes. The same expansion is also applied

to the pattern graphs except for the data-flow edges that represent
v ⇒ cp
use of input data.
Consequently, between the output datum of one match that
is the input datum to another match, there will be a copy node
that is not covered by either match. It is therefore assumed that a
special null-copy pattern, consisting of a single copy node and two value nodes, is
always included in the pattern set. A match derived from the null-copy pattern
has zero cost but requires, if selected, that both data are available in the same
location. This means that if the null-copy pattern can be used to cover some
copy node, then the two matches connected by that copy node both use the
same locations for its data. Hence there is no need for an additional instruction
to implement this copy. If the locations are not compatible, however, then a
match deriving the null-copy pattern cannot be selected as that would violate
the requirement that the locations must be the same. An actual instruction is
therefore necessary, and the restrictions on the loc(·) variables ensure that the
correct copy instruction is selected.

Fall-Through Branching. Branch instructions often impose constraints on

the distance that can be jumped from the instruction to a particular block. For
example, conditional branch instructions typically require that the false block
be located directly after the branch instruction, which is known as a fall-through.
 Modeling Universal Instruction Selection 619

This condition may also be imposed by instructions that span multiple blocks.
Consequently, the order in which the blocks appear in the generated code is
interconnected with the selection of certain matches.
The set of variables succ(b) ∈ B ∪ {bnull } models the successor of block b. A
valid block order then corresponds to a cycle formed by the succ(·) variables.
Using the global circuit constraint [27], this condition can be expressed as:
 
circuit ∪b∈B∪{bnull } {succ(b)} (9)

succ(bnull ) = bentry (10)

Hence, if the instruction corresponding to a match m performs a fall-through to a
block b, then this condition can be expressed as sel(m) ⇒ succ(place(m)) = b.
This approach can readily be extended to incorporate generic jump distances as
required by some processors.

Objective Function. The objective function depends on the desired charac-

teristics of the generated code. For example, if the compiler should maximize the
performance of the program under compilation, then the execution time of each
block should be minimized. In addition, the execution time should be weighted
with the relative execution frequency of each block. Thus, if freq(b) ∈ N1 denotes
the relative execution frequency of block b, then the objective function to be
minimized is:  
freq(b) × cost(m) (11)
b∈B m∈M s.t.
place(m)=b

where cost(m) ∈ N0 is the relative time it takes to execute the instruction

corresponding to a match m. Note that non-selected matches are always placed
in the bnull block, which is not part of the B set. If the compiler should optimize
for code size, then the weight freq(b) is dropped from Eq. 11 and cost(m) is
redefined as the size of the instruction corresponding to match m.

Implied Constraints. Similarly to Eq. 1, the set of data must be defined by

the set of selected matches:

sel(m) = 1 ∀d ∈ D (12)
m∈M s.t.
d∈defines(m)

For every block b in which a datum is defined, there must exist a selected
match that either is placed in b or spans b:

def (d) = b ⇒ sel(m) ∧ b ∈ {place(m)} ∪ spans(m)

(13)
∀b ∈ B, ∀d ∈ D, ∃m ∈ M

If two matches impose conflicting requirements on input or output data loca-

tions, or impose conflicting fall-through requirements, then at most one of these
matches may be selected.
620 G. Hjort Blindell et al.

Dominance Constraints. By analyzing the constraints on the loc(·) variables,

one can identify subsets S of values such that any solution with loc(d) = v and
v ∈ S can be replaced by an equivalent solution with loc(d) = max(S), for any
d ∈ D. Consequently, all values in S \ {max(S)} can be a priori removed from
the domains of all loc(·) variables.
Suppose that there are two mutually exclusive matches m and m with
cost(m) ≤ cost(m ) and the constraints imposed by m are compatible with and
no stricter than the constraints imposed by m . Then any solution that selects m
can be replaced by a solution of less or equal cost that selects m. Consequently,
sel(m ) = 0 can be set a priori for all such m . In case m and m have identical
cost and impose identical constraints, a lexicographic ordering rule is used.

Branching Strategy. Our branching strategy only concerns those sel(·) vari-
ables of matches not corresponding to copy instructions. Let M  denote this
set of matches. We branch on {sel(m) | m ∈ M  } ordered by non-increasing
| covers(m)|, trying sel(m) = 1 before sel(m) = 0. The intuition behind this is to
eagerly cover the operations. The branching on the remaining decision variables
is left to the discretion of the solver (see Sect. 5 for details).

Model Limitations. A constant value that has been loaded into a register for
one use cannot be reused for other uses. Consequently, the number of selected
matches may be higher than necessary. This problem is similar to spilling reused
temporaries in [11] and can be addressed by adapting the idea of alternative
temporaries introduced in [10].
For some processors, conditional jumps can be removed by predicating
instructions with a Boolean variable that determines whether the instruction
shall be executed [3]. For example, assume that the statement c = MAX in Fig. 1b
is implemented using a copy instruction. If this instruction can be predicated
with the Boolean value MAX < c, then the conditional jump to block bb4 becomes
superfluous. This is known as if-conversion. Such instructions can be described
using two pattern graphs: one representing the predicated version, and another
representing the non-predicated version. But because every operation must be
covered by exactly one match, the predicated version can only be selected if the
match implements all computations in the conditionally executed block.

5 Experimental Evaluation
The model is implemented in MiniZinc [31] and solved with CPX [1] 1.0.2, which
supports FlatZinc 1.6. The experiments are run on a Linux machine with an Intel
Core i7-2620M 2.70 GHz processor and 4 GB main memory using a single thread.
We use all functions (16 in total) from MediaBench [28] that have more than
5 LLVM IR instructions and do not contain function calls or memory computa-
tions (due to limitations in the toolchain but not in the constraint model). The
size of their corresponding program graphs ranges between 34 and 203 nodes. The
functions are compiled and optimized into LLVM IR files using LLVM 3.4 [26]
with the -O3 flag (optimizes execution time). These files serve as input to our
 Modeling Universal Instruction Selection 621

Fig. 6. Estimated execution speedup over LLVM for simple instructions.

toolchain. However, as LLVM currently lacks a method re-entering its backend

after instruction selection we cannot yet execute the code generated by our app-
roach. Instead the execution time is estimated using LLVM’s cost model.
In the following, the full runtimes comprising matching, flattening the MiniZ-
inc model to FlatZinc, and solving are summarized. Thus the time for producing
the LLVM IR files and transforming them into program graphs is not included.
The time spent on matching is negligible (less than 1% of the full runtime), while
the time spent on flattening is considerable (more than 84% of the full runtime).
The solving time is measured until proof of optimality. All runtimes are averaged
over 10 runs, for which the coefficient of variation is less than 6%.

Simple Instructions. For proof of concept, the processor used is MIPS32 [33]
since it is easy to implement and extend with additional instructions. The pat-
tern graphs are manually derived from the LLVM instruction set description (to
enable comparison), however the process could be easily automated as patterns
are already described as trees in LLVM. As greedy heuristics already generate
code of sufficient quality – in many cases even optimal – for MIPS32, we should
not, in general, expect to see any dramatic execution speedup.
The shortest full runtime for the benchmarks is 0.3 seconds, the longest is
83.2 seconds, the average is 27.4 seconds, and the median is 10.5 seconds.
Fig. 6 shows the estimated execution speedup of our approach compared to
LLVM. The geometric mean speedup is 1.4%; in average our approach is slightly
better than LLVM. As the figure reveals, our approach has the potential to gener-
ate better code than the state of the art even for simple and regular architectures
such as MIPS32, mainly due to its ability of moving computations to blocks with
lesser execution frequency. The cases where our approach generates code that is
worse than LLVM are due to the model limitations described in Sect. 4. Some of
these cases are aggravated by the fact that LLVM’s instruction selector is capa-
ble of, where appropriate, combining chains of binary ϕ-functions into a single
ϕ-function in order to reduce the number of branch operations. This feature has
yet to be implemented in our toolchain.
622 G. Hjort Blindell et al.

30%

20%

10%

0%
re

gs

gs

gs

gs

jp

gs

gs

gs

gs

gs

ul

gs

jr

gs

al
co

m_

m_

m_

m_

eg

m_

m_

m_

m_

m_

aw

m_

ou

m_

aw
ns

L_

mu

as

ad

_q

L_

L_

L_

mu

as

2l

ab

nd

su

2l
tr

su

lt

ua

as

ad

as

lt

in

_u

in
uc

_r

li

ea

ea
t

ty

r
_s
c.
Fig. 7. Estimated execution speedup by adding SIMD instructions to MIPS32.

An interesting aspect is that due to the branching described in Sect. 4 solv-

ing yields a very good first solution, which provides a tight upper bound. We
conjecture that this together with the lazy clause learning in the CPX solver is
the reason why we can prove optimality for all benchmarks.

SIMD Instructions. In the following experiments we extend the MIPS32

architecture with SIMD instructions for addition, shifting, and Boolean logic
operations as motivated by potential matches in the considered benchmark func-
tions. The SIMD instructions use the same registers as the simple instructions.
The shortest full runtime for the benchmarks is 0.3 seconds, the longest is
146.8 seconds, the average is 44.2 seconds, and the median is 10.5 seconds.
Fig. 7 shows the estimated execution speedup for MIPS32 together with
SIMD instructions. The geometric mean speedup compared to our approach
for basic MIPS32 is 3%. The best cases correspond to functions that are
computation-intensive (e.g. ulaw2linear) as more data parallelism can be
exploited by the SIMD instructions. The worst cases (no improvement) cor-
respond to control-intensive functions (e.g. gsm L asr), where there are either
no matches for SIMD instructions or matches that are not profitable. In the
latter case, the solver often discards SIMD matches that would require moving
operations to hot blocks which are estimated to be executed often. A SIMD
instruction in a hot block would be more costly than multiple primitive opera-
tions in the colder blocks where they originally reside. Our approach is unique
in that it reflects this trade-off accurately. A traditional approach – vectoriza-
tion first, then instruction selection – would greedily select the SIMD instruc-
tion and generate worse code. If a certain primitive operation must already be
executed in a hot block then the solver will bring in operations of the same
type from colder blocks to form a speculative SIMD instruction (this is the case
e.g. for alaw2linear).
Hence our approach can exploit sophisticated instructions (such as SIMD)
and has the potential to improve over traditional approaches since it accurately
reflects the trade-offs in instruction selection. We also expect to see further
benefits when integrated with register allocation and instruction scheduling.
 Modeling Universal Instruction Selection 623

6 Related Work
Several linear time, optimal algorithms exist for local instruction selection on
tree-based program and pattern graphs [2,22,32]. These have subsequently been
extended to DAG-based program graphs [17,18,25], but at the loss of optimal-
ity. Together with instruction scheduling or register allocation, the problem has
also been approached using integer programming (IP) [7,21,36] and constraint
programming (CP) [6,19,30]. Far fewer methods exist for global instruction selec-
tion, which so far only has been approached as a partitioned Boolean quadratic
problem [9,15,16]. Common among these techniques is that they are restricted
to pattern trees or pattern DAGs.
Many methods exist for selecting vector instructions separately from instruc-
tion selection, but attempts have been made at combining these two tasks using
IP [29,34] and CP [4]. Of these, however, only [34] takes the cost of data copying
into account, and none is global.
Global code motion has been solved both in isolation [12] as well as in inte-
gration with register allocation [5,24]. Both [5] and [12] use a program represen-
tation that is similar to ours, but where the data are not explicitly represented
as nodes. To the best of our knowledge, no previous attempt has been made in
combining global code motion with instruction selection.

7 Conclusions and Future Work

This paper introduces a universal graph-based representation for programs and
instructions that unifies control and data flow. From the representations a new
constraint model for instruction selection is derived. The paper shows that the
approach is more expressive and generates code of similar quality compared to
LLVM as a state-of-the-art compiler. The constraint model is robust for small
to medium-sized functions as well as expressive processor instructions and is
competitive with LLVM in code quality.
One line of future work is to extend the model to address the limitations
discussed in Sect. 4. Additional improvements include exploring more implied
and dominance constraints and pre-solving techniques to increase the model’s
robustness. We intend to explore both larger input programs as well as more
processor architectures with a more robust model.
Instruction selection is but one task in code generation. We will explore in
detail how this paper’s model can be integrated with a constraint model for
register allocation and instruction scheduling introduced in [10,11].

Acknowledgments. This research has been partially funded by LM Ericsson AB and

the Swedish Research Council (VR 621-2011-6229). The authors are grateful for helpful
discussions with Frej Drejhammar and Peter van Beek and for constructive feedback
from the anonymous reviewers.
624 G. Hjort Blindell et al.

References
1. Opturion CPX user’s guide: Version 1.0.2. Tech. rep., Opturion Pty Ltd (2013)
2. Aho, A.V., Ganapathi, M., Tjiang, S.W.K.: Code Generation Using Tree Match-
ing and Dynamic Programming. Transactions on Programming Languages and
Systems 11(4), 491–516 (1989)
3. Allen, J.R., Kennedy, K., Porterfield, C., Warren, J.: Conversion of Control Depen-
dence to Data Dependence. In: ACM SIGACT-SIGPLAN Symposium on Principles
of Programming Languages, pp. 177–189 (1983)
4. Arslan, M.A., Kuchcinski, K.: Instruction Selection and Scheduling for DSP Ker-
nels on Custom Architectures. In: EUROMICRO Conference on Digital System
Design (2013)
5. Barany, G., Krall, A.: Optimal and Heuristic Global Code Motion for Minimal
Spilling. In: Jhala, R., De Bosschere, K. (eds.) Compiler Construction. LNCS,
vol. 7791, pp. 21–40. Springer, Heidelberg (2013)
6. Bashford, S., Leupers, R.: Constraint Driven Code Selection for Fixed-Point DSPs.
In: ACM/IEEE Design Automation Conference, pp. 817–822 (1999)
7. Bednarski, A., Kessler, C.W.: Optimal Integrated VLIW Code Generation with
Integer Linear Programming. In: Nagel, W.E., Walter, W.V., Lehner, W. (eds.)
Euro-Par 2006. LNCS, vol. 4128, pp. 461–472. Springer, Heidelberg (2006)
8. Bruno, J., Sethi, R.: Code Generation for a One-Register Machine. Journal of the
ACM 23(3), 502–510 (1976)
9. Buchwald, S., Zwinkau, A.: Instruction Selection by Graph Transformation. In:
International Conference on Compilers, Architectures and Synthesis for Embedded
Systems, pp. 31–40 (2010)
10. Castañeda Lozano, R., Carlsson, M., Blindell, G.H., Schulte, C.: Combinatorial
spill code optimization and ultimate coalescing. In: Kulkarni, P. (ed.) Languages,
Compilers, Tools and Theory for Embedded Systems, pp. 23–32. ACM Press, Edin-
burgh, UK (2014)
11. Lozano, R.C., Carlsson, M., Drejhammar, F., Schulte, C.: Constraint-Based Reg-
ister Allocation and Instruction Scheduling. In: Milano, M. (ed.) CP 2012. LNCS,
vol. 7514, pp. 750–766. Springer, Heidelberg (2012)
12. Click, C.: Global Code Motion/Global Value Numbering. In: ACM SIGPLAN 1995
Conference on Programming Language Design and Implementation, pp. 246–257
(1995)
13. Cordella, L.P., Foggia, P., Sansone, C., Vento, M.: A (Sub)Graph Isomorphism
Algorithm for Matching Large Graphs. IEEE Transactions on Pattern Analysis
and Machine Intelligence 26(10), 1367–1372 (2004)
14. Cytron, R., Ferrante, J., Rosen, B.K., Wegman, M.N., Zadeck, F.K.: Efficiently
Computing Static Single Assignment Form and the Control Dependence Graph.
ACM TOPLAS 13(4), 451–490 (1991)
15. Ebner, D., Brandner, F., Scholz, B., Krall, A., Wiedermann, P., Kadlec, A.: Gen-
eralized Instruction Selection Using SSA-Graphs. In: ACM SIGPLAN-SIGBED
Conference on Languages, Compilers, and Tools for Embedded Systems, pp. 31–40
(2008)
16. Eckstein, E., König, O., Scholz, B.: Code Instruction Selection Based on SSA-
Graphs. In: Anshelevich, E. (ed.) SCOPES 2003. LNCS, vol. 2826, pp. 49–65.
Springer, Heidelberg (2003)
 Modeling Universal Instruction Selection 625

17. Ertl, M.A.: Optimal Code Selection in DAGs. In: ACM SIGPLAN-SIGACT Sym-
posium on Principles of Programming Languages, pp. 242–249 (1999)
18. Ertl, M.A., Casey, K., Gregg, D.: Fast and Flexible Instruction Selection with On-
Demand Tree-Parsing Automata. In: ACM SIGPLAN Conference on Programming
Language Design and Implementation, pp. 52–60 (2006)
19. Floch, A., Wolinski, C., Kuchcinski, K.: Combined Scheduling and Instruction
Selection for Processors with Reconfigurable Cell Fabric. In: International Confer-
ence on Application-Specific Systems, Architectures and Processors, pp. 167–174
(2010)
20. Garey, M., Johnson, D.: Computers and Intractability: A Guide to the Theory of
NP-Completeness. Freeman (1979)
21. Gebotys, C.H.: An Efficient Model for DSP Code Generation: Performance,
Code Size, Estimated Energy. In: International Symposium on System Synthesis,
pp. 41–47 (1997)
22. Glanville, R.S., Graham, S.L.: A New Method for Compiler Code Generation. In:
ACM SIGACT-SIGPLAN Symposium on Principles of Programming Languages,
pp. 231–254 (1978)
23. Hjort Blindell, G.: Survey on Instruction Selection: An Extensive and Modern
Literature Study. Tech. Rep. KTH/ICT/ECS/R-13/17-SE, KTH Royal Institute
of Technology, Sweden (October 2013)
24. Johnson, N., Mycroft, A.: Combined Code Motion and Register Allocation Using
the Value State Dependence Graph. In: International Conference of Compiler Con-
struction, pp. 1–16 (2003)
25. Koes, D.R., Goldstein, S.C.: Near-Optimal Instruction Selection on DAGs. In:
IEEE/ACM International Symposium on Code Generation and Optimization,
pp. 45–54 (2008)
26. Lattner, C., Adve, V.: LLVM: A compilation framework for lifelong program anal-
ysis & transformation. In: IEEE/ACM International Symposium on Code Gener-
ation and Optimization (2004)
27. Laurière, J.L.: A Language and a Program for Stating and Solving Combinatorial
Problems. Artificial Intelligence 10(1), 29–127 (1978)
28. Lee, C., Potkonjak, M., Mangione-Smith, W.H.: MediaBench: A tool for evaluating
and synthesizing multimedia and communications systems. In: IEEE MICRO-30,
pp. 330–335 (1997)
29. Leupers, R.: Code Selection for Media Processors with SIMD Instructions. In:
Conference on Design, Automation and Test in Europe, pp. 4–8 (2000)
30. Martin, K., Wolinski, C., Kuchcinski, K., Floch, A., Charot, F.: Constraint-Driven
Instructions Selection and Application Scheduling in the DURASE System. In:
International Conference on Application-Specific Systems, Architectures and Pro-
cessors, pp. 145–152 (2009)
31. Nethercote, N., Stuckey, P.J., Becket, R., Brand, S., Duck, G.J., Tack, G.R.: MiniZ-
inc: Towards a Standard CP Modelling Language. In: Bessière, C. (ed.) CP 2007.
LNCS, vol. 4741, pp. 529–543. Springer, Heidelberg (2007)
32. Pelegrı́-Llopart, E., Graham, S.L.: Optimal Code Generation for Expression Trees:
An Application of BURS Theory. In: ACM SIGPLAN-SIGACT Symposium on
Principles of Programming Languages, pp. 294–308 (1988)
626 G. Hjort Blindell et al.

33. Sweetman, D.: See MIPS Run, Second Edition. Morgan Kaufmann (2006)
34. Tanaka, H., Kobayashi, S., Takeuchi, Y., Sakanushi, K., Imai, M.: A Code Selection
Method for SIMD Processors with PACK Instructions. In: International Workshop
on Software and Compilers for Embedded Systems, pp. 66–80
35. Živojnović, V., Martı́nez Velarde, J., Schläger, C., Meyr, H.: DSPstone: A DSP-
Oriented Benchmarking Methodology. In: Conference on Signal Processing Appli-
cations and Technology, pp. 715–720 (1994)
36. Wilson, T., Grewal, G., Halley, B., Banerji, D.: An Integrated Approach to Retar-
getable Code Generation. In: International Symposium on High-Level Synthesis,
pp. 70–75 (1994)
 Optimizing the Cloud Service Experience
Using Constraint Programming

Serdar Kadioglu(B) , Mike Colena, Steven Huberman, and Claire Bagley

Oracle Corporation, Burlington, MA 01803, USA

{serdar.kadioglu,mike.colena,steven.huberman,
claire.bagley}@oracle.com

Abstract. This paper shows how to model and solve an important

application of the well-known assignment problem that emerges as part
of workforce management, particularly in cloud based customer service
center optimization. The problem consists of matching a set of highly
skilled agents to a number of incoming requests with specialized require-
ments. The problem manifests itself in a fast-paced online setting, where
the complete set of incoming requests is not known apriori, turning this
into a challenging problem where rapid response time and quality of
assignments are crucial for success and customer satisfaction. The key
contribution of this paper lies in the combination of a high-level con-
straint model with customizable search that can take into account var-
ious objective criteria. The result is an efficient and flexible solution
that excels in dynamic environments with complex, conflicting and often
changing requirements. The constraint programming approach handles
hundreds of incoming requests in real-time while ensuring high-quality
agent assignments.

1 Introduction

Introduced in early 1950s, the Assignment Problem is one of the fundamental

combinatorial optimization problems. It deals with the question of how to assign
agents (resources, jobs, etc.) to requests (tasks, machines, etc.) in the “best”
possible way on a one-to-one basis. Equivalently, the problem consists of finding
a maximum (or minimum) weighted matching in a weighted bipartite graph [1].
While the definition of the best depends on the particular application, it is
usually a value associated with each agent-request assignment representing time,
cost, profit, sales and etc.
The assignment problem enjoys numerous practical applications such as the
assignment of doctors/nurses to hospitals in the residency problem, students
to schools in the college admission problem, human organs to recipients in the
transplant matching problem, and aircrafts to gates as part of airline scheduling
problem (see [22] for a detailed survey). Recognizing the importance of the work
in this area, A. Roth and L. Shapley were awarded the Nobel Prize in Economic
Sciences in 2012, for their contributions to the theory of stable matchings.


c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 627–637, 2015.
DOI: 10.1007/978-3-319-23219-5 43
628 S. Kadioglu et al.

Fig. 1. An Example of the Agent Matching Problem.

In terms of solution methods, the assignment problem can be solved with the
simplex algorithm as it is a special case of linear programming. However, there
exist more efficient algorithms for this special structure. The prominent solution
method is the Hungarian algorithm (due to Hungarian mathematicians D. Kőnig
and J. Egerváry [17]). The algorithm runs in cubic time in the number of agents
(tasks) in the worst-case and is strongly polynomial [18], i.e., there exists no algo-
rithm for this structure with better time complexity. Other solutions include the
shortest augmenting path algorithms (see e.g., [13]), or flow-based approaches,
which have similar worst-case complexity.
Moving beyond the assignment problem to a more general setting, the next
section describes a number of important dimensions that extends the classical
definition. While these extensions bring the problem closer to real-life applica-
tions, it also becomes much harder to solve using a specialized approach such as
the Hungarian algorithm. To that end, this paper introduces a better solution
based on Constraint Programming (CP).

2 Service Center Optimization: Agent Matching

An important application of the assignment problem emerges as part of work-
force management, particularly in cloud based customer service center optimiza-
tion [12]. Conceptually, the problem consists of assigning a set of skilled agents
to a number of incoming requests with specialized requirements in a continuous,
24/7, fast-paced setting.

Definition 1 (The Agent Matching Problem (AMP)). Given a set of

Agents, A, and a set of Requests, R, the goal is to assign every request r ∈ R
to an agent a ∈ A such that: (i) the language of the agent matches the required
language of the request, (ii) preferably, the skill set of the agent contains the
skill demanded by the request, and (iii) each agent has a limit on the number
of requests that they can work on simultaneously; therefore, capacity constraints
must be respected at all times.
 Optimizing the Cloud Service Experience Using Constraint Programming 629

It is important to note that not all agents possess the same skill level. Given a
particular skill, there is a natural ordering among agents based on their expertise
and experience. Similarly, not every request is equally important. Depending on
the service level agreement, some requests may have priority over others. In
fact, there might even be multiple priority levels. The ultimate goal is to assign
the most skillful agent to the most important request while increasing agent
utilization and decreasing the wait time. This strategy is referred to as the best
agent matching.
Figure 1 presents an example of the AMP with 5 agents, shown in blue icons,
and 15 requests, shown in either green or yellow triangles depending on whether
they are standard or priority requests respectively. This is the initial state of the
problem before any matching has taken place. Each agent possesses a number of
technical skills and speaks one or more languages (see e.g., Agent 1). Similarly,
each request demands a certain technical skill and a language (see e.g., Request
3). Agents are ordered with respect to their skills based on their experience level
as shown in the Agent Skills list.

3 Why CP?

The AMP is a generalization of the assignment problem and differs from the clas-
sical definition in a number important ways. Unlike the original definition, the
number of tasks and requests might not align, leading to an unbalanced match-
ing problem. Secondly, this is an online problem as opposed to its one-time-only
offline counterpart, meaning that, the complete set of incoming requests cannot
be determined apriori. Consequently, the number of available agents changes
dynamically as the system proceeds and requests become active (or inactive)
through time. The agents are distinguishable in their skill set. Not every agent
can handle every request, i.e., there exists forbidden pairs of assignments. More-
over, agents are not bound to perform a single task as in the classical assignment
problem; they can work on multiple requests simultaneously. In addition, there
is no single objective to optimize; it is often the case that multiple conflicting
objectives are present. For example, while minimizing the number of agents is
important to reduce the operating cost, employing more agents reduces the wait
time, which in turn improves the customer satisfaction.
Such real-life requirements complicate the efficient algorithms that are
designed for the original setting. The Hungarian method, and similarly flow-
based approaches, have to generate a weighted combination of different objec-
tives and cost values for possible assignments. These values are not only hard
to predict but also need to be updated each time a constraint or preferences
change. Both the objective weighting schema and cost values highly influence
the actual solutions found. In practice, even the exact numeric values become
important for the stability and the convergence of the algorithm. Finally, intro-
ducing artificial high cost values to forbid infeasible agent-request assignments
or copies of agents (or tasks) are prohibitive for the scalability of the approach
as they have direct impact on the asymptotic complexity of the algorithm.
630 S. Kadioglu et al.

These considerations turn AMP into a challenging problem where rapid

response time and quality of assignments are crucial for success and customer
satisfaction. Poor decisions can lead to under-staffing, poor customer service,
and loss of revenue. The nature of the problem demands a fast real-time solu-
tion that can run iteratively in such a dynamic environment, and at the same
time, take into account various objective criteria. This paper proposes a solution
in this direction using constraint programming.

4 Added Value of CP

Expressiveness: As shown in Section 5.1, the CP model is high-level and offers a

succinct representation of the problem without decomposing it, e.g., into Boolean
assignment variables. There is almost no gap between the application and imple-
mentation, which is highly desired in a production environment and reduces the
time to solution.
Flexibility: The plug-and-play nature of the constraint model makes it easy
to add/remove restrictions to the problem. Customizable search heuristics, an
example of which is given in Section 5.2, allow expressing various preferences,
even conflicting ones.
Maintainability: There is a clear separation between the model and the search.
Each component can be developed on its own and can be modified without
changing the other parts. This is a strong property in an industrial setting.
Essentially, the same abstract model can be utilized to solve the problem while
different customers implement different preferences.
Efficiency: Powerful constraint engines are capable of handling hundreds of
requests per second while assigning the best possible agents. As shown in Section
7, this amounts to handling millions of requests an hour, which satisfies beyond
the current needs of service centers at peak hours. In addition, performing the
best agent selection incurs almost no additional cost on the runtime.

5 Solution Approach

In principle, a constraint model consists of three elements: a set of variables

each of which can take a value from a set of possible values, called domains,
and a set of restrictions on the possible combinations of values referred to as
constraints. [27].
The AMP lends itself naturally to a high-level model thanks to the declarative
nature of constraint programming. First, a high-level constraint model that can
capture the combinatorial structure of the problem is presented. Then, this model
is combined with a customizable greedy search heuristic to find a good solution
quickly.
 Optimizing the Cloud Service Experience Using Constraint Programming 631

5.1 The Constraint Model

The model uses two sets of decision variables; one that corresponds to the
requests, and one that accounts for the number requests assigned to an agent.

Request Variables: There is a one-to-one mapping between these variables and

requests. Each request is treated as an integer decision variable with a domain
that consists of only the agents that speak the required language. Notice how
language compatibility is implicitly guaranteed within this formalism via the
domains. There is no need to generate artificial high cost values to forbid infea-
sible agent-request assignments; those pairs simply do not exist. This reduces
the size of the problem considerably.
Given that some requests might not have an available agent at a particular
time, a wildcard agent, that is capable of serving any request, is added as the
last value to each domain. This way, the model is guaranteed to have a feasi-
ble solution while an assignment to the wildcard agent indicates an unmatched
request.

Cardinality Variables: In order to count the number of times an agent is

assigned to a request, a cardinality variable is created for each agent with a
domain from 0 (no assignment) to the amount of current work capacity of the
agent. The domain for the wildcard agent is set from 0 to the number of requests.
The CP formalism allows agents to have different capacities with ease depending
on their experience level. Similarly, the upper bound in each domain implicitly
enforces the maximum capacity of an agent.

Global Cardinality Constraint: The request and cardinality variables are

linked together via the (extended) global cardinality constraint [4,21,25]:

gcc(requestV ars, {Agents, wildCardAgent}, cardV ars)

5.2 The Search Heuristic

Variable Selection: Sorting the request variables, first, by their priority, and
then, by their wait time yields a natural variable ordering heuristic. Without loss
of generality, the variable heuristic can be extended into more levels of priorities
or other sorting schemas.

Value Selection: Given a request variable, the next step is to find an agent to
assign to it. Agents can be selected as directed by a “best” matching heuristic.
Several interesting alternatives exist; an exemplar value selection procedure is
outlined in Algorithm 1. The algorithm treats priority requests specially; if it
is possible, priority requests are assigned to the most skillful agent for the task
(line 6), otherwise, they are still matched with an agent with the same language
(line 10).
632 S. Kadioglu et al.

Input: request variable, var

Output: request-agent assignment, σ(var ← val)
1: val := −1
2: if var is a priority request then
3: agentsSortedBySkill = a best-first ordering of all agents wrt the desired skill
of var
4: for all v ∈ agentsSortedBySkill do
5: if v ∈ the domain of var then
6: val := v; break
7: end if
8: end for
9: if val == wildcard then
10: val := the first agent in the domain of var
11: end if
12: else
13: val := the first agent in the domain of var
14: end if
15: return σ(var ← val)
Algorithm 1: Value Selection: Best Agent Matching Heuristic

Figure 2 illustrates the assignments found when the incoming requests in

Figure 1 are matched to the agents without (on the left) and with (on the
right) the best agent matching heuristic. When the heuristic is in play, notice
how a priority request, namely Request 10, is assigned to Agent 4, who is the
most skilled representative on technical issues with iPads (see Agent Skills list
in Figure 1). Each agent works at capacity, i.e. two simultaneous requests, and
agent utilization is maximized. No priority request is waiting for service since
the heuristic tries to serve priority requests first. On the other hand, when the
heuristic is not active, two priority requests are placed on the waitlist, as shown
in red alerting triangles. Also, while Request 10 is still served in this case, it
is served by Agent 5, who speaks the same language, but not technically savvy
on iPad issues at all. This sharp contrast between the two solutions clearly
reveals the impact of customized search heuristics for improving the quality of
the assignments.
Again, without loss of generality, the value selection heuristic can be modified
to serve other purposes. For example, among the most skillful agents, the agent
with the least current load can be favored to balance the work among agents,
and so on.

5.3 A Note on the Filtering Power of GCC

It is important to note that customized search heuristics, such as the one

described in Algorithm 1, might assign values from the inner range of the domains
instead of the bounds. This leads to creating holes in the domains. Therefore,
such heuristics should be used with care when employing bounds consistency
 Optimizing the Cloud Service Experience Using Constraint Programming 633

Fig. 2. The Solution Quality. Solutions found for the AMP instance in Figure 1 when
the Best Agent Matching heuristic is not active (on the left) and is active (on the
right).

filtering on the cardinality variables (e.g., as in [15,24]). In the case of AMP, we

recognized the following behavior:

Remark 1. When the maximum cardinality of an agent is less than the number
of requests demanding the same skill, for which, this agent is the most skilled,
the best agent matching heuristic, in combination with a bounds consistent gcc
operator, might lead to thrashing.

To enable stronger filtering of cardinality variables, it is beneficial to con-

sider all the bounded variables and update their associated cardinalities. This
improvement is similar to the Assigned Var Removal optimization described in
the empirical gcc survey of [20].

6 Related Work

The existing approaches for solving the variants of the assignment problem
broadly fall into two main categories: (i) ad-hoc, heuristic implementations,
and (ii) generic methods. As opposed to the approaches in the first category,
which produce very specific solutions, lack modeling support, and hard to adapt
changes, our approach belongs to the second category, where, traditionally, Oper-
ational Research and Artificial Intelligence methods are employed. In particular,
constraint (logic) programming techniques were successfully used for assignment
problems, such as stand or counter allocation in airports, and berth allocation
to container ships (see e.g., [6,7,10,23]). CP is also applied to the special case
of personnel assignment as it is well-suited for changing work rules and labor
regulations, such as rostering nurses in hospitals, scheduling technicians, and
generating work plans to cover different skills (see e.g., [5,8,11]).
The work presented in [3,19,26] uses CP to assign highly-skilled profession-
als and employs similar decision variables. However, it focuses on one-to-one
matching, whereas, in our context, multiple assignments are allowed. Also, our
solution is geared toward a rapid operational setting whereas other approaches
634 S. Kadioglu et al.

Fig. 3. Scalability Experiments. The effect of increasing the number of agents, A, and
the number of requests, R.

seek long-term strategic or tactical planning. As in the work of [28], our approach
also avoids the costly optimality search and constructs a good solution instead.
Finally, it is possible to take into account not only the preferences of requests
but also the agents as shown in [2].

7 Numerical Results
The constraint model and the search heuristic are implemented using an in-house
constraint programming solver. All experiments were run on a Dell laptop with
Intel Core i5 CPU @2.5 GHz and 8.00 Gb RAM.
Benchmarks: Experiments were run on instances of varying complexity and size
based on the information gathered from our collaboration with other internal
groups. Each data set consists of 50 synthetic instances whereby the largest
instance deals with 250 agents and 500 request. Half of the requests selected
uniformly at random as priority, and are treated specially. Regarding the capacity
constraint, each agent is allowed to handle up to four requests simultaneously.
The goal of the experiments is to analyze the runtime and scalability of the
approach. Notice that, in real applications, there is a dynamic nature of the
problem: ongoing chat sessions remain active in the system for some unknown
period, and as agents become available, the system continues to serve the pre-
viously unmatched requests as well as the new incoming requests that arrived
in the meantime. In that sense, the following experiments stress test the initial
state of the problem.
Impact of Problem Size: In an online setting such as the AMP, the response
time is crucial, therefore, only a restricted runtime is allowed. As shown in
Figure 3, the CP approach spends minimal time, around half of a second on
average for the largest data set, while performing the best agent matching for
50% of the requests. If the proposed solution is deployed in a service center,
this would amount to dealing with more than 2 million requests per hour which
satisfies the business requirements of typical cloud based service centers under
heavy loads.
 Optimizing the Cloud Service Experience Using Constraint Programming 635

Next, the impact of increasing the number of agents and the number of
requests in isolation is studied. While the former leads to larger domains, the
latter corresponds to handling more decision variables. The tests revealed that
the complexity of the CP model depends more on the number of requests than
the number of agents.
Comparison with the Hungarian Algorithm: To test the CP approach
with the best-known algorithm, an in-house, specialized implementation of the
Hungarian method is considered. Experiments showed that the CP approach was
more than an order of magnitude faster on average for the same data sets. It
should be noted that the Hungarian algorithm solves an optimization problem,
albeit a fictitious one generated by imaginary agent-request assignment costs,
while the CP approach finds a greedy solution empowered with a constraint
model. The modification of the problem instances to fit into the right format
(such as, creating dummy nodes and copies of agents, generating artificial values
to express preferences and to forbid incompatible pairs) turned out to be pro-
hibitive in practice, especially for an online setting where multiple assignment
problems are solved successively. The same issue applies to flow-based approaches
as well. Moreover, this not only requires pre-processing the instances, but also,
post-processing to map the solution found back, which further complicates the
code and maintenance in the long run.
Impact of Best Agent Matching: Finally, the overhead of the special treat-
ment for priority requests is studied. The percentage of priority requests is
adjusted from none (0%) to all (100%) using increments of size 5. The num-
ber of agents is fixed to 200 and the number of requests is fixed to 400.
The experiments shown in Figure 4 reveals an interesting property of these
generated instances. The AMP problem goes through a transition when the
20(±3)% of requests are marked as priority, which might be attributed to an
easy-hard-less–hard pattern [9,14,16]. The behavior is due to the opposing forces
between the best agent matching heuristic, which favors most skilled agents,
versus the cardinality constraints, which allows only up to four simultaneous
requests. The combination stretches the cardinality constraint for validation.
Across all priority levels, the impact of the heuristic remains marginal. In a

Fig. 4. The Effect of the Best Agent Matching Heuristic.

636 S. Kadioglu et al.

sense, within the CP formalism, preference based search, i.e., the considered
best agent matching heuristic, comes at almost no additional cost, which makes
the approach more attractive.

8 Conclusion
Optimizing a service center involves many decisions one of which deals with
assigning agents in the best possible way. A good long-term hiring policy, skill-
based routing, and the right mix of employee contract types are essential for good
workforce management. This paper focused on runtime efficiency and assign-
ment quality in a real-time environment. The expressive modeling language of
constraint programming enabled a compact and flexible representation of this
combinatorial problem. This model provided an efficient solution that can handle
hundreds of requests per second, and more than 2 million per hour. The flexibility
of variable and value selection allowed customizable search that can account for
multiple objectives. The model and the search remained completely independent
from each other allowing practitioners to develop and modify one without alter-
ing the other. Overall, the proposed constraint programming solution, stands
out as an attractive approach to be considered for industrial applications of the
service center optimization problem.

References
1. Ahuja, R.K., Magnanti, T.L., Orlin, J.B., Reddy M.R.: Applications of Network
Optimization. In: Network Models. Handbooks in Operations Research and Man-
agement Science, vol. 7. North-Holland (1995)
2. Alsheddy, A., Tsang, E.P.K.: Empowerment scheduling for a field workforce.
J. Scheduling 14(6), 639–654 (2011)
3. Boni, O., Fournier, F., Mashkif, N., Naveh, Y., Sela, A., Shani, U., Lando, Z.,
Modai, A.: Applying Constraint Programming to Incorporate Engineering Method-
ologies into the Design Process of Complex Systems. In: Proceedings of the Twenty-
Fourth Conference on Innovative Applications of Artificial Intelligence (2012)
4. Bourdais, S., Galinier, P., Pesant, G.: hibiscus: A Constraint Programming Appli-
cation to Staff Scheduling in Health Care. In: Rossi, F. (ed.) CP 2003. LNCS, vol.
2833, pp. 153–167. Springer, Heidelberg (2003)
5. Chan, P., Heus, K., Veil, G.: Nurse scheduling with global constraints in CHIP:
Gymnaste. In: Proc. PACT 1998 (1998)
6. Chow, K.P., Perrett, M.: Airport counter allocation using constraint logic program-
ming. In: Proc. PACT 1997 (1997)
7. Chun, A.H.W., Chan, S.H.C., Tsang, F.M.F., Yeung, D.W.M.: Stand allocation
with constraint technologies at Chek Lap Kok international airport. In: Proc.
PACLP 1999 (1999)
8. Collignon, C.: Gestion optimisée de ressources humaines pour l’Audiovisuel. In:
Proc. CHIP users club (1996)
9. Crawford, J.M., Auton, L.D.: Experimental Results on the Crossover Point in
Random 3sat. Artificial Intelligence 81, 31–57 (1996)
 Optimizing the Cloud Service Experience Using Constraint Programming 637

10. Dincbas, M., Simonis, H.: APACHE - a constraint based, automated stand alloca-
tion system. In: Proc. of Advanced Software technology in Air Transport (ASTAIR
1991) (1991)
11. Dubos, A., Du Jeu, A.: Application EPPER planification des agents roulants. In:
Proc. CHIP users club (1996)
12. Durbin, S.D., Warner, D., Richter, J.N., Gedeon, Z.: RightNow eService Center:
Internet customer service using a self-learning knowledge base. In: Proceedings of
the Thirteenth Annual Conference of Innovative Applications of Artificial Intelli-
gence (IAAI 2002), pp. 815–821 (2002)
13. Fredman, M.L., Tarjan, R.E.: Fibonacci Heaps and Their Uses in Improved Net-
work Optimization Algorithms. J. ACM 34(3), 596–615 (1987)
14. Hogg, T., Huberman, B.A., Williams, C.P.: Phase Transitions and the Search Prob-
lem. Artif. Intell. 81(1–2), 1–15 (1996)
15. Katriel, I., Thiel, S.: Complete Bound Consistency for the Global Cardinality Con-
straint. Constraints 10(3), 191–217 (2005)
16. Kirkpatrick, S., Selman, B.: Critical Behavior in the Satisfiability of Random
Boolean Expressions. Science 264(5163), 1297–1301 (1994)
17. Kuhn, H.W.: The Hungarian method for the assignment problem. Naval Research
Logistics Quarterly 2(1–2), 83–97 (1955)
18. Munkres, J.: Algorithms for the Assignment and Transportation Problems. Journal
of the Society for Industrial and Applied Mathematics 5(1), 32–38 (1957)
19. Naveh, Y., Richter, Y., Altshuler, Y., Gresh, D.L., Connors, D.P.: Workforce opti-
mization: Identification and assignment of professional workers using constraint
programming. IBM Journal of Research and Development 51(3/4), 263–280 (2007)
20. Nightingale, P.: The extended global cardinality constraint: An empirical survey.
Artif. Intell. 175(2), 586–614 (2011)
21. Oplobedu, A., Marcovitch, J., Tourbier, Y.: CHARME: Un langage industriel de
programmation par contraintes, illustré par une application chez Renault. In: Pro-
ceedings of the Ninth International Workshop on Expert Systems and their Appli-
cations: General Conferencehnical, vol. 1, pp. 55–70 (1989)
22. Pentico, D.W.: Assignment problems: A golden anniversary survey. European Jour-
nal of Operational Research 176(2), 774–793 (2007)
23. Perrett, M.: Using constraint logic programming techniques in container port plan-
nings. ICL Technical Journal, 537–545 (1991)
24. Quimper, C.-G., Golynski, A., Lopez-Ortiz, A., van Beek, P.: An Efficient Bounds
Consistency Algorithm for the Global Cardinality Constraint. Constraints 10(2),
115–135 (2005)
25. Régin, J.C., Gomes, C.P.: The Cardinality Matrix Constraint. In: 10th Interna-
tional Conference on Principles and Practice of Constraint Programming, CP 2004,
pp. 572–587 (2004)
26. Richter, Y., Naveh, Y., Gresh, D.L., Connors, D.P.: Optimatch: applying con-
straint programming to workforce management of highly skilled employees. Int. J.
of Services Operations and Informatics 3(3/4), 258–270 (2008)
27. Rossi, F., van Beek, P., Walsh, T.: Handbook of Constraint Programming. Elsevier
(2006)
28. Yang, R.: Solving a Workforce Management Problem with Constraint Program-
ming. Technical Report, University of Bristol, Bristol (1996)
 Find Your Way Back: Mobility Profile Mining
with Constraints

Lars Kotthoff1(B) , Mirco Nanni2 , Riccardo Guidotti2 , and Barry O’Sullivan3

1
University of British Columbia, Vancouver, Canada
larsko@cs.ubc.ca
2
KDDLab – ISTI-CNR and CS Department, University of Pisa, Pisa, Italy
{mirco.nanni,riccardo.guidotti}@isti.cnr.it
3
Insight Centre for Data Analytics, University College Cork, Cork, Ireland
barry.osullivan@insight-centre.org

Abstract. Mobility profile mining is a data mining task that can be for-
mulated as clustering over movement trajectory data. The main challenge
is to separate the signal from the noise, i.e. one-off trips. We show that
standard data mining approaches suffer the important drawback that
they cannot take the symmetry of non-noise trajectories into account.
That is, if a trajectory has a symmetric equivalent that covers the same
trip in the reverse direction, it should become more likely that neither of
them is labelled as noise. We present a constraint model that takes this
knowledge into account to produce better clusters. We show the efficacy
of our approach on real-world data that was previously processed using
standard data mining techniques.

1 Introduction
Clustering is one of the fundamental tasks in data mining whose general aim
is to discover structure hidden in the data, and whose means consist in iden-
tifying the homogeneous groups of objects (the clusters) that emerge from the
data [7]. Typically, this translates into putting together objects that are similar
to each other, and keep separated as much as possible those that are different.
The applications of this basic task are many and varied, ranging from customer
segmentation for business intelligence to the analysis of images, time series, web
search results and much more.
This paper focuses on the application domain of mobility data mining, which
is concerned with the study of trajectories of moving objects and other kinds of
mobility-related data [6]. Trajectories are sequences of time-stamped locations
(typically longitude-latitude pairs) that describe the movements of an object. A
most relevant example, which will also be the reference context for this work,
is the sequence of GPS traces of cars collected by on-board devices, such as a
satellite navigation system or an anti-theft platform [5].
In this context, clustering can be used to identify popular paths followed by
several moving objects, such as common routes adopted by car drivers in a city.
A different perspective to the problem was proposed in recent years in [8]: the

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 638–653, 2015.
DOI: 10.1007/978-3-319-23219-5 44
 Find Your Way Back: Mobility Profile Mining with Constraints 639

objective of the analysis is not directly the collective mobility of a population,

but instead passes through an intermediate phase where the mobility of the
single individual is studied first. Clustering the trajectories of an individual has
the main purpose of highlighting which movements repeat consistently in his/her
life, and therefore can be considered an important and representative part of the
user’s mobility. In the realm of urban traffic and transportation engineering such
movements constitute the systematic component of the mobility of a person –
and then of a city, as a consequence.
Identifying such mobility profiles is an important current research direction
that has many applications, for example in urban planning [4]. If trips that
individuals make frequently in their cars can be identified, the schedule and
routes of the public transport system could be adjusted to provide people with
an incentive to leave the car at home and do something for the environment.
In this paper, we study the problem from a constraint programming point
of view. The trajectories that form part of mobility profiles have to conform to
certain restrictions that are hard to incorporate in traditional data mining algo-
rithms, but easy to express as constraints. We present a constraint programming
model and demonstrate its benefits over previous approaches.
There are a number of approaches that model data mining problems with
constraints (e.g. [1,2,9]), but to the best of our knowledge, this is the first time
that this particular problem is studied in this context.

2 Drawbacks of the Pure Data Mining Approach

Existing data mining approaches to extract mobility profiles are usually cus-
tomised algorithms that have been developed specifically for this purpose. The
approach proposed in [8] (schematically depicted in Figure 1) consists of two
ingredients: a trajectory distance function that encapsulates the context-specific
notion of similar trajectories and a clustering schema that takes care of group-
ing similar trajectories together (center of Figure 1). It also ensures that the
resulting groups have a certain minimum size, e.g. the cluster composed of the
single dashed trajectory is removed (center of Figure 1). Once clusters have
been formed, a representative trajectory for each of them is simply extracted by
selecting the most central element of the cluster, i.e. the one that minimizes the
sum of distances from the others (see the selected trajectories on the right of
Figure 1).
We impose the minimum size constraint because we aim to identify the trips
that are made regularly and are not interested in trips that occur regularly, but
infrequently (such as a trip to a holiday home). The clustering procedure yields
a list of trip groups, each of them essentially representing a mobility routine
followed by the user. The set of routines of a user form the mobility profile of
that user.
A very typical (and expected) result is that an individual’s mobility contains
at least two routines, one for the home-to-work systematic trips and one for
the symmetric work-to-home return journey. Indeed, symmetric routines (the
640 L. Kotthoff et al.

Fig. 1. The three steps of profile extraction: identify the trajectories (single trips)
from the GPS data, discover clusters by grouping trajectories that are close, refine the
clusters to remove the noise and extract representative routines.

home-work-home cycle being the most typical example) appear very naturally
in real life, and they are a very important component of an individual’s mobility.
Therefore, while it is simply important to discover all the routines hidden in an
individual’s mobility data, it is crucial to discover symmetric routines if they
exist.
From the viewpoint of a traffic management application for example, sym-
metric routines represent simple and regular mobility cycles that might neatly
fit a public transportation offer. Asymmetric routines on the other hand are
generally a symptom of a more complex daily mobility, for instance involving
bring-and-get activities interleaved within a home-work-home cycle, which are
much more difficult to serve by a public transportation service.
These observations lead to conclude that it would be advisable to equip
the routine extraction procedure with mechanisms that boost the discovery of
symmetric routines. Unfortunately, the plain clustering-based approach provided
so far in [8] follows a local perspective, focused on guaranteeing properties of
each single cluster – namely, mutual similarities of its members and a certain
minimum size of the cluster – without considering the relation between different
clusters or trajectories in different clusters.
Therefore, symmetric routines, which correspond to separate clusters even
though they are connected by this particular property, are treated in isolation.
What we would need to stimulate the discovery of symmetric routines, is a
way to link such separate clusters, and change the cluster construction criteria
accordingly. The purpose of this paper is to explore this way by exploiting a con-
straint formulation of the clustering problem, which can then be easily enriched
and customized to include the application-specific requirements we mentioned
above.

Setting the Minimum Cluster Size

Being able to take the symmetry of trajectories into account provides an addi-
tional benefit when clustering. The minimum size of a cluster is currently
enforced by providing a static value to the algorithm that applies to all clus-
ters. This value needs to be chosen carefully – if it is too high, no clusters at
all may be identified and if it is too low, too many clusters will be identified. In
particular, the value should be set such that trajectories that are noise as far as
a mobility profile is concerned are not identified as clusters.
 Find Your Way Back: Mobility Profile Mining with Constraints 641

In practice, symmetric trajectories are often part of clusters of different sizes.

There are many reasons for this. Consider for example the home-to-work routine,
where an individual drives straight from home to work every day, but on the way
back goes to the gym on certain days and shopping on others. Therefore, the
size of the home-to-work cluster will be larger than that of the symmetric work-
to-home cluster.
If the threshold for the minimum size of a cluster is set higher than the size
of the work-to-home cluster, it will be categorized as noise even though there is a
symmetric equivalent that supports it. Indeed, there could be cases like this with
symmetric clusters of different sizes where no fixed value for the minimum size
threshold will ensure that all symmetric clusters are identified while the noise
trajectories are not.
In a constraint model, this can be taken into account easily. Instead of con-
sidering only the size of a cluster to justify its existence, we can explicitly model
symmetric clusters as supports as well.

3 Constraint Model
We present the constraint model that takes all the considerations mentioned
above into account. The constraints encode when two trajectories should be in
the same cluster, when they should be in different clusters, when they are noise
and the conditions that apply when there are symmetric trajectories.
For the sake of simplicity, trajectories are represented by a start and an end
point. For our purposes, the route taken to get from the start to the end point
is secondary, as we are interested in the extraction of mobility profiles that do
not contain this information. For example, the way from home to work could be
different from the way from work to home because of traffic conditions, one-way
streets, or similar, but both are still part of the same routine. If the user stops,
e.g. to do some shopping, the trajectory is split into two parts with two separate
start and end points.
The model itself is general enough to accommodate routes represented by
more than two points as well, but would likely reduce the number of symmetric
trajectories as pairs would look less similar because of additional (irrelevant)
information. Overall, the model would become larger and more complex.
Individual trajectories x ∈ X are modelled as variables whose domains con-
sist of the clusters they can be assigned to. That is, each trajectory can be
assigned a cluster ID, including a special value which denotes that the trajectory
is noise. We use dist(x, y) to denote the distance (both spatial and temporal)
between two trajectories x and y. Note that for the temporal distance, only
the time of day is relevant, not the date. The reason for this is that we want
to cluster trips that occur regularly, e.g. the daily commute to work. The dis-
tance of two trajectories is defined as the spatial and temporal distance between
their start points and their end points, so dist(x, y) < A is short hand for
dist(xstart , ystart ) < A ∧ dist(xend , yend ) < A. T denotes the distance above
which two trajectories are considered to be not close and R the distance above
642 L. Kotthoff et al.

which they are considered far, T < R. If a trajectory is noise and should not be
part of any cluster, it is denoted x = noise.
Our formulation consists of two complementary parts. The first part is a set
of constraints that specify under which conditions two trajectories should belong
to different clusters or be labelled as noise. The second one is a multi-objective
optimization function that requires to cluster as many trajectories as possible
and to fit them into as few clusters as possible. This requires the identified
clusters to be as large as possible.
Our optimization function can be formulated as follows, where we use X
to denote the cardinality of the set X:

minimize unique({x | x ∈ X}) (no. distinct clusters)

(1)
minimize {x | x ∈ X, x = noise} (no. noise trajectories)

Our constraint model includes three kinds of constraints. First we specify

when two trajectories should not be in the same cluster:

∀ x, y ∈ X : dist(x, y) > R → x = y ∨ x = noise ∨ y = noise (2)

If two trajectories are far apart, they should be in different clusters or at
least one of them should be noise. In our application, we are trying to identify
mobility patterns which are defined by clusters of trajectories that are all close
together. We therefore want to avoid clustering trajectories that are close to a
set of intermediate trajectories, but far apart themselves.
Symmetric trajectories x and y are denoted as symm(x, y) and defined as
follows:

∀x, y ∈ X : symm(x, y) ≡ dist(xstart , yend ) ≤ T ∧

(3)
dist(xend , ystart ) ≤ T

We now introduce the symmetry constraints that represent the main advan-
tage of the constraint model over existing data mining algorithms. We denote
the threshold for the minimum number of trajectories in a cluster S. This is
a fixed value specified by the user and potentially prevents identifying clusters
that are small, but supported by a symmetric cluster. We therefore introduce the
threshold size for a pair of symmetric clusters S  , where S  > S. If a trajectory
x is symmetric to another trajectory y, they should belong to different clusters
if they are not noise and the sum of the sizes of the clusters they belong to must
be greater than S  .

∀ x, y ∈ X : symm(x, y) ∧ x = noise ∧ y = noise

(4)
∧ ({z | z ∈ X, z = x} + {z | z ∈ X, z = y} > S  ) → x = y

This set of constraints encodes most of the background knowledge on symmetric

trajectories. We both take symmetry into account explicitly and mitigate the
 Find Your Way Back: Mobility Profile Mining with Constraints 643

drawbacks of having a fixed minimum cluster size. The fixed threshold S is the
only means of separating signal from noise in the data mining application. In the
constraint model, we can relax this requirement in the presence of symmetric
clusters – if a symmetric pattern is identified, we require the sum of the cluster
sizes to be greater than the threshold for a pair of symmetric clusters, which is
greater than the threshold for a single cluster.
If, on the other hand, there is no trajectory that is symmetric to x, then x
is either noise or the size of the cluster that it is assigned to is greater than the
threshold S for the size of a single cluster.

∀ x ∈ X : {y | y ∈ X, symm(x, y)} = 0 →

(5)
x = noise ∨ (x = noise ∧ {z | z ∈ X, z = x} > S)

These constraints describe the requirements we place on trajectories to be

part of clusters as well as considering the relation between clusters by taking the
symmetry into account. We do not rely on any special constraints or esoteric
constructs – our model is generic and can be implemented and solved with
almost any constraint solver. In contrast, the data mining algorithm requires
a specialised implementation that, while being able to leverage existing work,
needs considerable efforts.

4 Model Implementation

We implemented the constraint model described above in the Minion constraint

solver [3]. Minion is a general-purpose constraint solver. Our choice of solver was
motivated by nothing but the authors’ familiarity with this particular solver and
its free availability.

4.1 Minion Model

Our Minion model is an almost direct translation of the model presented in the
previous section. For each trajectory, we add one variable whose domain values
represent the clusters the trajectory can be assigned to. We also add the dual
representation where each cluster is represented by an array of Booleans, one
for each trajectory. If a particular trajectory is in a particular cluster, the corre-
sponding Boolean value is set to 1, else 0. We also add the channelling constraints
between the one-variable-per-trajectory and the one-array-per-cluster represen-
tations. We require the dual representation to be able to post constraints on
cluster sizes.
The noise trajectories are assigned to a special cluster that is represented by
the value 0 in the domains of the trajectory variables. Modelling noise this way
allows us to treat it in the same way as other clusters.
The Minion input language is flat and does not support quantifications over
the variables. We therefore instantiate the free variables in Equations 2 to 5
and add constraints for each grounding. Note that the left hand side (before the
644 L. Kotthoff et al.

implication) in all of these equations except 4 can be computed statically before

solving the problem as it only depends on the trajectory data. We add constraints
corresponding to the right hand side of the implications to the Minion model
only if the left hand side is true. This preprocessing step helps to reduce the
size of the Minion model, which is potentially very large because of the nested
quantifications.
The implication on the right hand side of Equation 4 is modelled with the
help of auxiliary variables that represent the left hand side of the implication
being true. We need auxiliary variables here because Minion does not support
implications between constraints, but only between a constraint and a variable.
All variables and constraints are treated the same way by Minion and arc con-
sistency is enforced during search. We use Minion’s default parameters.
The thresholds T , R, S and S  are specified by the user. In the experiments
section we will see the effect of these parameters on the results.

4.2 Optimising the Clustering

The optimisation objective of mobility profile mining is to cluster as many tra-
jectories as possible, subject to the constraints, while minimising the overall
number of clusters. We cannot easily express this multi-objective optimisation
in our Minion model, but it turns out that we do not have to. We instead solve
the simpler satisfaction problem that is defined as follows. Given a number of
clusters, we assign values to trajectory variables in a descending manner. That
is, we start with the highest cluster and assign a trajectory to the noise cluster
(value 0) only if all other possibilities have been exhausted. This ensures that as
many trajectories as possible are clustered as non-noise. In Minion, we simply
specify a descending value ordering for the search.
We now minimise the number of clusters as follows. We start by specifying
only two clusters in addition to the noise cluster. If the constraint satisfaction
problem (CSP) has a solution, we have found our clustering. If there are more
symmetric trajectories that need to be in different non-noise clusters however,
this first CSP will have no solution. In this case, we increase the number of
clusters by one, generate a new CSP and try to solve it. This way, we identify
the solution with minimum number of clusters.
In practice, the number of clusters is usually small. The number of routine
trips an individual makes that we have observed empirically was fewer than five
in most cases, such that our method is likely to find the solution after only a
few iterations.

4.3 Scalability
The main reason why data mining and machine learning applications use spe-
cialised approximate algorithms instead of complete optimisation approaches
such as constraint programming is the scalability of such methods. Approxi-
mate methods are almost always orders of magnitude faster than corresponding
complete methods while delivering solutions of similar quality.
 Find Your Way Back: Mobility Profile Mining with Constraints 645

While scalability is a concern for our application as well, it is not an issue in

practice, due to the high modularity of the problem: indeed, each user can be
analyzed separately, making the complexity of the problem linear in the number
of users. The critical factor is, instead, the number of trajectories of a single user.
However, the vast majority of individuals have a relatively low number of trips
(in the dataset used for our experiments, which covers a time span of 5 weeks,
most users have fewer than 100 trajectories), and therefore the constraint models
can be solved in reasonable time. Our application is not time-critical, but puts
the main emphasis on the quality of the solutions. We are happy to spend more
time solving the problem as long as we get a better quality result.
The main factor that contributes to the size of the CSP apart from the
number of trajectories is the number of clusters. Each additional cluster increases
the size of the domains of the trajectory variables, requires an additional array
of Boolean values and additional constraints to account for the possibility of a
trajectory being assigned to that cluster. The iterative approach for setting the
number of clusters described above keeps the size of the CSP as small as possible
while not compromising on the quality of the solution.
We have found that this approach has a very significant impact on the effi-
ciency with which we can solve this clustering problem. Often, increasing the
number of clusters beyond the optimal number increases the solve time of the
model by orders of magnitude.

5 Experimental Evaluation

We evaluated the quality of the clusters that the constraint model finds on
real-world trajectories collected in Italy. We will show how our constraint model
improves on the existing solution for extracting mobility profiles, thus evaluating
the impact of taking symmetric clusters into account on the results. We briefly
describe the dataset used in the experiments, then summarize the results, and
finally show an example of profiles found on a single user with the different
approaches.

5.1 Dataset

Our dataset contains the trajectories of the cars of 150 users living around the
city of Pisa, Italy, collected over a period of 5 weeks (Figure 2). We cannot make
the actual data available because of privacy reasons; an anonymized version
is available at http://www.cs.ubc.ca/∼larsko/downloads/cp2015 anonymized
dataset.tar.gz.
The users were selected in such a way to have a good representation of differ-
ent mobility complexities, i.e. different numbers of trajectories. In our stratified
sample, the number of trajectories per user roughly follows a uniform distribu-
tion that ranges from 1 to 100 (see Figure 3).
Experiments were carried out adopting a spatio-temporal distance dist(x, y)
between two points that is computed as the standard Euclidean distance between
646 L. Kotthoff et al.

Fig. 2. Dataset of trajectories used in the experiments, containing 150 users centered
around Pisa, Italy.

Fig. 3. Distribution of number of trajectories per user used in the experiments (blue
bars) vs. distribution of original dataset (green line).
 Find Your Way Back: Mobility Profile Mining with Constraints 647

the spatial components of x and y if their temporal gap is less than a threshold
τ = 1hr, and ∞ otherwise. Therefore, we used the following ranges of values
for parameters S (the minimum number of trajectories in a cluster) and T (the
distance threshold between trajectories, expressed in meters): S ∈ {3, 4, 5, 6},
T ∈ {100, 200, 300, 400, 500}. Moreover, the parameters S  (the threshold for
the combined size of two symmetric clusters) and R (the maximum distance
between trajectories in a cluster) are set as functions of S and T , respectively:
S  = 1.5 · S, R = 3 · T . The coefficients used to compute S  and R have been
chosen empirically through a set of preliminary experiments over the dataset.
Different datasets may exhibit different properties and a recalibration of these
values may be required.
Computation over a medium-low-end desktop machine (i3 CPU at 3GHz,
with 4GB RAM, running Ubuntu 12) took an overall time below 15 minutes. In
particular, single users with 20 or less trajectories (which are the most common
in the original dataset) are dealt with in less than one second, while those with
50 to 100 trajectories (less common, yet over-represented in our sample set, to
evaluate their impact) took up to 15 seconds each.

5.2 Evaluation of Results

The main advantage of our constraint model over previous approaches is that it
allows to take symmetric clusters into account. Without the constraints that do
this, we should achieve results that are equivalent to the data mining solution
proposed in [8]. However, the constraint model should find an optimal solution
instead of a heuristic one.
In our first set of experiments, we compare both approaches without taking
symmetric clusters into account to get a baseline for the performance of our
new approach. This also allows to evaluate the impact of taking the symmetric
clusters into account later.
Figures 4 and 5 show a comparison of the two approaches for the different
parameter values (S and T ) in terms of the number of non-noise trajectories
(Figure 4) and of number of clusters found (Figure 5).
Both the number of non-noise trajectories and clusters increases as we relax
the constraints, i.e. allow fewer trajectories that are further apart in a cluster.
This is true for both the data mining and the CP approach. In general, both the
number of non-noise trajectories and clusters is larger for the CP approach, in
some cases significantly so. For example, the number of clusters for S = 3 that the
CP approach finds is almost twice the number the data mining approach finds. As
S is increased, the difference becomes smaller. The reason for this phenomenon
is that the data mining approach partitions the set of trajectories into clusters
through a heuristics that considers the distribution of data only locally, and
therefore can create small clusters that are later discarded, as shown in the
example of Figure 1. This kind of situations are avoided by the CP approach,
which instead considers the assignment of trajectories to clusters from a global
viewpoint.
648 L. Kotthoff et al.

Fig. 4. Number of clustered trajectories found by the data mining approach (dm) and
the constraint model without symmetry constraints (CPnosym ) for different distance
threshold and cluster size values T and S.

In our second set of experiments, we focus on evaluating the impact of taking

symmetric clusters into account. We adopt two measures for comparing the
complete constraint model with all constraints to the one that does not take
symmetric clusters into account as above.

– the trajectory coverage of method A over method B, defined as:

{x | x ∈ XA , x = noise} ∩ {x | x ∈ XB , x = noise}

tr cov(A, B) = (6)
{x | x ∈ XB , x = noise}

where XA (resp. XB ) represents the overall set of clustered trajecto-

ries obtained with method A (resp. B). When A completely covers B,
tr cov(A, B) = 1.
– the cluster coverage of method A over method B, defined as:

(Cu · cl covu (A, B))
cl cov(A, B) = u∈U  (7)
u∈U Cu

where U is the set of users and Cu the number of clusters found by A for
user u. Function cl covu (A, B), in turn, is computed over each user u as the
average trajectory coverage of each cluster of u found by A w.r.t. those found
by B. The matching between clusters is performed by taking the best match,
 Find Your Way Back: Mobility Profile Mining with Constraints 649

Fig. 5. Number of clusters found by the data mining approach (dm) and the constraint
model without symmetry constraints (CPnosym ) for different distance threshold and
cluster size values T and S.

i.e. maximizing the trajectory coverage limited to the two clusters compared.
Clusters left unpaired are assigned to an empty set. When A completely
covers B, cl cov(A, B) = 1.

While the trajectory coverage compares the overall ability to recognize sys-
tematic trips, the cluster coverage looks at how much each single cluster is dis-
covered by the other method. Figures 6 and 7 report the results obtained on
our dataset. Lines labelled with SCPsym represent the coverages of the model
with symmetry constraint (CPsym ) over that without symmetry (CPnosym ),
and SCPnosym the opposite.
The plots clearly show that the model that takes symmetric trajectories
into account completely covers the other one, both in global terms of trajectory
coverage (Figure 6) and in terms of (the more precise) cluster coverage (Figure 7),
since the corresponding values are always very close to 1. The model that does
not take symmetric trajectories into account on the other hand loses at least 50%
of trajectories and 70% of cluster coverage. These results are consistent across
the range of values for S and T that we consider here, with little variation.
Our results show that considering the symmetry of clusters has a large impact
on results. We are able to cluster a lot more trajectories that were mistakenly
identified as noise before, allowing to recover several backward trips that would
otherwise not be recognized as routines.
650 L. Kotthoff et al.

Fig. 6. Trajectory coverage for constraint models with symmetry constraint (CPsym )
and without (CPsym ) for different distance threshold and cluster size values T and S.

Fig. 7. Cluster coverage for constraint models with symmetry constraint (CPsym ) and
without (CPsym ) for different distance threshold and cluster size values T and S.
 Find Your Way Back: Mobility Profile Mining with Constraints 651

5.3 Qualitative Evaluation

The quality of the clustering is more difficult to assess, as there are no auto-
mated means to determine whether a set of clusters “makes sense”. In practice,
clusterings are evaluated through assessment by a human expert. We obviously
cannot do this for all of the users we consider in this paper and consider the
clusters for a single user as an example here.
While not all users have symmetric trajectories, the following example
demonstrates the motivation of our approach and the case that is difficult to
solve for specialised data mining methods. A clustering example that shows the
benefit of the constraint model is presented in Figure 8. The image was obtained
by using T = 100 and S = 6.
The red and the blue trajectories represent symmetric parts of the same
routine. However, there are fewer red than blue trajectories; in particular the
number of red trajectories is smaller than the threshold S. If the symmetry is not
taken into account, only the blue trajectories are clustered, whereas the red ones
are discarded as noise. The CP approach that does consider symmetry identifies
the red trajectories correctly as a non-noise cluster.

Fig. 8. Trajectory clusters found on a sample user (best seen in colors).

For the eastward journey, our user mainly follows the blue routine, but for
the return journey he/she often changes path, either for traffic reasons or for
personal needs, and the red trajectories represent the most frequent choice. The
figure clearly shows the benefit of being able to take symmetric trajectories into
account. Indeed, there are only three trajectories in the red cluster, which has
been discarded as noise by the other method.
652 L. Kotthoff et al.

Lowering S to 3 would have clustered the red trajectories as non-noise, but

at the same time erroneously identified a number of additional clusters that are
in fact noise for our purposes. Therefore, the model that does not consider sym-
metric trajectories (as well as the specialised data mining algorithm discussed
above) could not have identified the same set of clusters for this user.

6 Conclusion and Future Work

We have presented a constraint programming formulation of the problem of

mining mobility profiles from GPS trajectories. This is an important research
direction with many real-world applications that, to date, has been the domain
of specialised data mining algorithms. We have described the drawbacks of these
specialised algorithms and how they are alleviated with constraint programming.
The implementation of our model and its experimental evaluation on real-
world data showed that the model stands up to its promise and delivers clus-
terings of a higher quality than those that the previously-used specialised data
mining algorithms achieve. In particular, we are able to identify a much larger
number of non-noise trajectories and clusters. The experiments further demon-
strate that it is feasible to use complete optimisation methods for this problem.
The main benefit of constraint programming over similar technologies for
this application is the succinct representation of the model. In particular SAT
models would be much larger to accommodate the potentially large domains
encoding the possible cluster assignments. Similarly, ILP formulations would be
more complex because of nested constructs.
There are several avenues for future work. We did not add any symmetry-
breaking constraints to our model even though symmetries occur – for example
in the numbering of the clusters. Formulating and adding symmetry-breaking
constraints may increase the scalability of our approach further.
On the data mining side, there are additional problem characteristics that
are not taken into account in our current model. The day that particular trips
were taken for example is not considered, but it would be beneficial to do so.
We could for instance boost symmetric trajectories that occur on the same day,
as they are more likely to represent a recurring activity.
There are many additional extensions to the model presented in this paper.
Even our relatively basic formulation shows the promise of the approach by
delivering results of a higher quality than specialised approaches while being
easier to maintain and modify.

Acknowledgments. This work was supported by the European Commission as part

of the “ICON - Inductive Constraint Programming” project (contract number FP7-
284715). The Insight Centre for Data Analytics is supported by Science Foundation
Ireland under grant number SFI/12/RC/2289. We thank the anonymous reviewers for
their helpful feedback.
 Find Your Way Back: Mobility Profile Mining with Constraints 653

References
1. Davidson, I., Ravi, S.S., Shamis, L.: A SAT-based Framework for Efficient Con-
strained Clustering. In: SIAM International Conference on Data Mining, pp. 94–105
(2010)
2. De Raedt, L., Guns, T., Nijssen, S.: Constraint programming for data mining and
machine learning. In: AAAI, pp. 1671–1675 (2010)
3. Gent, I.P., Jefferson, C.A., Miguel, I.: MINION: a fast, scalable, constraint solver.
In: ECAI. pp. 98–102. IOS Press (2006)
4. Giannotti, F., Nanni, M., Pedreschi, D., Pinelli, F.: Trajectory pattern analysis for
urban traffic. In: Proceedings of the Second International Workshop on Computa-
tional Transportation Science, IWCTS 2009, pp. 43–47 (2009)
5. Giannotti, F., Nanni, M., Pedreschi, D., Pinelli, F., Renso, C., Rinzivillo, S.,
Trasarti, R.: Unveiling the complexity of human mobility by querying and min-
ing massive trajectory data. The VLDB Journal 20(5), 695–719 (2011). http://dx.
doi.org/10.1007/s00778-011-0244-8
6. Giannotti, F., Pedreschi, D. (eds.): Mobility, Data Mining and Privacy - Geographic
Knowledge Discovery. Springer (2008)
7. Tan, P.N., Steinbach, M., Kumar, V.: Introduction to Data Mining. Addison-Wesley
(2005)
8. Trasarti, R., Pinelli, F., Nanni, M., Giannotti, F.: Mining mobility user profiles for
car pooling. In: KDD, pp. 1190–1198 (2011)
9. Wagstaff, K., Cardie, C.: Clustering with Instance-level Constraints. In: Seventeenth
International Conference on Machine Learning, ICML 2000, pp. 1103–1110. Morgan
Kaufmann Publishers Inc., San Francisco, CA, USA (2000)
Joint Vehicle and Crew Routing and Scheduling

Edward Lam1,2(B) , Pascal Van Hentenryck1,3 , and Philip Kilby1,3

1
NICTA, Eveleigh, NSW 2015, Australia
edward.lam@nicta.com.au
2
University of Melbourne, Parkville, VIC 3010, Australia
3
Australian National University, Acton, ACT 2601, Australia

Abstract. Traditional vehicle routing problems implicitly assume only

one crew operates a vehicle for the entirety of its journey. However, this
assumption is violated in many applications arising in humanitarian and
military logistics. This paper considers a Joint Vehicle and Crew Routing
and Scheduling Problem, in which crews are able to interchange vehicles,
resulting in space and time interdependencies between vehicle routes and
crew routes. It proposes a constraint programming model that overlays
crew routing constraints over a standard vehicle routing problem. The
constraint programming model uses a novel optimization constraint that
detects infeasibility and bounds crew objectives. Experimental results
demonstrate significant benefits of using constraint programming over
mixed integer programming and a vehicle-then-crew sequential approach.

1 Introduction

A vehicle routing problem (VRP) aims to design routes for a fleet of vehicles that
minimize some cost measure, and perhaps, while adhering to side constraints,
such as time windows, or pickup and delivery constraints. In practice, however,
VRPs are not solved in isolation; they typically arise in a sequential optimization
process that first optimizes vehicle routes and then crew schedules given the vehi-
cle routes [33]. This sequential methodology has the advantage of reducing the
computational complexity. However, by assigning vehicle routes first, this app-
roach may lead to suboptimal, or even infeasible, crew schedules since decisions
in the first stage ignore crew constraints and objectives. Hence it is imperative to
simultaneously consider vehicle and crew constraints and objectives, particularly
in industries in which crew costs outstrip vehicle costs.
This paper considers a Joint Vehicle and Crew Routing and Scheduling Prob-
lem (JVCRSP) motivated by applications in humanitarian and military logistics.
In these settings, vehicles (e.g., airplanes) travel long routes, and serve a variety
of pickup and delivery requests under various side constraints. Vehicles must be
operated by crews who have limitations on their operating times. Crews are also
able to interchange vehicles at different locations and to travel as passengers
before and after their operating times. JVCRSPs are extremely challenging com-
putationally since vehicle and crew routes are now interdependent. In addition,
vehicle interchanges add an additional time complexity, since two vehicles must

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 654–670, 2015.
DOI: 10.1007/978-3-319-23219-5 45
 Joint Vehicle and Crew Routing and Scheduling 655

be synchronized to allow for the exchange to proceed. It is thus necessary to

decide whether vehicles wait, and the waiting duration, since both vehicles must
be present at the same location for an interchange to take place.
This paper proposes a constraint programming formulation of the JVCRSP
that jointly optimizes vehicle and crew routing in the hope of remedying some
limitations of a multi-stage model. The formulation overlays crew routing con-
straints over a traditional vehicle routing problem, and also adds a number of
synchronization constraints to link vehicles and crews. In addition, the formula-
tion includes a novel optimization constraint that checks feasibility and bounds
crew costs early in the search while the focus is on vehicle routing. The con-
straint programming model is solved using a large neighborhood search (LNS)
that explores both vehicle and crew neighborhoods.
The constraint program is compared to a mixed integer program and a two-
stage method that separates the vehicle routing and crew scheduling problems.
Experimental results on instances with up to 96 requests and different cost
functions indicate that (1) the joint optimization of vehicle and crew routing
produces considerable benefits over the sequential method, (2) the constraint
program scales significantly better than the mixed integer program, and (3)
vehicle interchanges are critical in obtaining high-quality solutions. These pre-
liminary findings indicate that it is now in the realm of optimization technology
to jointly optimize vehicle and crew routing and scheduling, and that constraint
programming has a key role to play in this promising direction.
The rest of this paper is organized as follows. Section 2 describes the Joint
Vehicle and Crew Routing and Scheduling Problem and some high-level mod-
eling decisions. Section 3 outlines the two-stage methodology, while Section 4
describes the mixed integer formulation of the JVCRSP. Section 5 details the
constraint programming formulation, two novel global constraints, and the search
procedure. Section 6 reports the experimental results. Section 7 reviews related
work and Section 8 concludes this paper.

2 Problem Description
The traditional vehicle routing problem with pickup and delivery and time win-
dows (VRPPDTW) [30] consists of a fleet of vehicles, stationed at a common
depot, that services pickup and delivery requests before returning to the depot.
Every pickup request has one associated delivery request, with an equal but
negated demand, that must be served after the pickup and by the same vehicle.
In addition, every request has a time window, within which service must begin.
Vehicles can arrive earlier than the time window, but must wait until the time
window opens before starting service. Given a route, fixed travel times dictate
the arrival and departure times of requests along the route.
The JVCRSP considered in this paper overlays, on top of the VRPPDTW,
crew routing constraints that explicitly capture the movements of crews. The
JVCRSP groups requests by location, and defines travel times between loca-
tions, thereby removing the one-to-one mapping between requests and locations
656 E. Lam et al.

1 3 5 7 9

2 4 6 8 10

Fig. 1. A location at which vehicle interchange occurs. Nodes 3 to 8 belong to the same
location. A vehicle (solid line) and a crew (dotted line) travel from node 1 to node 3,
and another vehicle and crew travel from node 2 to node 4. The two vehicles service
nodes 3 to 8, while the crews travel to nodes 7 and 8 to depart on a different vehicle.
The movements into nodes 3 and 4, and out of nodes 7 and 8 must be synchronized
between the vehicles and crews, but movements within the location are independent.

found in many routing problems. Vehicles require a crew when moving between
two locations, but require no crew when servicing requests within a location.
Similarly, crews require a vehicle to move between two locations, but can move
independently of vehicles within a location. Crews are restricted to at most one
driving shift, which has a maximum duration. No distance or time limitations
are placed on crews before and after the driving shift.
The JVCRSP allows crews to switch vehicles at any given location. This
idea is illustrated in Fig. 1. Vehicle interchange occurs when a crew moves from
a node that is serviced by one vehicle to a node that is serviced by another
vehicle, provided that the departure of the second vehicle is after the arrival
of the first. Hence vehicle interchange requires explicit modelling of arrival and
departure times. The JVCRSP retains the time windows on service start times
and permits early arrival, but also allows vehicles to wait after service to depart
at a later time. This functionality is necessary to facilitate vehicle interchange.
The objective function of the JVCRSP minimizes a weighted sum of the num-
ber of vehicles and crews used, and the total vehicle and crew travel distances.

3 A Two-Stage Method for the JVCRSP

One of the goals of this paper is to assess the benefits of jointly routing and
scheduling vehicles and crews. A baseline to determine the potential of joint
optimization is a two-stage approach that separates these two aspects by first
solving a vehicle routing problem and then a crew pairing problem. The first
stage can be solved with any appropriate technology, and this paper uses a
large neighborhood search over a constraint programming model (e.g., [29,21,1]).
Arrival and departure times are fixed at their earliest possible starting times
to produce vehicle schedules. The crew pairing problem is modeled with a set
covering problem typical in the airline industry [33]. It is solved using column
generation with a resource-constrained shortest path pricing problem, and then
a mixed integer programming solver to find integer solutions once the column
generation process has converged.
 Joint Vehicle and Crew Routing and Scheduling 657

4 A Mixed Integer Program for the JVCRSP

This section sketches a mixed integer programming (MIP) formulation for the
JVCRSP. The model consists of a vehicle component and a crew component.
The vehicle component is the traditional multi-commodity flow model of the
VRPPDTW, which is omitted for familiarity and space reasons. The crew com-
ponent contains similar routing constraints, but omits travel time and request
cover constraints, since these are already included in the vehicle component. The
vehicle and crew components are linked via synchronization constraints, which
are key aspects of the formulation.
Table 1 lists the relevant inputs and decision variables. The usual vehicle flow
variables vehv,i,j are duplicated for crews; the variable crewc,i,j denotes whether
crew c traverses the edge (i, j). Additionally, the driverc,i,j variable indicates
whether crew c drives the vehicle that traverses the arc (i, j), and is used to
synchronize the vehicles and crews in space.
Figure 2 describes the key synchronization constraints that coordinate the
vehicles, crews, and drivers in the MIP model. Equations (1) and (2) state that
when moving from one location to another, a crew requires a vehicle and a vehicle
requires a driver. Equation (3) restricts the driver along an arc to be one of the
crews that traverses that arc. Equations (4) and (5) link the arrival, service, and
departure times. Equation (6) constrains the start and end nodes to a common
arrival and departure time. Equations (7) to (9) synchronize time at request,
start, and end nodes, respectively. For every crew c that drives from a request
i to any request j, Eq. (10) bounds the start of driving to be no later than the
departure time at i. Similarly, Eq. (11) bounds the end of driving. Equation (12)
restricts the driving duration of each crew.

5 A Constraint Program for the JVCRSP

This section introduces the constraint programming (CP) formulation for the
JVCRSP. It describes the instance data, decision variables, and constraints of the
model, and the specifications of two global constraints, and the search procedure.
The data and decision variables are described in Table 2. Like the MIP model,
the CP model is divided into a vehicle component and a crew component.
The vehicle component, described in Figure 3, is based on the standard con-
straint programming modeling for vehicle routing (e.g., [21,28]). Successor vari-
ables succ(·) capture the sequence of requests along vehicle routes. By artificially
joining the end of one route to the start of another (Eqs. (13) and (14)), the
successor variables then describe a Hamiltonian cycle, which enables the use of
the Circuit global constraint for subtour elimination (Eq. (15)). Equations (16)
and (17) track a vehicle along its route, and Eq. (18) determines if a vehicle is
used. Equations (19) and (20) order the arrival, service, and departure times at
each request. Equation (21) restricts the start and end nodes to one common
arrival/departure time. Equation (22) enforces travel times. Symmetry-breaking,
load, and pickup and delivery constraints are omitted due to space limitations.
658 E. Lam et al.

Table 1. Instance data and decision variables of the MIP model.

Name Description

T >0 Time horizon.

T = [0, T ] Time interval.
V Set of vehicles.
C Set of crews.
T ∈T Maximum driving duration of a crew.
s Start node.
e End end.
R Set of all requests.
N = R ∪ {s, e} Set of all nodes, including the start and end nodes.
A Arcs to be traversed by the vehicles and crews, excluding
the arc (s, e) representing an unused vehicle/crew.
L Set of locations, including one depot location.
li ∈ L Location of i ∈ N .
ai ∈ T Earliest start of service at i ∈ N .
bi ∈ T Latest start of service at i ∈ N .
ti ∈ T Service duration of i ∈ N .

vehv,i,j ∈ {0, 1} 1 if vehicle v ∈ V traverses (i, j) ∈ A ∪ {(s, e)}.

crewc,i,j ∈ {0, 1} 1 if crew c ∈ C traverses (i, j) ∈ A ∪ {(s, e)}.
driverc,i,j ∈ {0, 1} 1 if crew c ∈ C drives on (i, j) ∈ A, li = lj .
arrv,i ∈ T Arrival time of vehicle v ∈ V at i ∈ N .
servv,i ∈ [ai , bi ] Start of service by vehicle v ∈ V at i ∈ N .
depv,i ∈ T Departure time of vehicle v ∈ V at i ∈ N .
crewTimec,i ∈ T Arrival/departure time of crew c ∈ C at i ∈ N .
driveStartc ∈ T Start time of driving for crew c ∈ C.
driveEndc ∈ T End time of driving for crew c ∈ C.
driveDurc ∈ [0, T ] Driving duration of crew c ∈ C.


crewc,i,j ≤ vehv,i,j , ∀c ∈ C, (i, j) ∈ A, li = lj , (1)
v∈V
 
vehv,i,j = driverc,i,j , ∀(i, j) ∈ A, li = lj , (2)
v∈V c∈C

driverc,i,j ≤ crewc,i,j , ∀c ∈ C, (i, j) ∈ A, li = lj ,

(3)
arrv,i ≤ servv,i , ∀v ∈ V, i ∈ R, (4)
servv,i + ti ≤ depv,i , ∀v ∈ V, i ∈ R, (5)
arrv,i = servv,i = depv,i , ∀v ∈ V, i ∈ {s, e}, (6)
arrv,i ≤ crewTimec,i ≤ depv,i , ∀v ∈ V, c ∈ C, i ∈ R, (7)
crewTimec,s − depv,s ≤ M1 (2 − vehv,s,j − crewc,s,j ), ∀v ∈ V, c ∈ C, j : (s, j) ∈ A, (8)
arrv,e − crewTimec,e ≤ M2 (2 − vehv,i,e − crewc,i,e ), ∀v ∈ V, c ∈ C, i : (i, e) ∈ A, (9)
⎛ ⎞

driveStartc − crewTimec,i ≤ M3⎝1 − driverc,i,j⎠, ∀c ∈ C, i ∈ R ∪ {s}, (10)
j:(i,j)∈A
⎛ ⎞

crewTimec,i − driveEndc ≤ M4⎝1 − driverc,h,i⎠, ∀c ∈ C, i ∈ R ∪ {e}, (11)
h:(h,i)∈A

driveDurc = driveEndc − driveStartc , ∀c ∈ C. (12)

Fig. 2. Core constraints of the MIP formulation.

 Joint Vehicle and Crew Routing and Scheduling 659

Table 2. Instance data and the decision variables of the CP model.

Name Description

T ∈ {1, . . . , ∞} Time horizon.

T = {0, . . . , T } Set of time values.
V ∈ {1, . . . , ∞} Number of vehicles.
V = {1, . . . , V } Set of vehicles.
C = {1, . . . , |C|} Set of crews.
C0 = C ∪ {0} Set of crews, including a 0 value indicating no crew.
T ∈T Maximum driving duration of a crew.
R ∈ {1, . . . , ∞} Total number of requests.
R = {1, . . . , R} Set of all requests.
S = {R + 1, . . . , R + V } Set of vehicle start nodes.
E = {R + V + 1, . . . , R + 2V } Set of vehicle end nodes.
N =R∪S∪E Set of all nodes.
N0 = N ∪ {0} Set of all nodes, including a dummy node 0.
s(v) = R + v Start node of vehicle v ∈ V.
e(v) = R + V + v End node of vehicle v ∈ V.
L Set of locations, including one depot location.
l(i) ∈ L Location of i ∈ N .
d(i, j) ∈ T Distance and travel time from i ∈ N to j ∈ N .
a(i) ∈ T Earliest start of service at i ∈ N .
b(i) ∈ T Latest start of service at i ∈ N .
t(i) ∈ T Service duration of i ∈ N .
w1 Cost of using one vehicle.
w2 Cost of using one crew.
w3 Cost of one unit of vehicle distance.
w4 Cost of one unit of crew distance.

vehUsed(v) ∈ {0, 1} Indicates if vehicle v ∈ V is used.

succ(i) ∈ N Successor node of i ∈ N .
veh(i) ∈ V Vehicle that visits i ∈ N .
arr(i) ∈ T Arrival time at i ∈ N .
serv(i) ∈ {a(i), . . . , b(i)} Start of service at i ∈ N .
dep(i) ∈ T Departure time at i ∈ N .
crewUsed(c) ∈ {0, 1} Indicates if crew c ∈ C is used.
crewDist(c) ∈ {0, . . . , ∞} Distance traveled by crew c ∈ C.
crewSucc(c, i) ∈ N0 Successor of i ∈ N0 for crew c ∈ C0 .
crewTime(i) ∈ T Arrival/departure time of every crew at i ∈ N .
driveStart(c) ∈ T Start time of driving for crew c ∈ C0 .
driveEnd(c) ∈ T End time of driving for crew c ∈ C0 .
driveDur(c) ∈ {0, . . . , T } Driving duration of crew c ∈ C.
driver(i) ∈ C0 Driver of vehicle veh(i) from i ∈ R ∪ S to succ(i),
with 0 indicating no driver.

The crew component, depicted in Fig. 4, overlays the vehicle component to

upgrade the VRP into a JVCRSP. Crew successor variables crewSucc(c, ·) model
a path for every crew c from any start node to any end node via a sequence of
arcs that cover the vehicle movements. To allow crews to start and end on any
vehicle, the model includes a dummy node 0; crews start and end at this dummy
node, whose successor must be a start node and whose predecessor must be
an end node. Equations (23) to (27) formalize this idea. The modeling uses the
Subcircuit constraint [14] to enforce subtour elimination, since a crew does not
visit all nodes. The Subcircuit constraint differs from the Circuit constraint
seen in the vehicle component by allowing some nodes to be excluded from the
Hamiltonian cycle. The successor of an excluded node is the node itself.
Every node i has an associated driver(i) variable that denotes the driver of
vehicle veh(i) when it moves from i to its successor succ(i) at a different location,
660 E. Lam et al.

succ(e(v)) = s(v + 1), ∀v ∈ {1, . . . , V − 1}, (13)

succ(e(V )) = s(1), (14)
Circuit(succ(·)), (15)
veh(s(v)) = veh(e(v)) = v, ∀v ∈ V, (16)
veh(succ(i)) = veh(i), ∀i ∈ R ∪ S, (17)
vehUsed(v) ↔ succ(s(v)) = e(v), ∀v ∈ V, (18)
arr(i) ≤ serv(i), ∀i ∈ R, (19)
serv(i) + t(i) ≤ dep(i), ∀i ∈ R, (20)
arr(i) = serv(i) = dep(i), ∀i ∈ S ∪ E, (21)
dep(i) + d(i, succ(i)) = arr(succ(i)), ∀i ∈ R ∪ S. (22)

Fig. 3. Key constraints of the vehicle component of the CP model.

crewSucc(c, 0) ∈ S ∪ {0}, ∀c ∈ C, (23)

crewSucc(c, i) ∈ R ∪ {i}, ∀c ∈ C, i ∈ S, (24)
crewSucc(c, i) ∈ R ∪ E, ∀c ∈ C, i ∈ R, (25)
crewSucc(c, i) ∈ {0, i}, ∀c ∈ C, i ∈ E, (26)
Subcircuit(crewSucc(c, ·)), ∀c ∈ C, (27)
l(succ(i)) = l(i) ↔ driver(i) = 0, ∀i ∈ R ∪ S, (28)
crewSucc(driver(i), i) = succ(i), ∀i ∈ R ∪ S, (29)
 
crewUsed(c) ↔ crewSucc(c, i) = i ↔ driver(i) = c, ∀c ∈ C, (30)
i∈N0 i∈R∪S

crewUsed(c) ≥ crewUsed(c + 1), ∀c ∈ {1, . . . , C − 1}, (31)

l(crewSucc(c, i)) = l(i) ∨ crewSucc(c, i) = succ(i), ∀c ∈ C, i ∈ R ∪ S, (32)
CrewShortcut(succ(·), crewSucc(·, ·), arr(·), dep(·), C, R, S, E, l(·)), (33)
arr(i) ≤ crewTime(i) ≤ dep(i), ∀i ∈ N , (34)
crewTime(i) ≤ crewTime(crewSucc(c, i)), ∀c ∈ C, i ∈ R ∪ S, (35)
CrewBound(crewDist(c), crewSucc(c, ·), crewTime(·), R, S, E, d(·, ·)), ∀c ∈ C, (36)
driveStart(driver(i)) ≤ dep(i), ∀i ∈ R ∪ S, (37)
driveEnd(driver(i)) ≥ arr(succ(i)), ∀i ∈ R ∪ S, (38)
driveDur(c) = driveEnd(c) − driveStart(c), ∀c ∈ C. (39)

Fig. 4. Crew component of the CP model.

 Joint Vehicle and Crew Routing and Scheduling 661

Fig. 5. A partial solution obtained at some stage during search. A crew is known to
traverse the arcs (1, 2) and (3, 4), but it is not yet known whether the start node S
connects (directly or via a path) to node 1 or node 3, and whether node 4 or node 2
connects to the end node. It is also not known whether such a path exists.

and takes the value 0 when the successor is at the same location, indicating that
no driver is necessary (Eq. (28)). Equation (29) requires drivers to move with
their vehicles. Equation (30) determines whether a crew is used. Equation (31)
is a symmetry-breaking constraint. Equation (32) is a space synchronization con-
straint that requires crews to move to a node at their current location, or move
with a vehicle (to a different location). The CrewShortcut global constraint
in Eq. (33) enforces the crew shortcut property, and is detailed in Section 5.2.
Equation (34) synchronizes time between vehicles and crews. The time window
in this constraint allows a crew to switch vehicle by moving to another node
at the same location, provided that this vehicle departs later than this crew.
Equation (35) forces crews to only move forward in time. Equation (36) is an
optimization constraint, described in Section 5.1, that bounds crew distances and
checks whether crews can return to the depot. Equations (37) to (39) determine
the driving time of each crew.
The objective function (Eq. (40)) minimizes a weighted sum of the vehicle
and crew counts, and the total vehicle and crew travel distances.
 
min w1 vehUsed(v) + w2 crewUsed(c)+
v∈V c∈C
 
w3 d(i, succ(i)) + w4 crewDist(c). (40)
i∈R∪S c∈C

5.1 Feasibility and Bounding of Crew Routes

Since crew routing is superimposed on a vehicle routing problem, it is important
to determine during the search whether each crew has a feasible route and to
compute a lower bound on the crew distance. Consider Fig. 5, which illustrates
a partial solution. Here, the crew has been allocated to travel on some arcs but
it is not known whether there exists a path from a start node to an end node
through these arcs, and how long this path is, if it exists.
To detect infeasibility and to compute lower bounds to crew distances, an
optimization constraint can be used [13,10,12,11]. It needs to find a shortest
662 E. Lam et al.

 
min d(i, j) · xi,j , (41)
i∈R∪S j∈cSucc(i)
 
xi,j = 1, (42)
i∈S j:(i,j)∈B
 
xh,i = 1, (43)
i∈E h:(h,i)∈B
 
xh,i = xi,j , ∀i ∈ R, (44)
h:(h,i)∈B j:(i,j)∈B

xi,j ≤ 1, ∀i ∈ R ∪ S, (45)
j:(i,j)∈B

ti + d(i, j) − tj ≤ M (1 − xi,j ), (i, j) ∈ B, (46)

xi,j ∈ [0, 1], (i, j) ∈ B, (47)
ti ∈ [min(cTime(i)), max(cTime(i))], i ∈ N . (48)

Fig. 6. The linear programming relaxation for optimization constraint CrewBound.

path through specified nodes, which is NP-hard [24,20,9,32]. Instead, the

CrewBound constraint solves a linear relaxation, which is presented in Fig. 6.
The set B represents the current set of arcs that can be traversed by the crew.
It is defined as

B = {(i, j) : i ∈ R ∪ S, j ∈ cSucc(i), i = j},

where the set cSucc(i) represents the domain of crewSucc(c, i) for every node
i ∈ R ∪ S and the appropriate crew c. The set cTime(i) represents the domain
of crewTime(i) for every node i ∈ N .
The xi,j variables indicate whether the crew traverses the arc (i, j) ∈ B, and
the ti variables represent the arrival time to node i.
The objective function minimizes the total distance (Eq. (41)). (Equa-
tions (42) to (44) are the flow constraints, which ensure the existence of a path
from a start node to an end node. Equation (45) is a redundant constraint that
strengthens the linear program. Equation (46) is the time-based subtour elimi-
nation constraint, where M is a big-M constant. Equations (47) and (48) restrict
the domains of the xi,j and ti variables respectively.

5.2 Symmetry-Breaking within Locations

The constraint program also uses the global constraint CrewShortcut to

remove symmetries in the crew routes. The intuition is best explained using
Fig. 1. Recall that, in this example, one crew moves from node 3 to 8 and the
other crew moves from node 4 to 7. Both crews thus shortcut the intermediate
nodes, which are only visited by vehicles. The global constraint CrewShortcut
 Joint Vehicle and Crew Routing and Scheduling 663

ensures that crews never visit intermediate nodes, but move directly from incom-
ing to outgoing nodes at a location. This pruning eliminates symmetric solutions
in which crews visit various sequences of intermediate nodes within a location.
Let pred(i) ∈ N be the predecessor of a node i ∈ N . The constraint imple-
ments simple propagation rules, such as,

l(pred(i)) = l(i) = l(succ(i)) ↔ crewSucc(c, i) = i, ∀i ∈ R, (49)
c∈C

which ensures that crews never visit a node whose predecessor and successor are
at the same location as the node itself.

5.3 The Search Procedure

This section briefly describes the labeling procedure and the LNS procedure.

The Labeling Procedure. The labeling procedure first assigns vehicle routes, and
then crew routes.
To determine vehicle routes, the search procedure considers each vehicle in
turn and assigns the successor variables from the start node to the end node.
The value selection strategy first chooses nodes at the same location ordered by
earliest service time, and then considers nodes at other locations, also ordered
by earliest service time.
To assign crew routes, the labeling procedure begins by ordering all nodes
i ∈ R ∪ S that have an unlabeled driver(i) variable by earliest departure time.
The first of these nodes is assigned a driver, and then a path from this node to
any end node is constructed for this crew by labeling both the crew successor
and the driver variables. This process is repeated until all driver(·) variables are
labeled. The value selection heuristic for crew successor variables favors the node
that is visited next by the current vehicle of the crew, provided that the crew
can drive from this node; otherwise, the successor is chosen randomly.

The LNS Procedure. The LNS procedure randomly explores two neighborhoods:

– Request Neighborhood: This neighborhood relaxes a collection of pickup

and delivery requests for insertion, in addition to all crew routes;
– Crew Neighborhood: This neighborhood fixes all vehicle routes, and
relaxes a number of crew routes.

6 Experimental Results

This section compares the solution quality and features of the MIP and CP
models, and the two-stage procedure.
664 E. Lam et al.

Table 3. The number of locations and requests in each of the ten JVCRSP instances.

A B C D E F G H I J
|L| 7 7 7 11 11 12 11 11 5 6
|R| 12 16 20 40 46 46 40 80 54 96

The Instances. The instances are generated to capture the essence of applications
in humanitarian and military logistics. These problems typically feature fewer
locations than traditional vehicle routing applications, but comprise multiple
requests at each location. Both the locations and the time windows are generated
randomly. Table 3 describes the size of the instances. The maximum number of
vehicles and crews are not limited in the instances.

The Algorithms. As mentioned previously, the two-stage method uses a con-

straint programming model for its first step. The model is solved using a large
neighborhood search with a time limit of one minute, and the first step takes
the best VRP solution out of ten runs. The crew pairing problem is solved opti-
mally using the output of the first stage. All instances are solved to optimality in
under 0.02 seconds. Hence the two-stage method requires a total of ten minutes.
Obviously, the VRP runs can be performed in parallel.
The MIP model is solved using Gurobi 6.0.0 on an eight-core CPU for twelve
hours with default parameters. The linear programming relaxation is rather weak
and the dual bounds are not informative, which is why they are not reported here.
An alternative MIP model that folds Eqs. (8) and (9) into Eq. (7) by modifying
the network is also tested, but is inferior to the one presented.
The CP model is implemented in Objective-CP [31], and initialized with
the same VRP solution as in the two-stage approach. This allows us to assess
the benefits of the joint optimization reliably, since the second stage and the
CP model are seeded with the same vehicle routes. The LNS procedure over
the constraint model is then performed ten times for five minutes each. The
model essentially requires a total of sixty minutes, which includes ten minutes
to produce the initial VRP solution. Of course, all these computations can be
executed in parallel in six minutes of wall clock time.

Solution Quality. Tables 4 to 6 depict the experimental results for various vehicle
and crew costs. The tables only report the CPU time for the MIP model, since
the two-stage and CP models have been given short execution times, as explained
above. The experimental results highlight a few important findings:
1. Jointly optimizing vehicle and crew routing and scheduling achieves signifi-
cant benefits. On the above instances, the average benefit of the CP model
over the two-stage approach is 12%. When crew costs are higher, the benefits
can exceed 45%.
2. The benefits obviously increase when the crew costs increase. However, even
when vehicle costs are five times as large as crew costs, the benefits are still
substantial and exceed 5% in more than half of the instances.
 Joint Vehicle and Crew Routing and Scheduling 665

Table 4. Solutions found by the three models with equal vehicle and crew costs (w1 =
w2 = 1000, w3 = w4 = 1). The objective value columns separate vehicle and crew costs
in parenthesis. Time (in seconds) is provided for the MIP model. The last column lists
the percentage improvement of the CP model compared against the two-stage method.

Two-Stage MIP CP

Objective Value Objective Value Time Objective Value %

A 8328 (2086 + 6242) 8318 (2086 + 6232) 5 8318 (2086 + 6232) -0.1%
B 7198 (2065 + 5133) 6154 (2063 + 4091) 27132 6167 (2068 + 4099) -16.7%
C 8216 (2079 + 6137) 6192 (2088 + 4104) 37254 6200 (2079 + 4121) -32.5%
D 14879 (3298 + 11581) - - 13855 (3302 + 10553) -7.4%
E 18105 (2275 + 15830) - - 17181 (2278 + 14903) -5.4%
F 26334 (4406 + 21928) - - 24570 (4412 + 20158) -7.2%
G 14110 (3351 + 10759) - - 13074 (3355 + 9719) -7.9%
H 25617 (6585 + 19032) - - 24975 (6587 + 18388) -2.6%
I 8894 (2316 + 6578) - - 7795 (2309 + 5486) -14.1%
J 16932 (3710 + 13222) - - 13877 (3710 + 10167) -22.0%

Table 5. Best solutions from the three models with higher crew costs (w1 = 1000,
w2 = 5000, w3 = w4 = 1).

Two-Stage MIP CP

Objective Value Objective Value Time Objective Value %

A 32328 (2086 + 30242) 32318 (2086 + 30232) 5 32318 (2086 + 30232) 0.0%
B 27198 (2065 + 25133) 22154 (2056 + 20098) 8352 22167 (2068 + 20099) -22.7%
C 32216 (2079 + 30137) 28198 (3086 + 25112) 4869 22200 (2079 + 20121) -45.1%
D 58879 (3298 + 55581) - - 53905 (3317 + 50588) -9.2%
E 78105 (2275 + 75830) - - 73175 (2280 + 70895) -6.7%
F 110334 (4406 + 105928) - - 100502 (4413 + 96089) -9.8%
G 54110 (3351 + 50759) - - 43996 (3353 + 40643) -23.0%
H 97617 (6585 + 91032) - - 92928 (6585 + 86343) -5.0%
I 32894 (2316 + 30578) - - 27817 (2326 + 25491) -18.3%
J 64932 (3710 + 61222) - - 49930 (3710 + 46220) -30.0%

Table 6. Best solutions from the three models with higher vehicle costs (w1 = 5000,
w2 = 1000, w3 = w4 = 1).

Two-Stage MIP CP
Objective Value Objective Value Time Objective Value %

A 16328 (10086 + 6242) 16318 (10086 + 6232) 7 16318 (10086 + 6232) -0.1%
B 15198 (10065 + 5133) 14154 (10056 + 4098) 2327 14167 (10068 + 4099) -7.3%
C 16216 (10079 + 6137) 14200 (10091 + 4109) 32135 14200 (10079 + 4121) -14.2%
D 26879 (15298 + 11581) - - 24946 (15303 + 9643) -7.7%
E 26105 (10275 + 15830) - - 25172 (10278 + 14894) -3.7%
F 42334 (20406 + 21928) - - 40597 (20412 + 20185) -4.3%
G 26110 (15351 + 10759) - - 24008 (15353 + 8655) -8.8%
H 49617 (30585 + 19032) - - 48911 (30585 + 18326) -1.4%
I 16894 (10316 + 6578) - - 15800 (10326 + 5474) -6.9%
J 28932 (15710 + 13222) - - 25930 (15710 + 10220) -11.6%
666 E. Lam et al.

3. On the smaller instances, the MIP marginally outperforms the CP model

when run over a significantly longer period of time.
4. The MIP model contains numerous big-M synchronization constraints that
weaken its linear relaxation, making it difficult to solve. It can only prove opti-
mality on the trivial instance A, and it also fails to scale to large instances,
contrary to the CP model, which quickly finds high-quality solutions.

The Impact of Vehicle Interchanges. One of the main advantages of jointly opti-
mizing vehicle and crew routing and scheduling is the additional flexibility for
vehicle interchanges. For example, instance C in Tables 4 to 6 highlight some of
the benefits of this flexibility. The CP model is able to delay vehicles to favor
interchanges, thereby decreasing crew costs significantly.
Table 7 quantifies vehicle interchanges for the instances with equal vehicle
and crew costs. It reports the total number of crews required in the solution, the
number of crews participating in vehicle interchanges, the total number of vehicle
interchanges, and the average number of interchanges per crew that switches
vehicle. These results indicate that the constraint programming model allows
more vehicle interchanges on average. For the larger instances, each crew travels
on almost three vehicles on average. The total number of vehicle interchanges
is roughly the same for the CP model and the two-stage approach, but the CP
model produces solutions with fewer crews.

The Impact of the Joint Vehicle and Crew Routing and Scheduling. It is also
interesting to notice in Tables (4) to (6) that the CP model often prefers vehicle
routes whose costs are higher than those of the two-stage approach. This is the
case despite the fact that the CP search is initialized with the first-stage VRP
solution. This indicates that the CP model trades off crews costs for vehicle
costs. Running a two-stage version of the CP model, in which the vehicle routes

Table 7. Statistics on vehicle interchange in the solutions to the runs with equal vehicle
and crew costs. For both the two-stage and the CP models, this table provides the
number of crews used in the solution, the number of crews that participates in vehicle
interchange, the total number of interchanges, and the average number of interchanges
for each crew that undertakes interchange.

Two-Stage CP

Crews Intchng Vehicle Average Crews Intchng Vehicle Average

Used Crews Intchngs Intchngs Used Crews Intchngs Intchngs

A 6 2 2 1.0 6 3 3 1.0
B 5 2 2 1.0 4 3 3 1.0
C 6 3 3 1.0 4 4 4 1.0
D 11 8 12 1.5 10 7 7 1.0
E 15 9 14 1.6 14 12 21 1.8
F 21 20 35 1.8 19 19 36 1.9
G 10 8 9 1.1 9 8 11 1.4
H 18 17 28 1.6 17 16 29 1.8
I 6 5 5 1.0 5 3 5 1.7
J 12 11 15 1.4 9 8 15 1.9
 Joint Vehicle and Crew Routing and Scheduling 667

(but not the vehicle schedules) are fixed, also confirms these benefits. The CP
model significantly outperforms this two-stage CP approach on a number of
instances. These results indicate that both the flexibility in interchanges and the
joint optimization are key benefits of the CP model.

7 Related Work
JVCRSPs have not attracted very much interest in the literature at this point,
probably due to their inherent complexity [5]. This contradicts with the numer-
ous studies of VRPs and related problems (e.g., [30,23,2]). This section reviews
the most relevant work.
A JVCRSP is considered in [22], in which vehicles transport teams to service
customers, and are able to move without any team on board. The problem
features three types of tasks that must be serviced in order, and all customers
have one task of each type. Each team can only service one compatible type
of task. The MIP formulation has variables indexed by five dimensions, and is
intractable. The paper also develops a simple iterative local search algorithm
embedded within a particle swarm metaheuristic. This approach is very specific
and cannot accommodate side constraints easily.
A JVCRSP modeling a mail distribution application is solved using a two-
stage heuristic column generation approach in [19]. This model features multiple
depots at which vehicle routes begin and end. In the first stage, trips that begin
and end at these depots are computed. The second stage takes a scheduling
approach and assigns vehicles and crews to the trips. Vehicle interchange can only
occur at the depots at the start or end of a trip. In addition, the model features
a 24-hour cyclic time period, and variables indexed by discretized time blocks.
Since the solution method is based on two stages, it cannot jointly optimize
vehicles and crews. Observe also that the CP formulation in our paper does not
require time-indexed variables, and allows for vehicle interchange at all locations.
Another JVCRSP considers European legislation and includes relay stations
where drivers are allowed to rest and interchange vehicles [8]. Upon reaching a
relay station, the vehicle must wait a fixed amount of time. Vehicle interchange
can only occur if other drivers arrive at this relay station during this time interval.
The problem also provides a shuttle service, separate to the fleet of vehicles,
that can be used to move drivers between relay stations. The problem is solved
using a two-stage LNS method. In the first stage, a VRP is solved and the
resulting routes form the customers of a VRP in the second stage, in which the
crews perform the role of vehicles. Observe that this approach also cannot jointly
optimize vehicle and crew routing. The model features several fixed parameters,
such as the duration during which a vehicle waits at a relay station, and a
search procedure that only explores a limited number of nearby relay stations.
Both these restrictions reduce the search space, but negatively impact vehicle
interchange, leading the authors to conclude that allowing for vehicle interchange
does not significantly improve the objective value. Our model proves the contrary
since it does not fix the duration that a crew waits, and can explore all locations
for vehicle interchange.
668 E. Lam et al.

JVCRSPs belong in the class of problems known as the VRPs with multiple
synchronization constraints [5]. Synchronization is a feature present in certain
VRPs, in which decisions about one object (e.g., vehicle, route, request) imply
actions that may or must be taken on other objects. A complex VRP with
synchronization is the VRP with Trailers and Transshipments (VRPTT) [7,4]. It
features two vehicle classes: lorries which can move independently, and trailers
which must be towed by an accompanying lorry. All lorries begin at a single
depot with or without a trailer. A lorry can detach its trailer at transshipment
locations in order to visit customers who are unable to accommodate a trailer
(e.g., due to size). Lorries can transfer load into and/or attach with any trailer
at any transshipment location. A lorry that has detached its trailer can also
return to the depot without reattaching a trailer, leaving its trailer behind at a
transshipment location to be collected by another lorry at a future time. Several
sophisticated MIP formulations are presented, which are solved using branch-
and-cut on instances with up to eight customers, eight transshipment locations
and eight vehicles. The VRPTT can be reformulated into a JVCRSP by casting
crews as lorries with zero capacity, and vehicles as trailers [6].
The VRPTT is most closely related to our JVCRSP model, since lorry routes
and trailer routes are jointly computed, and the search space is not artificially
limited. However, one major difference is that the VRPTT incorporates load
synchronization, which the JVCRSP does not consider.
Finally, observe that vehicle and crew scheduling problems (VCSPs) are thor-
oughly studied (e.g., [18,27,17,15,16,3,25,26]). VCSPs aim to assign vehicles and
crews to a predetermined set of trips, with each trip consisting of a fixed route,
and usually fixed arrival and departure times. Trips in VCSPs correspond to
parts of a route in JVCRSPs, which are not available a priori, but instead, must
be computed during search, thereby increasing the computational challenges.

8 Conclusion

Motivated by applications arising in humanitarian and military logistics, this

paper studied a Joint Vehicle and Crew Routing and Scheduling Problem, in
which crews are able to interchange vehicles, resulting in interdependent vehicle
routes and crew routes. The paper proposed a constraint programming model
that overlays crew routing constraints over a standard vehicle routing problem.
The model uses a novel optimization constraint to detect infeasibility and to
bound crew objectives, and a symmetry-breaking global constraint. The model
is compared to a sequential method and a mixed integer programming model.
Experimental results demonstrate significant benefits of the constraint program-
ming model, which reduces costs by 12% on average compared to the two-stage
method, and scales significantly better than the mixed integer programming
model. The benefits of the constraint programming model are influenced by
both the additional flexibility in vehicle interchanges and in trading crew costs
for vehicle costs. Future work will investigate more sophisticated search tech-
niques and global constraints to scale to larger instances.
 Joint Vehicle and Crew Routing and Scheduling 669

References
1. Bent, R., Van Hentenryck, P.: A two-stage hybrid local search for the vehicle rout-
ing problem with time windows. Transportation Science 38(4), 515–530 (2004)
2. Berbeglia, G., Cordeau, J.F., Gribkovskaia, I., Laporte, G.: Static pickup and deliv-
ery problems: a classification scheme and survey. TOP 15(1), 1–31 (2007)
3. Cordeau, J.F., Stojković, G., Soumis, F., Desrosiers, J.: Benders decomposition for
simultaneous aircraft routing and crew scheduling. Transportation Science 35(4),
375–388 (2001)
4. Drexl, M.: On some generalized routing problems. Ph.D. thesis, RWTH Aachen
University, Aachen (2007)
5. Drexl, M.: Synchronization in vehicle routing–a survey of VRPs with multiple
synchronization constraints. Transportation Science 46(3), 297–316 (2012)
6. Drexl, M.: Applications of the vehicle routing problem with trailers and transship-
ments. European Journal of Operational Research 227(2), 275–283 (2013)
7. Drexl, M.: Branch-and-cut algorithms for the vehicle routing problem with trailers
and transshipments. Networks 63(1), 119–133 (2014)
8. Drexl, M., Rieck, J., Sigl, T., Press, B.: Simultaneous vehicle and crew routing and
scheduling for partial- and full-load long-distance road transport. BuR - Business
Research 6(2), 242–264 (2013)
9. Dreyfus, S.E.: An appraisal of some shortest-path algorithms. Operations Research
17(3), 395–412 (1969)
10. Focacci, F., Lodi, A., Milano, M.: Embedding relaxations in global constraints for
solving TSP and TSPTW. Annals of Mathematics and Artificial Intelligence 34(4),
291–311 (2002)
11. Focacci, F., Lodi, A., Milano, M.: Optimization-oriented global constraints. Con-
straints 7(3–4), 351–365 (2002)
12. Focacci, F., Lodi, A., Milano, M.: Exploiting relaxations in CP. In: Milano, M.
(ed.) Constraint and Integer Programming, Operations Research/Computer Sci-
ence Interfaces Series, vol. 27, pp. 137–167. Springer, US (2004)
13. Focacci, F., Lodi, A., Milano, M., Vigo, D.: Solving TSP through the integration of
OR and CP techniques. Electronic Notes in Discrete Mathematics 1, 13–25 (1999)
14. Francis, K.G., Stuckey, P.J.: Explaining circuit propagation. Constraints 19(1),
1–29 (2014)
15. Freling, R., Huisman, D., Wagelmans, A.: Applying an integrated approach to vehi-
cle and crew scheduling in practice. In: Voß, S., Daduna, J. (eds.) Computer-Aided
Scheduling of Public Transport. Lecture Notes in Economics and Mathematical
Systems, vol. 505, pp. 73–90. Springer, Berlin Heidelberg (2001)
16. Freling, R., Huisman, D., Wagelmans, A.: Models and algorithms for integration
of vehicle and crew scheduling. Journal of Scheduling 6(1), 63–85 (2003)
17. Freling, R., Wagelmans, A., Paixão, J.: An overview of models and techniques for
integrating vehicle and crew scheduling. In: Wilson, N. (ed.) Computer-Aided Tran-
sit Scheduling. Lecture Notes in Economics and Mathematical Systems, vol. 471,
pp. 441–460. Springer, Berlin Heidelberg (1999)
18. Haase, K., Desaulniers, G., Desrosiers, J.: Simultaneous vehicle and crew scheduling
in urban mass transit systems. Transportation Science 35(3), 286–303 (2001)
19. Hollis, B., Forbes, M., Douglas, B.: Vehicle routing and crew scheduling for
metropolitan mail distribution at Australia Post. European Journal of Operational
Research 173(1), 133–150 (2006)
670 E. Lam et al.

20. Ibaraki, T.: Algorithms for obtaining shortest paths visiting specified nodes. SIAM
Review 15(2), 309–317 (1973)
21. Kilby, P., Prosser, P., Shaw, P.: A comparison of traditional and constraint-based
heuristic methods on vehicle routing problems with side constraints. Constraints
5(4), 389–414 (2000)
22. Kim, B.I., Koo, J., Park, J.: The combined manpower-vehicle routing problem for
multi-staged services. Expert Systems with Applications 37(12), 8424–8431 (2010)
23. Laporte, G.: What you should know about the vehicle routing problem. Naval
Research Logistics (NRL) 54(8), 811–819 (2007)
24. Laporte, G., Mercure, H., Norbert, Y.: Optimal tour planning with specified nodes.
RAIRO - Operations Research - Recherche Opérationnelle 18(3), 203–210 (1984)
25. Mercier, A., Cordeau, J.F., Soumis, F.: A computational study of benders decompo-
sition for the integrated aircraft routing and crew scheduling problem. Computers
& Operations Research 32(6), 1451–1476 (2005)
26. Mercier, A., Soumis, F.: An integrated aircraft routing, crew scheduling and flight
retiming model. Computers & Operations Research 34(8), 2251–2265 (2007)
27. Mesquita, M., Paias, A.: Set partitioning/covering-based approaches for the inte-
grated vehicle and crew scheduling problem. Computers & Operations Research
35(5), 1562–1575 (2008), part Special Issue: Algorithms and Computational Meth-
ods in Feasibility and Infeasibility
28. Rousseau, L.M., Gendreau, M., Pesant, G.: Using constraint-based operators to
solve the vehicle routing problem with time windows. Journal of Heuristics 8(1),
43–58 (2002)
29. Shaw, P.: Using Constraint Programming and Local Search Methods to Solve
Vehicle Routing Problems. In: Maher, M.J., Puget, J.-F. (eds.) CP 1998. LNCS,
vol. 1520, pp. 417–431. Springer, Heidelberg (1998)
30. Toth, P., Vigo, D.: The Vehicle Routing Problem. Society for Industrial and Applied
Mathematics (2002)
31. Van Hentenryck, P., Michel, L.: The Objective-CP Optimization System. In:
Schulte, C. (ed.) CP 2013. LNCS, vol. 8124, pp. 8–29. Springer, Heidelberg (2013)
32. Volgenant, T., Jonker, R.: On some generalizations of the travelling-salesman prob-
lem. The Journal of the Operational Research Society 38(11), 1073–1079 (1987)
33. Yu, G.: Operations Research in the Airline Industry. International Series in Oper-
ations Research & Management Science: 9, Springer, US (1998)
 Constructing Sailing Match Race Schedules:
Round-Robin Pairing Lists

Craig Macdonald, Ciaran McCreesh, Alice Miller, and Patrick Prosser(B)

University of Glasgow, Glasgow, Scotland

patrick.prosser@glasgow.ac.uk

Abstract. We present a constraint programming solution to the prob-

lem of generating round-robin schedules for sailing match races. Our
schedules satisfy the criteria published by the International Sailing Fed-
eration (ISAF) governing body for match race pairing lists. We show
that some published ISAF schedules are in fact illegal, and present cor-
responding legal instances and schedules that have not previously been
published. Our schedules can be downloaded as blanks, then populated
with actual competitors and used in match racing competitions.

1 Introduction
This work describes a round-robin competition format that arises from sailing,
known as match racing. Competitors, i.e. skippers, compete against each other in
a series of matches taking place in rounds, known as flights. A match is composed
of two skippers, with each skipper in a boat on their own. Skippers in a match set
off together, one on the port side, the other starboard, and first home wins that
match. Criteria published by the International Sailing Federation (ISAF) dictate
what makes a legal match race schedule and what should be optimized. This is
a rich set of criteria, 13 in all, and as far as we know, all published schedules
have been produced by hand. This is a daunting task. A close inspection of
these schedules reveals that most are illegal, violating many of the match race
scheduling criteria, many schedules are missing, and some that are legal are far
from the ISAF definition of optimality.
This paper describes the scheduling problem. We present a constraint pro-
gramming solution and document how this model was constructed incrementally.
We believe this illustrates how adaptable constraint programming is: a robust
model can be developed incrementally to address all of the required criteria.
Our schedules are then constructed in a sequence of stages, not dissimilar to the
hybrid and decomposition based approaches by Lombardi and Milano [3]. Some
ISAF published schedules are presented along with some of our own, some of
these being new. These schedules are essentially blank forms, made available to
event organizers, which can be filled in with names of actual competitors, i.e.
these schedules are reusable.
The paper is organized as follows: we introduce the 13 match racing criteria,
then explain the constraint models and four stages of schedule construction, and
then schedules are presented. We finish with discussion and a conclusion.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 671–686, 2015.
DOI: 10.1007/978-3-319-23219-5 46
672 C. Macdonald et al.

2 Problem Definition: Round-Robin Pairing Lists

The guidelines for running match racing events [2] places a number of criteria
on how the competing skippers are scheduled into flights and matches, known as
pairing lists. Several of these criteria are applicable when the number of skippers
exceeds the number of available boats, in which case skippers have to change in
and out of boats, and are allowed time in the schedule to check and fine-tune
the boat they change into. Typically, matches set off at 5 minute intervals with
each match taking about 20 minutes. Therefore, if we have 10 skippers and 10
boats we have 45 matches, 9 flights each of 5 matches, and if each flight takes
about 50 minutes, eight or nine flights a day is a reasonable achievement for most
events [2]. Consequently, the number of boats and skippers involved is typically
small.
The criteria for match racing schedules (pairing lists) are given in ISAF
“International Umpires’ and Match Racing Manual” [2], section M.2 “Recom-
mended Criteria for Round Robin Pairing Lists”, and are detailed below.

Principal Criteria in Order of Priority:

1. Each skipper sails against each other skipper once.

2*. When skippers have an even number of matches, they have the same
number of port and starboard assignments.
3*. When skippers have an odd number of matches, the first half of the
skippers will have one more starboard assignment.
4. No skipper in the last match of a flight should be in the first match of
the next flight.
5. No skipper should have more than two consecutive port or starboard
assignments.
6. Each skipper should be assigned to match 1, match 2, etc. in a flight as
equally as possible.
7. In flights with five or more matches, no skipper should be in the next-
to-last match in a flight and then in the first match of the next flight.
8. If possible, a skipper should be starboard when meeting the nearest
lowest ranked skipper (i.e. #1 will be starboard against #2, #3 will be
starboard against #4).
9. Close-ranked skippers meet in the last flight.
10. Minimize the number of boat changes.
11. Skippers in the last match of a flight do not change boats.
12. Skippers in new boats do not sail in the first match of the next flight.
13. Skippers have a reasonable sequence of matches and blanks.

Note that criteria 10, 11 and 12 only apply when there are fewer boats than
skippers; 11 and 12 override 6 when changes are required, and 13 applies
when there are more boats than skippers.
 Constructing Sailing Match Race Schedules: Round-Robin Pairing Lists 673

We have rephrased criteria 2 and 3 (denoted *) to clarify an error in the man-

ual [2]. Note that the order of matches within a flight is significant and is con-
strained, as is the order of flights within a schedule and the position of skippers
within a match. This permits fair schedules that provide sufficient changeover
time for skippers changing boats, etc.
Criterion 6 allows us to measure the balance of a schedule. Perfect balance
would be when a skipper has as many matches in first position in flights as in
second position in flights, and so on. For example, if we had 9 skippers, each
skipper would have 8 matches and could be in the first, second, third or fourth
match of a flight. In perfect balance, each of the 8 skippers would appear twice
in each position. Balance is one of our optimization criteria, and we do this
unconditionally, even when there are more skippers than boats.
Criterion 13 uses the term blanks where conventionally we use the term bye.
That is, a blank is a bye and a bye is a flight where a skipper does not compete
(and is ashore).
Criterion 12 discusses boat changes: if in flight i this skipper is not in a match
(i.e. it is a bye for this skipper) but is in a match in flight i + 1 then he has to
get into a boat, rearrange all his kit and set the boat to his preference before
racing commences, and this is a boat change.
Next, we note that skippers can be ranked prior to an event, based on their
performance in previous events1 , which is used to seed the skippers. The ordering
of skippers within a match signifies their starting position (port or starboard),
as skippers allocated the starboard starting position gain a small competitive
advantage in that match—criterion 8 accomplishes the seedings.
Criterion 10 is ambiguous: this could be interpreted on a schedule basis: i.e.
to minimize the overall number of changes in the entire pairing list schedule; or
alternatively as well as a fair, but minimal number of changes across all skippers.
In our work, we minimize the overall number of changes.
There are some conflicts inherent in the criteria. Take criteria 4 and 11,
assume we have 4 boats, and in flight i the last match is the pair of skippers
(x, y). Criterion 11 dictates that skippers x and y must appear in the next
flight, i + 1, and criterion 4 that neither can be first in the next flight. This
forces skippers x and y to compete in flight i + 1 as the last match, violating
criterion 1. Therefore, although not explicitly stated, it is not possible to satisfy
every criteria with fewer than 6 boats.

3 The Constraint Models

Our approach produces schedules in four stages. The first stage produces a
schedule that respects criteria 1, 4, 11 and 12, and minimizes criteria 10 (boat
changes). The second stage constructs a schedule that respects criteria 1, 4, 11,
12 (again), has criterion 10 as a hard constraint generated from first stage, and
minimizes criterion 6 (balance). This results in a multi-objective schedule that

1
http://www.sailing.org/rankings/match/
674 C. Macdonald et al.

Flight Matches
Table 1. A round-robin schedule with 7 skippers,
0 (0,1) (2,3) (4,5) 6 boats and 7 flights. Skippers are numbered 0 to
1 (0,2) (4,6) (1,5) 6. Note that the order of skippers within a flight is
2 (2,6) (0,5) (1,3) significant, as is position within a match (port or
3 (5,6) (0,3) (1,4) starboard). This schedule is the result of stages 1
4 (3,5) (1,6) (2,4) and 2, and has yet to be renumbered and oriented
5 (3,6) (0,4) (2,5) (stages 3 and 4).
6 (3,4) (1,2) (0,6)

attempts to minimize boat changes and balance. The third stage satisfies crite-
rion 9, and is a simple translation of the schedule produced in stage 2. The final
stage orients skippers within matches to address criteria 2, 3, 5 and 8.
We now describe the constraint models and processes used in each stage. In
the text below we assume there are n skippers and b boats, with m = b/2 matches
in a flight. In total there are t = n(n−1)/2 matches and f = n(n−1)/m flights.
We use the schedule in Table 1, for 7 skippers and 6 boats, to help illustrate the
constraint models.

3.1 Stage 1: Minimizing Boat Changes

Modeling skippers: The first thing we model, in the first half of Figure 1, is a
skipper. Each skipper σ has both a temporal view of their schedule (“who, if
anyone, am I racing in the match at time t?”), and a state view (“where am I
racing within flight f ?”). The temporal view is via the array timeSlot defined in
(V1). If variable timeSlot[i] = k and k ≥ 0 then this skipper is in a match with
skipper k at time i.
Variable , [i] (V2) gives the state of this skipper in flight i, corresponding to
time slots timeSlot[m · i] to timeSlot[m(i + 1) − 1]. The cases in (C1) state that
a skipper can be in a match in the first time slot in the flight, or in the middle
of the flight2 , or in the last time slot of the flight. Alternatively, if all time slots
in a flight are equal to −1 then the skipper is in a bye (i.e. not in a match in
this flight), and if all time slots in a flight are equal to −2 then the skipper has
finished all matches.
We must then ensure that each skipper σ is in a match with all other skip-
pers {0, ... , n − 1} \ {σ}. This is (C2), which is enforced by imposing a global
cardinality constraint [5] on the array timeSlot.

State transitions: The , variables are then used to impose match race criterion 4
(if last in flight i then not first in flight i + 1), criterion 11 (if last in flight i
then not in a bye in flight i + 1) and criterion 12 (if flight i is a bye then flight
i + 1 is not first). These criteria are imposed in (C3) by the deterministic finite
automaton (DFA) shown in Figure 2 using Pesant’s regular constraint [4]. The
2
i.e. not first and not last.
 Constructing Sailing Match Race Schedules: Round-Robin Pairing Lists 675

A copy of these variables and constraints is created for each skipper σ:

∀τ ∈ {0 ... t−1} : timeSlot[τ ] ∈ {−2 ... t−1} \ {σ} (V1)

∀i ∈ {0 ... f −1} : state[i] ∈ {First, Mid, Last, Bye, End} (V2)
∀i ∈ {0 ... f −1} : (C1)
state[i] = First ⇔ timeSlot[m · i] ≥ 0
state[i] = Mid ⇔ ∃j ∈ {m · i + 1 ... m · (i + 1) − 2} : timeSlot[j] ≥ 0
state[i] = Last ⇔ timeSlot[m · (i + 1) − 1] ≥ 0
state[i] = Bye ⇔ ∀j ∈ {m · i ... m · (i + 1) − 1} : timeSlot[j] = −1
state[i] = End ⇔ ∀j ∈ {m · i ... m · (i + 1) − 1} : timeSlot[j] = −2
eachOccursExactlyOnce(timeSlot, {0, ... , n−1} \ {σ}) (C2)
regular(state, Figure 2) (C3)
∀i ∈ {0 ... f −2} : change[i] ∈ {0, 1} (V3)
totalChanges ∈ N (V4)
∀i ∈ {0 ... f −2} : change[i] = 1 ⇔ state[i] = Bye ∧ state[i + 1] = Bye (C4)

totalChanges = change (C5)

Match and temporal perspectives:

∀i ∈ {0 ... n−2} : ∀j ∈ {i+1 ... n−1} :

match[i, j] ∈ {0 ... t−1} (V5)
match[j, i] ≡ match[i, j]
∀k ∈ {0 ... t−1} :
match[i, j] = k ⇔ σ[i].timeSlot[k] = j ∧ σ[j].timeSlot[k] = i (C6)
∀i ∈ {0 ... n−2} : ∀j ∈ {i+1 ... n−1} :
modMatch[i, j] ∈ {0 ... f −1} (V6)
modMatch[j, i] ≡ modMatch[i, j]
modMatch[i, j] = match[i, j]/m (C7)
∀i ∈ {0 ... n−1} : allDifferent(modMatch[i]) (C8)
∀τ ∈ {0 ... t−1} :
time[τ ] ∈ {(0, 1) ... (n−2, n−1)} (V7)
∀i ∈ {0 ... n−2} : ∀j ∈ {i+1 ... n−1} : time[τ ] = (i, j) ⇔ match[i, j] = τ (C9)

totalBoatChanges = σ.totalChanges (V8)
minimise(totalBoatChanges) (C10)

Fig. 1. Our constraint model, from a skipper perspective (top) and a match and tem-
poral perspective (below). The number of skippers is n, and m is the number of matches
in a flight. The number of flights is f =
n(n − 1)/m, and there are t = n(n − 1)/2
matches (and time slots) in total. We define N to include zero.
676 C. Macdonald et al.

Skipper 5:

4 1 0 6 3 2

0
1
2
3
4
5
6 (3,4) (1,2) (0,6)

Fig. 2. A pictorial representation of a skipper (skipper 5) with multicoloured state and

grey timeslots. The schedule for skipper 5 is in bold and the DFA for criteria 4, 11
and 12 is drawn with state Start in white, First in blue, Mid in yellow, Last in
green, Bye in pink and End in red. The edge marked  is explained in the text.

arcs in Figure 2 represent possible transitions from one state to another. The
DFA is encoded as a set of transition objects qi , ι, qj , where a transition is
made from state qi to state qj when encountering input ι. In addition we specify
the accepting states, which are all states except the start state (white) and the
bye state (pink). This constraint also ensures that if a skipper has finished all
matches in flight i (i.e. is in state End) then that skipper will also have finished
all matches in flight i + 1 (and so on). Note that when we have as many boats
as skippers, the directed edge (Bye, Last) (marked ) becomes bidirectional.
Figure 2 also shows the skipper-oriented variables and constraints corre-
sponding to skipper 5 in the schedule shown in Table 1. The state array is
presented as coloured boxes, and below that is the timeSlot array. The arrows
represent the channeling constraints (C1). The state array is coloured with green
representing Last, yellow for Mid, blue for First, red for Last and (not shown)
pink for Bye. The schedule of Table 1 is reproduced with the matches for skipper
5 emboldened.

Boat changes: A boat change occurs when a skipper has been in a Bye state in
flight i and then is in a match in the next flight i + 1. This is encoded using the
array of zero/one variables (V3), and the constraint (C4), and accumulated into
totalChanges (V4, C5).
 modMatch
0 1 2 3 4 5 6
0 - 0 1 3 5 2 6
1 0 - 6 2 3 1 4
schedule
2 1 6 - 0 4 5 2
3 3 2 0 - 6 4 5 Flight Matches
4 5 3 4 6 - 0 1 0 (0,1) (2,3) (4,5)
5 2 1 5 4 0 - 3 1 (0,2) (4,6) (1,5)
2 (2,6) (0,5) (1,3)
6 6 4 2 5 1 3 -
3 (5,6) (0,3) (1,4)
4 (3,5) (1,6) (2,4)
match 5 (3,6) (0,4) (2,5)
0 1 2 3 4 5 6 6 (3,4) (1,2) (0,6)
0 - 0 3 10 16 7 20
1 - 19 8 11 5 13
2 - 1 14 17 6
3 - 18 12 15
4 - 2 4
5 - 9
6 -

time (0,1) (2,3) (4,5) (0,2) (4,6) (1,5) (2,6) (0,5) (1,3) (5,6) (0,3) (1,4) (3,5) (1,6) (2,4) (3,6) (0,4) (2,5) (3,4) (1,2) (0,6)
Constructing Sailing Match Race Schedules: Round-Robin Pairing Lists

Fig. 3. A pictorial representation of the entire model of the schedule for 7 skippers and 6 boats. The schedule is reproduced on the right.
In the centre we have the 7 skippers, on left the modMatch and match arrays, and along the bottom the time array. The arrows signify
channeling between parts of the model.
677
678 C. Macdonald et al.

Match perspective: We are now in a position to model criterion 1 (each skip-

per sails against every other skipper) and optimization criterion 10 (minimize
boat changes). In the second half of Figure 1 we present a match perspective
of the schedule, using a two dimensional array of variables match (V5), where
match[i, j] is the time slot in which skippers σ[i] and σ[j] meet in a match. Only
the half above the diagonal is represented, and the lower half is made up of
exactly the same variables, i.e. match[i, j] is exactly the same constrained vari-
able as match[j, i]. Constraint (C6) states that a match between skippers σ[i]
and σ[j] takes place at time k (i.e. match[i, j] = k) if and only if skipper σ[i]’s
k th time slot is skipper σ[j] and conversely that skipper σ[j]’s k th time slot is
skipper σ[i]. Variables (V6) and constraints (C7) then convert this from time
slots to flights, i.e. modMatch[i, j] is the flight in which skippers σ[i] and σ[j]
meet for a match. Finally, constraint (C8) ensures that each skipper’s match
occurs in different flights. Also, since modMatch[i, j] ≡ modMatch[j, i], setting
rows to be all different also ensures that columns are all different.

Temporal perspective: We now take a temporal view (V7), such that time[τ ] is
a pair (i, j), stating that at time τ skippers σ[i] and σ[j] are in a match. We
channel between the match perspective and the temporal perspective using (C9).

Optimization criteria: Finally we have the optimization criterion (criterion 10)

to minimize the number of boat changes. This is handled by (V8) and (C10).

Decision variables: The decision variables are time[0] to time[t − 1], i.e. for each
time slot we decide what match takes place. A symmetry break is made at the
top of search by forcing the first round to contain the matches (0, 1), (2, 3), ... ,
i.e. ∀i ∈ {0 ... m − 1} : match[2i, 2i + 1] = i. (This is independent of criterion 9,
which will be enforced in a subsequent stage by renumbering.)

Figure 3 gives a pictorial view of the entire model, using the 7 skipper and 6
boat problem. On the right we have the 7 skippers with their states and time
slots. Again, on the right we give the schedule actually produced. On the left we
have the modMatch and match arrays, and at the bottom the time array. The
arrows show the channeling between parts of the model.
The schedule presented has 6 boat changes: there are no boat changes in
flight 0 (first flight), in flight 1 skipper 6 has a boat change, in flight 2 skipper 3
has a boat change, in flight 3 skipper 4, in flight 4 skipper 6, in flight 5 skipper
0, and flight 6 skipper 1.

3.2 Stage 2: Minimizing Imbalances

Having produced a schedule that minimizes boat changes, we then minimize
imbalance (i.e. apply criterion 6) by extending the model in Figure 4. Assuming
the first stage produces a schedule with β boat changes we post this as a hard
constraint (C11) in stage 2. For each skipper σ we introduce a zero/one variable
position[i, j] (V9) which is 1 if and only if skipper σ is in a match in time m·j +i,
 Constructing Sailing Match Race Schedules: Round-Robin Pairing Lists 679

The objective value from stage 1 is used as a hard constraint:

totalBoatChanges ≤ β (C11)

A copy of these variables and constraints is created for each skipper σ:

∀i ∈ {0 ... m−1} : ∀j ∈ {0 ... f −1} :
position[i, j] ∈ {0, 1} (V9)
position[i, j] = 1 ⇔ timeSlot[m.j + i] ≥ 0 (C12)
∀i ∈ {0 ... m−1} :
imbalance[i] ∈ N (V10)
 
 f −1 
n − 1  
imbalance[i] =  − position[i, j] (C13)
 m 
j=0

maxImbalance ∈ N (V11)
maxImbalance = maximum(imbalance) (C14)

We minimize the maximum imbalance over all skippers:

∀i ∈ {0 ... n−1} : imbalance[i] ≡ σ[i].maxImbalance (V12)
maxImbalance ∈ N (V13)
maxImbalance = maximum(imbalance) (C15)
minimise(maxImbalance) (C16)

Fig. 4. Additions to the constraint model in Figure 1 for stage 2. The constant β is
the minimum number of boat changes found in stage 1.

i.e. in the ith match of the j th flight (C12). Imbalance (V10) is then computed
for each position, as the absolute amount over or under the expected presence
in a given position (C13). Finally, we compute the maximum imbalance over all
positions for this skipper (V11, C14).
We then capture the maximum imbalance over all skippers (V12, V13
and C15) and minimize this (C16). As before, the decision variables are time[0]
to time[t − 1] and the same symmetry breaking is imposed at top of search.

3.3 Stage 3: Renaming Skippers

Stage 3 then takes as input the schedule produced in stage 2 and renames skip-
pers in the last round of the last flight to be (0, 1), satisfying criterion 9. This
is a simple linear process. The schedule produced for 7 skippers and 6 boats is
shown in Table 2.
680 C. Macdonald et al.

Flight Matches
Table 2. A round-robin schedule with 7 skippers,
0 (0,6) (2,3) (4,5) 6 boats and 7 flights, which is the result of stages
1 (0,2) (4,1) (6,5) 1, 2, and 3. It has 6 boat changes and imbalance
2 (2,1) (0,5) (6,3) 1. An example of imbalance is skipper σ[0] with 2
3 (5,1) (0,3) (6,4) matches in position 0 (first), 3 in position 1 and 1
4 (3,5) (6,1) (2,4) (mid) in position 2 (last). Perfect balance would be
5 (3,1) (0,4) (2,5) 2 matches in each position, only achieved by σ[2].
6 (3,4) (6,2) (0,1)

3.4 Stage 4: Orienting Matches

The final stage, stage 4, orients skippers within matches, such that they are either
port (red) or starboard (green). This is done to satisfy criteria 2, 3, 5 and 8. A
sketch of the constraint model is given with a supporting diagram, Figure 5.
A two dimensional array of constrained integer variables orient[i, j] ∈ {0, 1}
has orient[i, j] = 0 if and only if skipper σ[i] is on port side in his j th match
(starboard otherwise). For example, taking the schedule in Table 2, skipper
σ[2] meets skippers in the order 3,0,1,4,5,6; skipper σ[3] meets skippers in the
order 2,6,0,5,1,4 and σ[5] meets skippers in the order 4,6,0,1,3,2. Therefore
orient[2, 0] = 1 and orient[3, 0] = 0 (criterion 8), orient[2, 4]
= orient[5, 5] and
orient[3, 3]
= orient[5, 4]. Summation constraints on each skipper enforce crite-
ria 2 and 3 (equal number of port and starboard matches). Criterion 8 is enforced
by posting constraints such that in a match (σ[i], σ[j]) where |i − j| = 1, the
higher indexed skipper (lower ranked) is on the port side. A regular constraint is
posted for criterion 5 (restricting sequences of port and starboard assignments).
In Figure 5, port is red and starboard green. The sequence of zero/one vari-
ables for each skipper is shown on the left. Bottom right is the DFA, and top
right is the final schedule produced from stages 1 to 4. The schedule has 6 boat
changes and a maximum imbalance of 1. The schedule presented here is new,
not appearing in the example schedules provided in the manual [2].

4 Sample Schedules
We present four pairs of schedules. The first pair is for 8 skippers and 6 boats,
one published by ISAF, the other generated by us. Next, we present two new
schedules, not appearing in the ISAF manual, for 9 skippers and 6 boats and 9
skippers with 8 boats, then 10 skippers and 6 boats, again both published by
ISAF and generated by us. Finally, we show what is a relatively easy problem, 8
skippers and 8 boats. For each of our schedules (with the exception of the latter
“easy-8” problem) we used 1 day CPU time for the first stage (minimizing boat
changes) then a further 2 days to minimize imbalance. We were unable to prove
optimality in this time, therefore some of these schedules might be improved.
The processor used was an Intel (R) Xeon (R) E5-2660 at 2.20GHz. Note, stages
3 and 4 (renaming and orientation) completed in less than a second.
 Constructing Sailing Match Race Schedules: Round-Robin Pairing Lists 681

Fig. 5. A pictorial representation of the orientation process, stage 4. Top right is the
actual oriented schedule. Below that is the DFA to satisfy criterion 5. On the left is the
zero/one variables, a row representing a skipper’s sequence of competitors. An edge
between variables corresponds to a matched pair, that must take different values.

Flight Matches Flight Matches

0 (5,2) (4,3) (1,6) 0 (2,7) (3,0) (5,4)
1 (4,2) (6,5) (3,1) 1 (0,2) (4,3) (7,5)
2 (6,4) (2,1) (3,5) 2 (0,4) (7,6) (5,3)
3 (6,2) (5,0) (1,7) 3 (4,6) (5,0) (3,1)
4 (5,1) (0,6) (2,7) 4 (6,5) (3,2) (1,4)
5 (7,5) (2,0) (6,3) 5 (6,3) (4,7) (2,1)
6 (0,7) (4,1) (3,2) 6 (7,3) (1,6) (4,2)
7 (7,4) (0,3) 7 (7,1) (2,5) (6,0)
8 (4,0) (7,3) 8 (5,1) (0,7) (6,2)
9 (5,4) (7,6) (1,0) 9 (1,0)

Fig. 6. Schedules for 8 skippers and 6 boats. On the left, the published ISAF schedule
(illegal), and on the right is our schedule.

8 skippers and 6 Boats: We first analyze the ISAF schedule, Figure 6 on the
left. Criterion 4 (skippers in last match in a flight cannot be first in next flight)
is violated on 3 occasions (skipper 1 in flight 3, skipper 7 in flight 4, skipper
0 in flight 7). Criterion 5 (no more than two consecutive port or starboard
assignments) is violated for skipper 6 (flights 1 to 3) and skipper 7 (flights 7 to
9). Criterion 6 (imbalance) is violated for skippers 6 and 8. Criterion 12 (new
boats do not sail first) is violated in flight 9 for skipper 5. Finally, the schedule
has 8 boat changes. Our schedule, Figure 6 on the right, respects all criteria and
has 6 boat changes and a maximum imbalance of 1.
682 C. Macdonald et al.

Flight Matches Flight Matches

0 (0,7) (3,2) (5,4) 0 (0,6) (3,2) (5,4) (1,7)
1 (0,2) (7,4) (5,3) 1 (2,0) (6,3) (4,1) (8,7)
2 (4,0) (7,3) (2,5) 2 (6,4) (0,5) (7,2) (3,8)
3 (3,0) (2,4) (6,5) 3 (4,7) (1,8) (0,3) (5,2)
4 (4,3) (5,1) (8,6) 4 (3,1) (7,5) (8,0) (2,6)
5 (3,1) (6,2) (4,8) 5 (3,7) (4,8) (2,1) (6,5)
6 (6,3) (1,4) (8,2) 6 (8,2) (1,6) (5,3) (0,4)
7 (4,6) (2,7) (1,8) 7 (5,1) (7,0) (6,8) (4,3)
8 (7,6) (0,8) (2,1) 8 (8,5) (2,4) (7,6) (1,0)
9 (8,7) (5,0) (6,1)
10 (8,5) (1,7) (0,6)
11 (7,5) (3,8) (1,0)

Fig. 7. Two new schedules. On the left, 9 skippers and 6 boats, and on the right 9
skippers and 8 boats.

9 skippers: Two new schedules are presented in Figure 7, on the left for 6 boats
and on the right for 8 boats. Neither of these schedules appear in the ISAF
Umpires’ Manual. Both schedules respect all criteria. For 6 boats there are 8
boat changes (no skipper making more than 1 boat change) and a maximum
imbalance of 2. For 8 boats there are again 8 boat changes, no skipper with
more than 1 boat change, and each skipper 4 times on starboard side and 4 on
port side.

10 skippers and 6 boats: Figure 8 shows the ISAF schedule on the left for 10
skippers and 6 boats and on the right, our schedule. The ISAF schedule violates
criterion 5 (no more than 2 consecutive port or starboard assignments) ten times
(twice for skippers 0, 1 and 6 and once for skippers 3, 7, 8 and 9). Criterion 12
(new boats do not sail first) is violated 7 times (in flight 3 for skippers 6 and
2, flight 5 for skipper 8, flight 6 for skipper 7, flight 8 for skipper 1, flight 9 for
skipper 4 and flight 12 for skipper 4 again). There are 22 boat changes, with the
minimum changes for a skipper being 1 and the maximum 3. Our schedule, right
of Figure 8, satisfies all criteria, has 12 boat changes with the minimum changes
for a skipper being 0 and the maximum 2, and a maximum imbalance of 1.

8 skippers and 8 boats (easy-8): Figure 9 shows the ISAF schedule on the right
for 8 skippers and 8 boats and on the right, our schedule. The ISAF schedule
violates criterion 4 (last match in a flight cannot be first in next flight) for
skipper 5 in flights 1 and 2. Also, criterion 5 (no skipper should have more than
2 consecutive port or starboard assignments) is violated for skippers 0, 1, 3, 4 and
6. Furthermore, skipper 1 appears in the last match of four flights resulting in
significant imbalance. Our schedule, on the right, satisfies all criteria, is optimal
(maximum imbalance of 1) and took less than 10 seconds in total (all stages) to
produce.
 Constructing Sailing Match Race Schedules: Round-Robin Pairing Lists 683

Flight Matches Flight Matches

0 (8,3) (1,7) (0,9) 0 (4,7) (3,2) (0,5)
1 (7,3) (0,8) (1,9) 1 (2,4) (7,0) (3,5)
2 (0,7) (9,3) (1,8) 2 (0,4) (7,3) (5,2)
3 (6,2) (5,0) (7,4) 3 (4,3) (0,2) (5,7)
4 (5,2) (6,4) (3,0) 4 (3,0) (6,5) (2,7)
5 (8,6) (2,9) (4,1) 5 (6,3) (7,1) (8,2)
6 (7,2) (4,8) (9,6) 6 (7,6) (2,1) (9,8)
7 (2,4) (0,6) (7,5) 7 (2,6) (8,7) (4,9)
8 (6,1) (5,9) (8,2) 8 (6,8) (1,4) (7,9)
9 (9,4) (8,7) (6,5) 9 (8,1) (5,4) (9,6)
10 (9,7) (6,3) (8,5) 10 (5,1) (9,2) (4,6)
11 (7,6) (3,1) (9,8) 11 (1,9) (6,0) (8,4)
12 (4,3) (5,1) (2,0) 12 (0,9) (3,8) (6,1)
13 (3,5) (4,0) (2,1) 13 (8,0) (9,5) (1,3)
14 (5,4) (3,2) (1,0) 14 (5,8) (9,3) (1,0)

Fig. 8. Schedules for 10 skippers and 6 boats. On the left, the ISAF published schedule
(illegal), and on the right, our schedule.

Flight Matches Flight Matches

0 (5,2) (4,3) (1,6) (0,7) 0 (4,3) (2,0) (1,5) (7,6)
1 (2,4) (0,6) (1,7) (5,3) 1 (3,2) (0,4) (6,1) (5,7)
2 (0,5) (7,2) (6,3) (4,1) 2 (4,1) (7,3) (5,2) (0,6)
3 (7,3) (5,1) (4,0) (6,2) 3 (1,7) (6,5) (3,0) (2,4)
4 (3,0) (4,7) (6,5) (2,1) 4 (0,5) (4,6) (7,2) (3,1)
5 (6,4) (7,5) (2,0) (3,1) 5 (7,0) (2,1) (6,3) (5,4)
6 (7,6) (5,4) (3,2) (1,0) 6 (6,2) (5,3) (4,7) (1,0)

Fig. 9. Schedules for 8 skippers and 8 boats. On the left, the published ISAF schedule
(illegal), and on the right, our schedule.

5 Discussion

It was anticipated that this problem would be easy to model. This naı̈vety was
due to the use of the term “round-robin” for these schedules, leading us to believe
that we could use some of the models already in the literature [1,6,7]. This
naı̈vety slowly dissolved as we addressed more and more of the ISAF criteria. A
number of models we produced were scrapped due to misunderstanding about the
actual problem, i.e. a communication problem between the authors. Eventually
this was resolved by presenting the majority of the criteria as a deterministic
finite automaton. This became our Rosetta Stone, with the surprising benefit
that it not only improved the team’s communications, it was also a constraint
(the regular constraint).
The staged approach was taken cautiously. We expected that isolating the
orientation of schedules as a post-process might leave us with a hard or insoluble
684 C. Macdonald et al.

problem. So far, every schedule we have produced has been oriented with very
little search, typically taking less than a second.
We used the Choco constraint programming toolkit and one of our goals was
to use only toolkit constraints, i.e. we wanted to see how far we could go without
implementing our own specialized constraints. We are pleased to say that we did
not have to do anything out of the box, although we did feel that we had used
just about every constraint there was.
There are many redundant variables in our model. One of our first valid mod-
els was essentially that shown in Figure 2, and did not use modMatch and match
variables. Performance was hopeless, struggling to produce a 7 skipper 6 boat
schedule in days. The modMatch and match variables were added. This improved
domain filtering, and a flattened match array was used as decision variables.
That is, the decisions where “when do we schedule this match?”. At this point
we had not yet produced a 9 skipper 6 boat schedule and we suspected that
the combined criteria might exclude such a schedule. A non-constraint model
was developed, primarily to attempt to prove that there was no schedule for 9
skippers with 6 boats. This program used a backtracking search with decision
variables being positions within flights, i.e. time slots being assigned to matches.
At the top of search, the matches in the first flight were set, and symmetry
breaking was used to reduce the number of legal second flights. A solution was
found in seconds! With this experience we added in the time variables, channeled
these into the existing model and used these as decision variables, i.e. the ques-
tion now was “what match will we do now?”. We also anchored the matches in
the first flight. With these changes we began to produce schedules in acceptable
time. The model was then built upon, incrementally adding criteria. This took a

Fig. 10. Scene from an imaginary dinghy race

 Constructing Sailing Match Race Schedules: Round-Robin Pairing Lists 685

surprisingly short amount of time, sometimes minutes of coding to add in a new

feature. The model was then enhanced so that it would address optimization
rather than just satisfaction, again a trivial programming task.
We have only reported a handful of our schedules, however there are missing
schedules i.e. unpublished and not yet produced by us. Examples of these are 10
skippers and 8 boats, 11 skippers with fewer than 10 boats, 12 skippers and 8
boats, 13 skippers and 8 boats and 14 skippers with fewer than 10 boats. None of
these schedules have been published by the ISAF, although we expect that they
can be produced by selectively violating some of constraints. We have also not
yet encoded criterion 7, addressing the situation of 10 or more boats. To do this
we will have to modify the DFA used by the regular constraint in the first two
stages. In addition, we have yet to address the format of two-group round-robin
schedules as also found in the manual [2].

6 Conclusion

We have produced new and better match race schedules. These schedules can
be used by anyone who competes under the criteria published by ISAF. Our
schedules can be downloaded as blank schedules and then populated with the
names of the actual competing skippers.
So, why did we use constraint programming? The answer is obvious: we are
constraint programmers, or to borrow Mark Twain’s words “To a man with a
hammer, everything looks like a nail”. But actually, constraint programming has
been a good way to go. From an engineering perspective, it has allowed us to
prototype solutions quickly and to build solutions incrementally. There is also an
advantage that we might exploit in the future: we can now investigate the effect
different criteria have on our ability to produce schedules and how relaxation of
those affect optimization criteria. That is, the constraint program might be used
to design the next set of criteria for match race schedules, or allow the event
organizer to decide which criteria to sacrifice to ensure a faster schedule with
fewer boat changes.

References
1. Henz, M., Müller, T., Thiel, S.: Global constraints for round robin tournament
scheduling. European Journal of Operational Research 153(1), 92–101 (2004)
2. International Sailing Federation. ISAF International Umpires’ and Match Racing
Manual (2012)
3. Lombardi, M., Milano, M.: Optimal methods for resource allocation and scheduling:
a cross-disciplinary survey. Constraints 17(1), 51–85 (2012)
4. Pesant, G.: A regular language membership constraint for finite sequences of vari-
ables. In: Proceedings of the 10th International Conference on Principles and
Practice of Constraint Programming, CP 2004, Toronto, Canada, September 27-
October 1, pp. 482–495 (2004)
686 C. Macdonald et al.

5. Quimper, C.-G., van Beek, P., López-Ortiz, A., Golynski, A., Sadjad, S.B.S.: An
Efficient Bounds Consistency Algorithm for the Global Cardinality Constraint. In:
Rossi, F. (ed.) CP 2003. LNCS, vol. 2833, pp. 600–614. Springer, Heidelberg (2003)
6. Rasmussen, R.V., Trick, M.A.: Round robin scheduling - a survey. European Journal
of Operational Research 188(3), 617–636 (2008)
7. Trick, M.A.: Integer and Constraint Programming Approaches for Round-Robin
Tournament Scheduling. In: Burke, E.K., De Causmaecker, P. (eds.) PATAT 2002.
LNCS, vol. 2740, pp. 63–77. Springer, Heidelberg (2003)
 Design and Evaluation of a Constraint-Based
Energy Saving and Scheduling Recommender
System

Seán Óg Murphy1,2(B) , Óscar Manzano1,2 , and Kenneth N. Brown1,2

1
International Energy Research Centre, Cork, Ireland
2
Insight Centre for Data Analytics, Department of Computer Science,
University College Cork, Cork, Ireland
{seanog.murphy,oscar.manzano,ken.brown}@insight-centre.org
http://www.insight-centre.org

Abstract. Development of low-cost and inter-operable home sensing

products in recent years has motivated the development of consumer-
level energy and home monitoring software solutions to exploit these new
streams of data available to end-users. In particular, this opens up the
home energy space as an area of high potential for the use of consumer-
level energy optimisation with home-owners actively engaged with data
about their energy use behaviour. We describe the development of a
tablet-based home energy cost saving and appliance scheduling system
which calculates behaviour change suggestions that save occupants on
their energy bills while avoiding disruption to their regular routines. This
system uses a Constraint Satisfaction Problem Solver to compute savings
based on real-world sensor data, and to generate revised schedules in a
user-friendly format, operating within a limited computing environment
and achieving fast computation times.

1 Introduction
In-home networked sensing has, in recent years, reached a level of cost, reliability
and protocol maturity that enables consumer-level home sensing to a degree
not possible before. The increasing variety and availability of Zigbee-compatible
wireless sensing allows for rich sensor coverage of homes with minimal impact or
installation, providing rich data streams from around the home, monitoring use
of space, appliance use, temperature and more. Providing these data streams in
an easy to understand format is of substantial benefit to home-owners to help
understand inefficient heating of spaces, electricity efficiency and overall energy
costs [24].
Furthermore, developments in energy infrastructure and generation have lead
to a greater variety of energy production sources and tariff schemes, leading to a
greater incentive (both in terms of cost and in environmental impact) for home-
owners to take pay attention to their energy consumption over different time
periods. Some regions use variable tariffs based on the time of day [25], and

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 687–703, 2015.
DOI: 10.1007/978-3-319-23219-5 47
688 S.Ó. Murphy et al.

some schemes feature time-variable peak-demand limits with punitive charges

incurred where total concurrent energy use exceeds a limit at a particular time
of the week [1].
In this work we describe a Constraint-based system that recommends adjust-
ments to energy use patterns to provide targetted cost savings in the home, based
on historical and current observation of appliance use patterns in that home. The
data is obtained from a deployed sensor network and energy monitoring system.
The adjusted appliance use pattern accepted by the home-owner then forms
the basis of a appliance use scheduling application, which recommends a set of
activation times that correspond to observed historical behaviour while respect-
ing limits of overall capacity and variable demand pricing schemes. This system
operates on a low-power Android tablet device, as part of a Reporting Tool app
which provides robust, easy-to-understand feedback to home-owners from the
data streams provided by the home sensor deployment.

2 AUTHENTIC Smart Home Project

AUTHENTIC is a multi-institutional initiative to design and deliver a Home

Area Network (HAN) infrastructure capable of supporting opportunistic decision
making for effective energy management within the home.
The Authentic HAN consists of a collection of Zigbee wireless sensors com-
municating over a multi-hop wireless network with a low-power gateway Linux
PC. This hub stores the sensor readings in a MySQL database. The readings
are retrievable through a REST (Representational State Transfer) interface [19]
in the form of JSON (Javascript Object Notation) Objects [6] which are trans-
formed and can be displayed to the users via a reporting tool (discussed below).
The first phase of deployments included smartplug sensors attached to power
sockets which report appliance activations and energy consumption, passive
infrared motion detectors reporting space occupancy, temperature, light and
humidity sensors, and contact sensors reporting when doors and windows are
opened or closed (Figure 1). Future phases include wireless smart meters for
measuring total electricity, gas and oil consumption, sensors for monitoring the
use and settings of heating, ventilation, air conditioning and water heating, and
sensors and meters for monitoring the available of energy from renewable sources.
In order to assist the users to monitor and change their energy consump-
tion, and to track their energy bills, a Reporting Tool app has been developed
for Android tablets. The app provides direct and instant feedback, converting
complex sensed data into customisable human-readable reports; sample screen-
shots are shown in Figure 2. Responding to user queries for specified time peri-
ods, the app can create visualisation of, for example, total energy consumption,
consumption by room, consumption by appliance, room occupancy patterns,
correlations between room occupancy and temperature or appliance use, and
comparisons between different appliances and different time periods. The app
includes a facility for real-time alerts when specified or anomalous patterns occur
(e.g. windows opening in a heated room, or instantaneous consumption above
 Design and Evaluation of a Constraint-Based Energy Saving 689

Fig. 1. AUTHENTIC sensors (movement, light, humidty, contact and smartplug sen-
sors)

(a) Appliance Use History

(b) Plan Recommendation (number of appliance acti-

vations)

Fig. 2. Android reporting tool

690 S.Ó. Murphy et al.

a threshold). Finally, the app provides facilities for guiding the user to change
their behaviour. At a basic level, the app allows the user to set annual energy
goals, and to track progress towards those goals. More sophisticated feedback,
including advice on how to achieve specified reductions, is the main topic of this
paper, and is discussed in the remaining sections.

3 Requirements
Users need guidance to help change their energy consumption patterns in accor-
dance with their own goals and preferences. We consider two high level goals:
reducing cost, and reducing total energy consumption1 . The first steps are under-
standing how much energy is consumed by each activity and what cost is incurred
by those activities at different times. Reducing consumption itself is easy – appli-
ances can simply be switched off – but the difficulty is in balancing the reduction
in consumption with the users’ preferences for safe and comfortable living. Low-
ering a heating thermostat by 2 degrees may be unsafe in winter, while switching
off an entertainment console may be impractical.
The aim is to propose changes to a user’s behaviour which achieve their
energy goals, while still being acceptable for their health, comfort and enjoyment.
Further, those proposed changes should be personal, tailored to the users in
a specific home. Thirdly, the interaction load imposed on the users must be
manageable: systems which require extensive elicitation of preferences and utility
functions under different scenarios before any proposal can be made are simply
impractical in most domestic settings. Finally, the system should respect the
privacy of the users, and should be capable of operating entirely within the home,
on devices with limited computing power, without releasing data or relying on
data obtained from other users outside the home.
Our intention is to guide the users based on the observed behaviour of those
users in their home. The data streams from the sensors and smart plugs provide
a history of appliance use, from which we can extract daily and weekly patterns.
Given a goal (e.g. reduce the energy bill by 10%), we will then search for minimal
changes to those patterns which satisfy the goal. Our proposals will be on two
levels. The first level will recommend the total number (or duration) of activa-
tions for each appliance over a specified period, and if differential tariffs are in
use, the total number of appliance activations in each price band.
The second level will propose a schedule of use for each day, ensuring the
total simultaneous load is within any threshold, again representing a minimal
change to observed behaviour patterns. Thus in each case an initial proposal is
tailored to each home, without requiring any interactive preference elicitation
apart from setting the initial high level goal. In each case, the users should then
be able to interact with the system, imposing new constraints or objectives if
the proposal is not satisfactory, and requesting a new recommendation. Further,
the system will then monitor energy consumption within the home, alerting the
1
Additional goals of reducing carbon footprint and reducing energy consumption from
non-renewable sources are being developed.
 Design and Evaluation of a Constraint-Based Energy Saving 691

user when targets are not going to be met, and offering recomputation of the
schedules or activation levels.
Our problem thus involves multiple levels of decision support and optimisa-
tion, over constrained variables. The problems range from relatively simple linear
assignment problems to constrained scheduling, and in each case the problems
may be extended by additional user constraints. Thus we choose to use constraint
programming as the single technology for all problem variants because of the
dedicated support for modelling cumulative scheduling problems and because
of its flexibility in adding what ultimately could be arbitrary side constraints
and preferences. The Authentic system also imposes some technical limitations,
which influence our choice of solver. The Reporting Tool app (Section 2) runs
under Android, and is limited to JDK 1.6 for library compatibility issues. Thus
we model and solve the decision and optimisation problems using the Choco
2.1.5 Java library, as it is also JDK1.6 compatible. This places a restriction on
the variable types and constraints we use.

4 Constraint-Based Recommender System

4.1 System Overview

The overall system is structured as follows. The Authentic HAN produces a

database of sensor readings for the home. This data is transformed into appli-
ance activation reports. Information is presented to the user, who specifies an
energy or cost goal for the next period. From the appliance reports, the system
constructs a constraint optimisation model, and proposes a high-level energy
consumption plan to the user. It iterates with the user until an acceptable plan
is agreed. From the high-level plan, the system then creates a scheduling problem
and proposes a schedule to the user that respects the plan and requires a minimal
change to previous behaviour. Again, the user iterates with the system until an
acceptable schedule is agreed. The user is in control of all activations, and the
system monitors compliance, alerts the user when targets are to be missed, and
propose new plans or schedules in response. We describe the constraint-based
modules in Sections 4 and 5, and evaluate performance in Section 6.

4.2 Data Preprocessing (Data Analysis Module)

To pre-process the sensor data, we developed a Data Analysis module which
interfaces with the Authentic RESTful API [2] [18] to retrieve home sensor data
which it discretises into forms more amenable to analysis and optimisation.
Sensor readings and events are retrieved in the form of JSON Objects which
are then parsed by the module. By parsing the sensor information, data manipu-
lation can be performed to convert this information into alternative forms more
suitable for optimisation problems or efficiency analysis. In general, the facility
is provided to convert Occupancy, Temperature, Light, Humidity and Appliance
Activation information into arrays of samples (at a user-configurable sampling
692 S.Ó. Murphy et al.

rate). This allows for the like-for-like comparison of readings between arbitrary
timepoints and aribtrary sensing types (e.g. plotting Occupancy vs Television,
Light vs Lamp, or Humidity vs Dryer).
An additional function provided by this module is the generation of “Acti-
vationTuples”, data objects representing the usage pattern for appliances in
the home which we use as input for the Appliance Use Optimisation Module
(Section 4.3).
ActivationTuples contain the following information:
– Appliance Name
– Average Consumption Per Activation (Watt-Hours)
– Number of Activations per Tariff (in the case of Day/Night or other variable
tariffs)
– Power Draw Profile (time series, Watts)

Activations are considered from the time an appliance sensor reports instan-
taneous demand above a minimum threshold until the instantaneous demand
returns to below that threshold. The average consumption is based on the watt-
hour (WH) consumption per activation period. In the case of appliances with
variable activation profiles (for example, different washing machine settings), we
can subdivide the appliance into seperate sub-appliances, one for each setting
or pattern. The Power Draw Profile represents the instantaneous power demand
in watts over a time series with configurable sampling rate (e.g. 5 minutes).
The Power Draw Profile is used with the Scheduling Module to ensure that the
total concurrent power load in a home or on a circuit at a given time can be
accounted for in the constraint satisfaction problem. As the reporting interval of
the smart-plug sensors is generally 5 minutes, higher resolution sampling rates
are not required.

4.3 Appliance Use Optimisation Module (Solver)

The aim of the Appliance Use Optimisation module (“Solver”) is to propose
high-level patterns of appliance use which will achieve desired cost or energy
savings while minimising disruption to the household. These patterns are simply
counts of each appliance use in each price tariff. The module retrieves usage
patterns from the analysis module in the form of activation tuples, and uses
these to create a constraint satisfaction or optimisation problem. In this work
we consider three tariffs (high, medium, low) available each day, although the
solver is flexible and any number of price bands can be represented. The solver
then searches for a modified pattern of use, by moving activations to lower
price bands or reducing the number of activations, with a constraint on total
cost. To avoid disruption, and to promote acceptance of the plan, we include an
objective to minimise the deviation between the historical activation use pattern
and the proposed pattern, represented as the sum of the squared differences of
the appliance use counts). The constraint model is shown in Tables 1 and 2.
 Design and Evaluation of a Constraint-Based Energy Saving 693

Table 1. Appliance Use Optimisation variables

Variable Description
Name
Constants
InputCost Cost of the historical period
T arget Reduction in cost required (e.g. 0.9 of the previous cost)
H, L, M Price per unit(wH) of electricity at High, Medium and Low tar-
iffs
Ai InputAct Original Number of activations of Appliance i in input
Ai Cons Average Consumption of Appliance i (wH) per activation
Ai InputH Original number of activations at High tariff for Appliance i in
input
Ai InputM Original number of activations at Medium tariff for Appliance i
in input
Ai InputL Original number of activations at Low tariff for Appliance i in
input
Variables
Ai HAct Number of activations of Appliance i at High tariff times
Ai M Act Number of activations of Appliance i at Medium tariff times
Ai LAct Number of activations of Appliance i at Low tariff times
Auxilliary Variables
Ai HDif f The difference between historical High use and the new plan for
Appliance i
Ai M Dif f The difference between historical Medium use and the new plan
for Appliance i
Ai LDif f The difference between historical Low use and the new plan for
Appliance i
Ai HCost Total Cost for Appliance i at High tariff
Ai M Cost Total Cost for Appliance i at Medium tariff
Ai LCost Total Cost for Appliance i at Low tariff
Ai T otalCost Total cost for Appliance i
Ai T otalDif f Sum of the squared differences for Appliance i
T otalCost Total cost of the new plan
TotalDiff Objective variable. Sum of squared differences
Ai Act Total Number of activations of Appliance i

The only input required from the user for the first solution is to specify the
required energy or cost reduction, and thus the interaction load on the user is
minimal. However, it is likely that the proposed changes may not satisfy the
user’s requirements. Thus after the first schedule is proposed, the user is invited
to specify additional constraints. The main type of user constraint is a domain
restriction, specifying for a single appliance the acceptable range of activations in
any price band or in total. Additional constraints, including for example, that the
total number of activations for one appliance must be not less than the number
for another can easily be handled by the solver, but are not yet implemented in
the interface. The intention is to regard these user constraints as critiques of the
694 S.Ó. Murphy et al.

Table 2. Appliance Use Optimisation constraints

Constraints Description
Ai HCost = Ai HAct ∗ H Set the cost at H for Appliance i
Ai M Cost = Ai HAct ∗ M Set the cost at M for Appliance i
Ai LCost = Ai HAct ∗ L Set the cost at L for Appliance i
Ai HDif f = (Ai HAct − Ai InputH)2 Set the Squared Difference in H.
2
Ai M Dif f = (Ai M Act − Ai InputM ) Set the cost at Squared Difference
in M
Ai LDif f = (Ai LAct − Ai InputL)2 Set the cost at Squared Difference
in L
Ai T otalDif f = Ai HDif f + Ai M Dif f + Ai LDif f Set the sum Difference for Appli-
ance i
Ai T otalCost = Ai HCost + Ai M Cost + Ai LCost Set the total cost for Appliance i
Ai Act = Ai HAct + Ai M Act + Ai LAct Set the total number of activa-
tions for Appliance i

n
T otalDif f = (Ai T otalDif f ) Set the total sum of Differences
i=1

n
T otalCost = (Ai T otalCost) Set the total Cost
i=1
T otalCost ≤ InputCost ∗ T arget Ensure the TotalCost is below the
target price
M inimise(T otalDif f ) Objective is to minimise the sum
of squared differences

proposed solutions, and thus the user is not required to specify any preference
that is already satisfied. However, the added user constraints are stored, and
will be applied to future solving instances (and the user is invited to remove any
constraints that no longer apply).

5 Scheduling Module
The output from the optimisation module is a high level plan specifying the
number of activations of each appliance in each price band over a specified period.
The user is in control, and will choose when to activate an appliance. However,
as discussed above, the system will monitor the activations and alert the user
if the observed activations are not on track to meet the goals at the end of the
period. In cases where there are multiple price bands during a day, or where
there are peak power thresholds which invove a higher price, it may be difficult
for the user to balance their appliance use appropriately. Thus, we also provide a
task scheduler, which recommends start times for appliance use over a number of
days in order to respect the cost constraints. As with the high level plan, there
is an objective to find appliance start times which are as close as possible to
the observed historical patterns for the user. The system allows users to modify
 Design and Evaluation of a Constraint-Based Energy Saving 695

aspects of the schedule, by adding or modifying tasks and constraints. The user
is in control, but as before, the system will monitor appliance use compared to
the schedule and alert the user when thresholds are likely to be breached, and
offer the option of rescheduling.
The first aim is to generate a schedule which respects the cost constraints.
The number of activations for each appliance is obtained from the high-level plan,
and the required number of task instances are generated. In order to handle the
time granularity for different constraints, each task instance is then broken up
into multiple sub-tasks each of fixed duration, and sequencing constraints are
applied between the subtasks. Disjunctive constraints[8] are imposed on multi-
ple instance of the same appliance, to ensure that they cannot be scheduled to
operate simultaneously. The high level plan also specifies the number of activa-
tions to take place under each tariff, and so for each task instance, we designate
its highest permissible tariff and impose constraints preventing that task run-
ning in a higher tariff’s time period. Note that this is a permissive approach,
allowing the solver to find schedules with a lower cost than expected.
The next constraint to be considered is the peak power threshold. We handle
this by imposing a cumulative constraint [3] over all tasks. We associate heights
with each appliance task relative to their power demand (and in cases where
the power demand varies over the activation cycle, the heights vary across the
different sub-tasks). The cumulative constraint height parameter is set to to
the maximum power threshold, and in the case of variable maximum demand
schemes we create dummy tasks with appropriate heights to represent temporary
lower thresholds over appropriate time periods.
To ensure each schedule mimics prior behaviour, we analyse historic usage
patterns and infer a time series of the frequency with which an appliance was
active at a given time. The aim is then to minimise the abnormality of appliance
activation times; that is, to avoid recommending appliance use at times with
very low historic use frequency. To achieve this, we divide each day of the week
into timeslots based on the sub-task time granularity (in this work, 5 minutes).
For each appliance activation in the historical data, we increment a count in each
timeslot that matches the day and time (e.g. for a 15-minute activation we might
increment Tuesday 4:00. Tuesday 4:05 and Tuesday 4:10). Having done this for
several weeks worth of historical data, we now have a time series representing the
frequency any appliance was in use at any given time during the 7-day week. We
invert this time series (where F = the highest frequency in the series, all entries E
are changed to F-E, Figure 5) which we use to produce a set of dummy constant
Task Variables (each of height equal to the appropriate inverted frequency) for
use with another cumulative constraint.
The peak value for the cumulative constraint is then a variable, with the aim
being to minimise that peak, subject to the lower bound equal to the maximum
height of the inverse frequencies. For each appliance activation, the height of
all of its sub-tasks are scaled to be the gap between the highest and lowest
frequencies for that appliance, and the sub-tasks are added to the cumulative
constraint (Figure 5). Thus, if we schedule a task so that the full duration of
696 S.Ó. Murphy et al.

Fig. 3. Frequency Time Series and inversion to produce Masking Gradient

Fig. 4. Scheduling appliances using Masking Gradient

a task (i.e. each of its sub-task components) is at a time of highest frequency,

the cumulative height variable is set to the lower bound; if the appliance task is
scheduled so that it is active at the time of lowest historic frequency, then the
cumulative height is at its maximum. All appliances are then normalised, and
the optimisation objective is to minimise the sum of the appliance cumulative
heights; that is, we minimise the sum of the ’abnormalities’ of the activation
times over each appliance use.
 Design and Evaluation of a Constraint-Based Energy Saving 697

The initial solution to the scheduling can now be generated and displayed to
the user. Although it may seem counter-intuitive to schedule uses of appliances
like televisions or games consoles, it is important to account for their expected
use, in order to manage the use of other appliances around that time. Thus we
learn when those devices are most often used, and build that expected usage
into the schedule. This initial solution does not require any interaction with the
user to elicit preferences or objectives. We then allow the user to interact with
the scheduler to improve the schedule, if necessary. The user can add new tasks
or delete existing tasks for a particular day, can extend the duration of a task,
can fix the start time of a task (by sliding the task position on the displayed
schedule), and can add limited temporal constraints between tasks (for example,
a specific dryer task must start within 30 minutes of the completion of specific
washing task, or that the shower and washing machine cannot be active at
the same time), and then can request a rescheduling. For new tasks, the same
historical frequency pattern is applied; for user task time changes, the start time
is fixed, and the frequency mask is adjusted to that no penalty is applied.

Fig. 5. Scheduling three appliances accounting for concurrent power demand

6 Performance Evaluation

A series of experiments were performed to determine the performance of the

optimisation module on an Android tablet (Samsung Galaxy Tab 4 , Android
version 4.4.2). Three months of data from a real-world home deployment formed
the basis for the historical appliance use in these experiments, and the computa-
tion time taken for the Savings module was typically approximately 900ms and
the Scheduler completed its task in approximately 17 seconds (using 30 minute
time slots). These timings are well within the limits for acceptable response
time, and demonstrate the success of the constraint programming formulation
698 S.Ó. Murphy et al.

of the problems. As there is the facility for users to influence the savings prob-
lem through the addition of constraints on the number of appliance activations
at different tariffs (i.e. the user can specify more limited ranges for the number
of activations of appliances during particular tariffs, or to specify limits on the
amount of adjustment made), we investigated the impact of the introduction of
preferences on the running time and solvability of the savings problem. We also
investigate the impact of larger and smaller time-slots on scheduler performance,
with and without some user constraints applied.

6.1 Solver Performance

To investigate performance on the extreme end of user interaction with the appli-
ance savings module, we performed experiments where we gradually restrict the
scope of adjustment available to the solver. From savings targets of 10% through
to 60%, we performed the savings routine while gradually restricting the range
of acceptable values (or “freedom”) for appliance activations at each tariff. For
instance, at “40%” freedom, the solver is free to adjust the number of activations
for appliances at any particular tariff band to within 40% of the historical values.
As the freedom is reduced, solutions at high savings targets become unavailable,
and eventually when the scope for adjustment is very low (low freedom value)
no solutions can be found at any savings target, as shown in Figure 6. In these
results, we observe that the imposition of user-specifiable restrictions on scope
for adjustment to appliance use has little-to-no impact on the computation time
(all solutions found took between 900 and 1150ms to discover), with solutions
unavailable where the limits on value adjustment imposed by preferences prevent
the discovery of solutions in the range of the target savings amount. Where no
solution is available, the solver takes a consistent 780ms to determine this.

6.2 Scheduler Performance

To evaluate the performance of the scheduling module, we took the output of the
activations saving module (without preferences) as the basis for scheduling the
next week of appliance activations. In this evaluation, we schedule using a range
of time-slot resolution values, from 5 minutes up to 60 minutes, and observe
the computation time. As the users can introduce custom constraints to the
scheduler, we re-perform the scheduling preocedure at each time-slot resolution
level with the addition of an Inter-Appliance Disjunction constraint between two
appliances (two particular appliances cannot operate simultaneously), and the
further addition of a temporal link between two appliances (wherever one appli-
ance ends, a particular other appliance should start shortly afterwards). The
results of these experiments are shown in Figure 7. We observe that the perfor-
mance with the addition of user-specified constraints remains approximately the
same as without (“Default”), and that while computation time for small time-
slot resolution is extreme (up to 400 seconds), with timeslots of 20 minutes or
larger the computation time is much more reasonable (from around 30 seconds
computation time at 20-minute slots down to 5 seconds at 60 minute slots).
 Design and Evaluation of a Constraint-Based Energy Saving 699

Fig. 6. Savings module performance under restricted scope

Fig. 7. Scheduler Results

700 S.Ó. Murphy et al.

7 Related Work
A number of examples of work in optimisation of energy use to balance demand
and reduce costs exist intended to operate at the supply side. These approaches
are typically intended to integrate with “smart” appliances in the home to allow
for load balancing across the so-called “smart grid”. As these approaches typi-
cally operate centrally at the utility, Mixed Integer Linear Programming (MILP)
is commonly used to compute power use schedules based on energy availability
[4,5,13,14,22,23]. MILP approaches take some time to compute (several min-
utes generally), and as they are solving over several homes and with respect to
the utility company’s energy supply requirements, there is limited scope for the
home-owner to understand their personal energy use.
While not applicable to operating on tablets, Lim et al [17] describes a mixed-
integer programming approach to air conditioning scheduling based on scheduled
meetings in buildings, demonstrating the integration of energy-based require-
ments with real-world user requirements in scheduling. Scott et al [21] present
a MILP-based framework featuring the concept of “comfort” in addition to the
consumption and demand considerations that we also consider.
Felfernig & Burke[12] investigated the position of constraint-based models
as the basis of Recommender Systems, noting that limited attention or under-
standing of complex recommender systems on the part of the user is a major
consideration in the design of recommender systems. In our work, we use a
simple-to-understand tablet-based interface to allow users to enter a dialogue
with the system, introducing requirements which are interpreted as constraints,
but without requiring particular expertise on the part of the user. Like our work,
Ha et al [15] proposed a constraint-based approach for demand-side appliance
scheduling, though this work was demonstrated using randomly generate syn-
thetic data and compution time was in the order of minutes.
He et al [16] consider the case where electricity pricing is highly variable
in time and the active use of energy by homes impacts the pricing, a potential
future direction for energy production. They developed a market model combined
with optimisation of community electricity use (using both constraint program-
ming and mixed integer programming) for demand management and cost sav-
ing. Tushar et al [26] described a game-theory-based approach to demand-side
energy management, modeling the interaction between home-owners and utilities
through the trade and consumption of variable energy supplies and costs.
Darby et al [11] and Carroll [9] note that end-user perception has a significant
role to play in motivating changes in energy use behaviours, with greater engage-
ment and savings achieved simply through making more information available to
the homeowner as found in smart-metering, from variable tariff trials in Ireland
[10]. Bu et al [7] presented an MILP-based adaptive home energy scheduling app-
roach with an initial discretisation step similar to the role of data pre-processing
model in our work, though using synthetic test data rather than real-world data
streams, and with relatively high computation time and resources compared to
our approach. De Sá Ferreira et al [20] investigate highly variable tariff design
at the utility level, generating optimised tariff pricing on a day-to-day basis.
 Design and Evaluation of a Constraint-Based Energy Saving 701

Highly variable tariffs generated through such a scheme could function well com-
bined with the energy saving and scheduling described in our work, allowing for
dynamic schedule generation on a daily basis while respecting the home-owner’s
typical behaviour patterns.

8 Conclusions and Future Work

We presented a constraint-based energy saving recommender system which
uses real-world appliance energy use data to generate recommended behaviour
changes to achieve energy saving goals in the home. The system is computation-
ally efficient and operates on a low powered tablet. Integrating user preferences
allows for a continuous, interactive optimisation process with users introducing
additional requirements as constraints in the problem, and through periodic re-
calculation of solutions when the sensor-observed energy use behaviour differs
from the suggested adjustments. This aspect allows for users to interact with the
model and solutions without any special knowledge, and could lead to greater
understanding on the part of the home-owner as to their energy use habits and
their impact on costs and emissions.
In future we will expand the deployments and models, incorporating add
room and water heating sensing, and the capacity to remotely actuate appli-
ances attached to smart-plugs. Remote actuation would allow for the optimised
schedules to be automatically implemented in the home without requiring the
users to physically attend the devices, and would substantially ease the burden
on the occupant to conform to the scheduled plans. As the AUTHENTIC project
expands, further home deployments and long-term feedback from users will moti-
vate expanding the range of user-specified constraints and further investigation
into the performance of the scheduler in these scenarios. We will expand on the
peak-load model for cases where temporally-restricted renewable energy or home-
energy-storage is a feature. In the case that a home has a large energy storage
solution (e.g. an electric car battery), a limited capacity of cheap energy could
be stored overnight and used to power appliances the next day during higher
tarriff times. Similarly, solar panels could augment the energy supply depending
on the weather, and this could motivate opportunistic appliance scheduling in
reaction to volatile cheap energy availability.

Acknowledgments. This paper is funded by Enterprise Ireland through grant

TC20121002D. We are grateful to the IERC and our academic and industry part-
ners in Authentic for the design and development of the Authentic HAN, and to the
anonymous occupants of our pilot homes for trialling the Authentic HAN, providing
their data, and testing the Reporting Tool and Recommender System.

References
1. Application of delivery prices, orion group energy. http://www.oriongroup.co.nz/
downloads/ApplicationOfDeliveryPrices.pdf (issued February 2, 2015, accessed
April 21, 2015)
702 S.Ó. Murphy et al.

2. Authentic restful api. http://authentic.ucd.ie/documentation (accessed: December

1, 2014)
3. Aggoun, A., Beldiceanu, N.: Extending chip in order to solve complex scheduling
and placement problems. Mathematical and Computer Modelling 17(7), 57–73
(1993)
4. Agnetis, A., de Pascale, G., Detti, P., Vicino, A.: Load scheduling for house-
hold energy consumption optimization. IEEE Transactions on Smart Grid 4(4),
2364–2373 (2013)
5. Bajada, J., Fox, M., Long, D.: Load modelling and simulation of household electric-
ity consumption for the evaluation of demand-side management strategies. In: 2013
4th IEEE/PES Innovative Smart Grid Technologies Europe (ISGT EUROPE),
pp. 1–5. IEEE (2013)
6. Bray, T.: The javascript object notation (json) data interchange format (2014)
7. Bu, H., Nygard, K.: Adaptive scheduling of smart home appliances using fuzzy
goal programming. In: The Sixth International Conference on Adaptive and Self-
Adaptive Systems and Applications, ADAPTIVE 2014, pp. 129–135 (2014)
8. Carlier, J.: The one-machine sequencing problem. European Journal of Operational
Research 11(1), 42–47 (1982)
9. Carroll, J., Lyons, S., Denny, E.: Reducing household electricity demand through
smart metering: The role of improved information about energy saving. Energy
Economics 45, 234–243 (2014)
10. CER11080a: Electricity smart metering customer behaviour trials (cbt findings
report) (2011)
11. Darby, S.J., McKenna, E.: Social implications of residential demand response in
cool temperate climates. Energy Policy 49, 759–769 (2012)
12. Felfernig, A., Burke, R.: Constraint-based recommender systems: technologies and
research issues. In: Proceedings of the 10th International Conference on Electronic
Commerce, p. 3. ACM (2008)
13. Georgievski, I.: Planning for coordination of devices in energy-smart environments.
In: Proceedings of the Twenty-Third International Conference on Automated Plan-
ning and Scheduling (ICAPS 2013) (2013)
14. Ha, D.L., Joumaa, H., Ploix, S., Jacomino, M.: An optimal approach for electrical
management problem in dwellings. Energy and Buildings 45, 1–14 (2012)
15. Ha, L.D., Ploix, S., Zamai, E., Jacomino, M.: Tabu search for the optimization of
household energy consumption. In: 2006 IEEE International Conference on Infor-
mation Reuse and Integration, pp. 86–92. IEEE (2006)
16. He, S., Liebman, A., Rendl, A., Wallace, M., Wilson, C.: Modelling rtp-based resi-
dential load scheduling for demand response in smart grids. In: International Work-
shop on Constraint Modelling and Reformulation (Carlos Ansotegui 08 September
2014 to 08 September 2014), pp. 36–51 (2014)
17. Lim, B.P., van den Briel, M., Thiébaux, S., Bent, R., Backhaus, S.: Large Neighbor-
hood Search for Energy Aware Meeting Scheduling in Smart Buildings. In: Michel,
L. (ed.) CPAIOR 2015. LNCS, vol. 9075, pp. 240–254. Springer, Heidelberg (2015)
18. O’Sullivan, T., Muldoon, C., Xu, L., O’Grady, M., O’Hare, G.M.: Deployment of
an autonomic home energy management system. In: The 18th IEEE International
Symposium on Consumer Electronics (ISCE 2014), pp. 1–2. IEEE (2014)
19. Richardson, L., Ruby, S.: RESTful web services. O’Reilly Media, Inc. (2008)
20. de Sá Ferreira, R., Barroso, L.A., Rochinha Lino, P., Carvalho, M.M., Valenzuela,
P.: Time-of-use tariff design under uncertainty in price-elasticities of electricity
demand: A stochastic optimization approach. IEEE Transactions on Smart Grid
4(4), 2285–2295 (2013)
 Design and Evaluation of a Constraint-Based Energy Saving 703

21. Scott, P., Thiébaux, S., van den Briel, M., Van Hentenryck, P.: Residential Demand
Response under Uncertainty. In: Schulte, C. (ed.) CP 2013. LNCS, vol. 8124,
pp. 645–660. Springer, Heidelberg (2013)
22. Sou, K.C., Kordel, M., Wu, J., Sandberg, H., Johansson, K.H.: Energy and co 2 effi-
cient scheduling of smart home appliances. In: 2013 European Control Conference
(ECC), pp. 4051–4058. IEEE (2013)
23. Sou, K.C., Weimer, J., Sandberg, H., Johansson, K.H.: Scheduling smart home
appliances using mixed integer linear programming. In: 2011 50th IEEE Con-
ference on Decision and Control and European Control Conference (CDC-ECC),
pp. 5144–5149. IEEE (2011)
24. Sweeney, J.C., Kresling, J., Webb, D., Soutar, G.N., Mazzarol, T.: Energy saving
behaviours: Development of a practice-based model. Energy Policy 61, 371–381
(2013)
25. Torriti, J.: Price-based demand side management: Assessing the impacts of time-
of-use tariffs on residential electricity demand and peak shifting in northern italy.
Energy 44(1), 576–583 (2012)
26. Tushar, W., Zhang, J.A., Smith, D.B., Thiebaux, S., Poor, H.V.: Prioritizing con-
sumers in smart grid: Energy management using game theory. In: 2013 IEEE Inter-
national Conference on Communications (ICC), pp. 4239–4243. IEEE (2013)
 Scheduling Running Modes of Satellite
Instruments Using Constraint-Based
Local Search

Cédric Pralet1(B) , Solange Lemai-Chenevier2 , and Jean Jaubert2

1
ONERA – The French Aerospace Lab, F-31055 Toulouse, France
cedric.pralet@onera.fr
2
CNES Toulouse, Toulouse, France
{solange.lemai-chenevier,jean.jaubert}@cnes.fr

Abstract. In this paper, we consider a problem involved in the man-

agement of Earth observation satellites. The input of the problem is a
time-stamped observation plan and the output is a schedule of transi-
tions between running modes of instruments available on board. These
transitions must be scheduled so as to effectively realize the acquisitions
requested, while satisfying several constraints, including duration con-
straints, non-overlapping constraints, and resource constraints over the
thermal consumption of instruments. Criteria such as the minimization
of the number of times instruments are switched off are also considered,
for long-term reliability issues. The scheduling problem obtained needs
to be solved several times per day, and the requirement is to produce
good quality solutions in a few seconds. To produce such solutions, we
propose a specific constraint-based local search procedure. Experiments
are performed on realistic scenarios involving hundreds of observations,
and the approach is compared with other techniques.

1 Problem Description

The study described in this paper is performed in relation with a future space
program involving an observation satellite orbiting around the Earth at a low-
altitude orbit (several hundreds of kilometers). The task of this satellite is to
realize acquisitions over specific Earth areas using three optical instruments I1,
I2, I3 available on board, and to download collected data towards ground recep-
tion stations. For such a satellite, each orbit takes about one hour and a half, and
different areas are overflied during successive orbits due to the rotation of the
Earth on itself. The set of areas selected for observation is defined by a ground
mission center, which produces so-called acquisition plans for the satellite. Each
acquisition in these plans is defined by its start and end times, by a set of instru-
ments to be used, and by instrument configurations, which correspond to specific
settings regarding instrument signal processing.
In addition to acquisition plans, the ground mission center must also pro-
vide the satellite with low-level plans specifying in more detail the operations

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 704–719, 2015.
DOI: 10.1007/978-3-319-23219-5 48
 Scheduling Running Modes of Satellite Instruments 705

to execute in order to make instruments ready for acquisition when requested.

These low-level plans must define, for each instrument, the successive transitions
to make between several possible running modes: the (O)ff mode, the (S)tandby
mode, the (H)eat mode, and the (A)cquisition mode. The possible transitions
between these modes are described in Fig. 1a. To effectively realize an acquisition
a over time interval [t1 , t2 ], each instrument required for a must be in the acqui-
sition mode over [t1 , t2 ]. For this, low-level plans must specify when instruments
must perform a transition from the off mode to the standby mode, when the
right instrument configuration must be loaded, when a transition to the heat
mode must be performed to warm up the instrument, and when a transition
to the acquisition mode must be triggered to have the instrument ready at t1 .
Between two acquisitions, the instrument may come back to the heat, standby,
or off mode depending on the time and resources available. Each transition will
then result in a sequence of even more basic commands executed on board.
In this paper, we consider this problem of scheduling on the ground the tran-
sitions between instrument modes, from a time-stamped acquisition plan given
as an input. In this problem, several kinds of constraints must be taken into
account, including (1) duration constraints, enforcing a minimum duration spent
in each mode and a fixed duration for each mode transition; (2) non-overlapping
constraints over transitions; more precisely, transitions over instruments I1 and
I2 cannot overlap except when they correspond to the same transition; the reason
for this is that instruments I1 and I2 are managed on board by the same moni-
tor, which is able to send a common sequence of basic commands to both I1 and

(a) chgConf chgConf

OFF SBY HEAT ACQ

(b) I1 O S S H A H A H A H S H A H S O

I2 O S S H A H S H A H A H S S H A H A H S O

I3 O S S H A H H A H S H A H S O

acqs a1 a2 a3 a4 a5 a6

orbit 1 orbit 2

Fig. 1. (a) Mode automaton associated with each instrument; chgConf stands for a
change of the configuration of the instrument; (b) example of a schedule defining mode
transitions for instruments I1, I2, I3, given an acquisition plan involving 6 acquisitions
a1 to a6 over 2 orbits; transitions correspond to empty rectangles between modes; tran-
sitions between the same two modes correspond to changes in instrument configuration;
an acquisition such as a1 uses all instruments, while a4 uses only I2.
706 C. Pralet et al.

I2, but not two distinct command sequences in parallel; similarly, for instrument
monitoring reasons, it is not possible to perform a transition between off and
standby over I3 in parallel to a transition between off and standby over I1 or I2;
moreover, no transition must be performed during acquisitions to avoid distur-
bances; (3) resource constraints, imposing that the total thermal consumption
induced by the activities performed on each instrument over each orbit cannot
exceed a maximum limit. In terms of optimization criterion, the main objec-
tive is to limit the number of times instruments are switched off, for long-term
reliability issues. For a given number of times instruments are switched off, a
secondary objective is to minimize the total thermal consumption of the instru-
ments (better resource usage). In the end, the goal is to produce schedules such
as the one provided in Fig. 1b.
This mode transition scheduling activity is the last step of the complete mis-
sion programming process which may be done several times a day, depending on
the station opportunities to upload a new plan to the satellite. It is required to
be solved in less than a few seconds. Constraint Programming (CP) comes into
play because the realization of consistent mode transition schedules becomes
more and more challenging for space engineers, due to the increasing complex-
ity of satellites. One expected benefit in using CP is a better exploration of
the search space, at least better than with hard-coded decision rules. Another
advantage is that in first design phases as the one we are in, the declarative
nature of CP is more flexible in case of changes in the satellite specifications,
and it allows to easily assess the impact of new design choices.
The paper is organized as follows. We first give a formal model of the prob-
lem (Sect. 2). We then present a specific constraint-based local search encoding
(Sect. 3) and a dedicated local search procedure (Sect. 4). Finally, we describe
experiments on realistic scenarios involving hundreds of acquisitions (Sect. 5).

2 Problem Modeling

Data. The set of possible running modes of the instruments is M =

{O, S , H , A}: O for Off, S for Standby, H for Heat, A for Acqui-
sition. The set of possible transitions between these modes is T =
{OS , SO, SH , HS , HA, AH , SS , HH }. In this set, mm stands for a transition
from mode m to mode m , transitions mm being associated with instrument
configuration loading. We also define as S = M ∪ T the set of possible states of
each instrument at each time. We then introduce several input data:

– a set of contiguous orbits O; each orbit o ∈ O has a start time TsOrbo and
an end time TeOrbo ; the planning horizon considered is [Ts, Te] with Ts
the start time of the first orbit and Te the end time of the last orbit;
– a set of instruments I; for each instrument i ∈ I and each mode m ∈ M,
DuMini,m denotes a minimum duration for i in mode m, and for each
transition of type t ∈ T , Dui,t denotes the fixed duration of transition t
for instrument i; a thermal consumption rate ThRatei,s is associated with
 Scheduling Running Modes of Satellite Instruments 707

each state s ∈ S of instrument i; last, ThMaxi denotes a maximal thermal

consumption allowed for instrument i over each orbit;
– a set of acquisitions A; each acquisition a ∈ A has a fixed start time TsAcqa
and a fixed duration DuAcqa ; for each instrument i ∈ I, Conf a,i denotes
the configuration required by acquisition a on instrument i (value nil when
instrument i is not used in the realization of a); without loss of generality,
we assume that each instrument is used at least once in the acquisition plan;
– a set of incompatibilities Inc between mode transitions; each element in Inc
is a 4-tuple (i, i , t, t ) ∈ I 2 × T 2 specifying that transition t on instrument i
cannot be performed in parallel to transition t on instrument i ; for instance,
(I1, I2, SH , OS ) ∈ Inc because transition SH over I1 cannot be performed
in parallel to transition OS over I2.

For each instrument i ∈ I, several additional data can be derived, such as the
number of acquisitions Nacqsi for which i is used. For every u ∈ [1..Nacqsi ],
Acqi,u ∈ A denotes the acquisition associated with the uth use of instrument i.

Allowed State Sequences. In order to produce plans such as the one given in
Fig. 1b, it is necessary to determine, for each instrument, the sequence of its con-
tiguous states over planning horizon [Ts, Te], as well as the exact start and end
times associated with these states. As the desired state sequences must contain
the acquisitions, the main question is actually how to handle each instrument
outside these acquisitions, that is (1) from the beginning of the planning horizon,
where instruments are switched off, to the first acquisition, (2) between two suc-
cessive acquisitions which use the same instrument configuration, (3) between
two successive acquisitions which use distinct configurations, and (4) from the
last acquisition to the end of the horizon, where instruments must all be switched
off for safety reasons. These four types of global transitions between successive
uses of instruments must be composed of state sequences that are consistent
with the automaton given in Fig. 1a.
To limit the search space, one key point is to avoid considering state sequences
such as [A,AH,H,HS,S,SH,H,HS,S,SH,H,HA,A], in which the instrument is
warmed up and cooled down (part SH,H,HS) in the middle of two acquisitions.
To do this, we define beforehand a set of allowed state sequences for each global
transition between instrument uses. The decompositions allowed are provided in
Fig. 2. It can be seen that from the beginning of the horizon to the first acqui-
sition, there are only two relevant state sequences, performing the configuration
loading operation in the standby mode and heat mode respectively. For global
transitions between acquisitions which use the same instrument configuration,
there are four relevant decompositions: one which comes back to the heat mode
between the acquisitions, one which comes back to the standby mode, and two
which come back to the off mode and differ in the place where the configuration
is loaded (the loaded configuration is lost when using the off mode). Similarly,
there are six possible decompositions for global transitions between acquisitions
which use distinct configurations. For the last global transition, there is a unique
possible decomposition.
708 C. Pralet et al.

conf 0 conf 0 conf 1

TransInit HA A AH TransSameConf HA A AH TransChangeConf HA A AH TransEnd
decomp decomp decomp decomp

O OS S SS S SH H H H HH H H HS S SO O

O OS S SH H HH H H HS S SH H H HH H HS S SH H

H HS S SO O OS S SS S SH H H HS S SS S SH H

H HS S SO O OS S SH H HH H H HS S SH H HH H

H HS S SO O OS S SS S SH H

H HS S SO O OS S SH H HH H

Fig. 2. Global transitions between instrument uses and their possible decompositions

Decision Variables. In the following, we define a state interval as a temporal

interval over which a unique state is active. Formally, a state interval itv is
defined by a boolean presence variable pres(itv), by an integer start time variable
start(itv), by an integer end time variable end (itv), by a fixed state S(itv) ∈ S,
and by a fixed instrument I(itv) over which it is placed. In case of acquisition
intervals, I(itv) contains all instruments required for the acquisition. We consider
integer dates because the equipments which send mode transition commands
to the instruments work with a discrete time clock. We also consider global
transition intervals gitv, which are defined by an integer start time variable
start(gitv), by an integer end time variable end (gitv), and by a fixed type among
{TransInit, TransSameConf , TransChangeConf , TransEnd }.
Then, to describe schedules at a global level (upper part in Fig. 2), we intro-
duce the following elements:

– for each acquisition a ∈ A, one state interval itvAcqa of state A is intro-

duced for representing the temporal interval during which a is realized;
– for each instrument i and each use u ∈ [1..Nacqsi ] of this instrument, two
mandatory state intervals itvHAi,u and itvAHi,u of respective states HA
and AH are introduced for representing the transitions between the heat
mode and the acquisition mode just before and just after acquisition Acqi,u ;
– for each instrument i and each use u ∈ [1..Nacqsi +1] of this instrument, one
global transition interval itvTransi,u is defined for representing the global
transition performed before the uth use of the instrument (use Nacqsi + 1
corresponds to the last global transition before the end of the horizon);
– for each instrument i, if [a1 , . . . , an ] denotes the sequence of acquisitions
performed by i, the sequence of intervals associated with i is defined by:
Seqi : [itvTransi,1 , itvHAi,1 , itvAcqa1 , itvAHi,1 ,
...
itvTransi,n , itvHAi,n , itvAcqan , itvAHi,n , itvTransi,n+1 ].
 Scheduling Running Modes of Satellite Instruments 709

To describe schedules at a finer level (lower part in Fig. 2), we introduce the
following elements:

– for each instrument i and each use u ∈ [1..Nacqsi + 1] of i, one integer deci-
sion variable decompi,u ∈ [1..Ndeci,u ] is introduced for representing the
index of the decomposition chosen for realizing global transition itvTransi,u ,
among those given in Fig. 2; parameter Ndeci,u stands for the number of
possible decompositions of itvTransi,u ;
– for each instrument i, for each use u ∈ [1..Nacqsi + 1] of i, and for each
decomposition d ∈ [1..Ndeci,u ] of itvTransi,u , one sequence of state inter-
vals Deci,u,d = [itv1 , . . . , itvk ] is introduced; the latter is determined follow-
ing the decomposition schemes provided in Fig. 2; state intervals in sequence
Deci,u,d are optional (they will be present only in case decompi,u = d).

The set of state intervals defined in the model is then Itv = {itvAcqa | a ∈
A} ∪ {itvHAi,u | i ∈ I, u ∈ [1..Nacqsi ]} ∪ {itvAHi,u | i ∈ I, u ∈ [1..Nacqsi ]} ∪
{itv ∈ Deci,u,d | i ∈ I, u ∈ [1..Nacqsi +1], d ∈ [1..Ndeci,u ]}. Each interval itv in
this set has a minimum start time Min(itv) and a maximum end time Max (itv),
both obtained from the known dates of acquisitions and from the minimum
durations of state intervals. The set of intervals in Itv which may overlap orbit
o and which concern instrument i is denoted by ItvOrb i,o .

Constraints. We now express the set of constraints holding over the integer
variables, intervals, and sequences of intervals previously introduced. These con-
straints are given in Eq. 1 to 15. Each constraint introduced is active only when
the intervals over which it holds are present. Constraint 1 expresses that all
intervals involved in sequence Seqi must be present. Constraints 2-3 express
that the start and end times of this sequence must coincide with the start and
end times of the planning horizon. Constraint 4 imposes that successive inter-
vals belonging to sequence Seqi must be contiguous (no period during which
the state of the instrument is undefined). Constraint 5 expresses that for a given
global transition, only state intervals involved in the decomposition chosen for
this transition are present. Constraints 6-7 impose that the start and end times
of the sequence of intervals associated with a decomposition must coincide with
the start and end times of the global transition from which it is derived. Con-
straint 8 specifies that successive intervals belonging to sequence Deci,u,d must
be contiguous (no period during which the state of the instrument is undefined).
Constraints 9-10 define a minimum duration for state intervals labeled by a mode
and a fixed duration for state intervals labeled by a transition. Constraints 11-
12 express that acquisition intervals must start at the required start time and
have the required duration. Constraint 13 imposes that no transition must occur
during acquisitions. Constraint 14 expresses that incompatible transitions must
not overlap. Constraint 15 imposes a maximum thermal consumption over each
orbit of the satellite for each instrument. In this constraint, the left term of the
710 C. Pralet et al.

inequality takes into account only the part of intervals which overlaps the orbit
considered.

∀i ∈ I,
∀itv ∈ Seqi , pres(itv) = 1 (1)
start(f irst(Seqi )) = Ts (2)
end (last(Seqi )) = Te (3)
∀itv ∈ Seqi | itv = last(Seqi ), end (itv) = start(next(itv, Seqi )) (4)
∀i ∈ I, ∀u ∈ [1..Nacqsi + 1], ∀d ∈ [1..Ndeci,u ],
∀itv ∈ Deci,u,d , pres(itv) = (decompi,u = d) (5)
start(f irst(Deci,u,d )) = start(itvTransi,u ) (6)
end (last(Deci,u,d )) = end (itvTransi,u ) (7)
∀itv ∈ Deci,u,d | itv = last(Deci,u,d ), end (itv ) = start(next(itv , Deci,u,d ))(8)
∀itv ∈ Itv s.t. S(itv) ∈ M, duration(itv) ≥ DuMinI(itv),S(itv) (9)
∀itv ∈ Itv s.t. S(itv) ∈ T , duration(itv) = DuI(itv),S(itv) (10)
∀a ∈ A,
start(itvAcqa ) = TsAcqa (11)
duration(itvAcqa ) = DuAcqa (12)
∀itv ∈ Itv | S(itv) ∈ T , noOverlap(itv, itvAcqa ) (13)
∀itv, itv  ∈ Itv s.t. (I(itv), I(itv  ), S(itv), S(itv  )) ∈ Inc, noOverlap(itv, itv  ) (14)

∀i ∈ I, ∀o ∈ O, pres(itv) · d(itv) · ThRatei,S(itv) ≤ ThMaxi (15)
itv∈ItvOrb i,o

with d(itv) = 0 if (end (itv) ≤ TsOrbo ) ∨ (start(itv) ≥ TeOrbo )

min(end (itv), TeOrbo ) − max(start(itv), TsOrbo )) otherwise

Optimization Criteria. The first objective is to minimize the number of times

instruments are switched off, for long-term reliability issues:

minimize card{itv ∈ Itv | (S(itv) = O) ∧ (pres(itv) = 1)} (16)

The second (and less important) criterion is to minimize the total thermal
consumption over all instruments:

minimize pres(itv) · (end (itv) − start(itv)) · ThRateI(itv),S(itv) (17)
itv∈Itv

The two criteria defined are antagonistic because for instance the addition
of off periods into the plan reduces the total thermal consumption. Criteria
imposing a fair sharing between instruments could also be considered, as well as
a minimization of the number of off intervals for a particular instrument.

Problem Analysis. The decision problem obtained is a constraint-based schedul-

ing problem involving temporal constraints, resource constraints, and optional
 Scheduling Running Modes of Satellite Instruments 711

tasks [1]. It can be related with standard Job Shop Scheduling Problems
(JSSP [2]). In our case, there would be one job per instrument, and the (optional)
operations associated with each job would be the set of state intervals associ-
ated with each instrument. However, there are several differences with JSSP. For
instance, some operations are shared between jobs, namely the acquisition tasks
that require using several instruments simultaneously. Also, successive opera-
tions must be contiguous. Contiguity is important for ensuring that the state of
the instrument is defined at every time and for getting the right thermal con-
sumption. This consumption depends on the duration of state intervals, which
is not fixed for state intervals associated with modes.
Another significant feature of the problem is the presence of global transi-
tion tasks which can be decomposed in several possible ways. Such a hierarchical
aspect is also found in the framework of Hierarchical Task Networks (HTN [3])
developed in the planning community. In this framework, some abstract tasks
are specified and the goal is to decompose these tasks into basic sequences of
actions, through a set of available decomposition methods. In CP, the possible
decompositions of global transitions could be expressed using an alternative con-
straint [4]. In another direction, the set of allowed sequences of state intervals
could be represented using a regular constraint [5].
Last, the criteria considered are not standard criteria such as the makespan
or the tardiness. On this point, it is worth noting that some transitions should
be scheduled as early as possible to reduce the resource consumption, whereas
others should be scheduled as late as possible for the same reason (more details
on these points later in the paper).

3 Specific Problem Encoding

To obtain an efficient approach, several difficulties must be overcome. We present

these difficulties and we show how they are handled.

3.1 Problem Complexity Versus Allowed Running Times

The problem obtained could be modeled and solved using CP solvers such as
IBM ILOG CpOptimizer, which contains efficient scheduling techniques. It could
also be represented using Mixed Integer Programming [6], which would be par-
ticularly adapted to deal with linear expressions related to thermal aspects.
However, we must be able to solve in a few seconds large instances, which can
involve several hundreds of acquisitions generating several thousands or tens of
thousands candidate state intervals.
To overcome this difficulty, we use Constraint-Based Local Search (CBLS [7]),
which has the capacity to quickly produce good quality solutions for large-size
problems. As in standard constraint programming, CBLS models are defined
by decision variables, constraints, and criteria. One distinctive feature is that
in CBLS, all decision variables are assigned when searching for a solution, i.e.
the approach manipulates complete variable assignments. The search space is
712 C. Pralet et al.

then explored by performing local moves which reassign some decision variables,
and it is explored more freely than in tree search with backtracking. One speci-
ficity of CBLS models is that they manipulate so-called invariants. The latter
are one-way constraints x ← exp where x is a variable (or a set of variables)
and exp is a functional expression of other variables of the problem, such as
x ← sum(i ∈ [1..N ]) yi . During local moves, these invariants are efficiently main-
tained thanks to specific procedures that incrementally reevaluate the output of
invariants (left part) in case of changes in their inputs (right part).
For tackling the application, we use the InCELL CBLS engine [8]. The lat-
ter offers several generic building blocks, including a large catalog of CBLS
invariants together with incremental reevaluation techniques, and generic ele-
ments for defining specific local search procedures. In InCELL, a temporal
interval itv is represented by a presence variable pres(itv), a start time-point
start(itv), and an end time-point end (itv). For incrementally managing tem-
poral constraints over intervals, InCELL uses the Simple Temporal Network
invariant [9]. The latter takes as input an STN [10], which is a set of simple
temporal constraints over time-points, and it maintains as an output the tem-
poral consistency of this STN as well as, for each time-point t, two variables
earliestTime(t) and latestTime(t) giving its earliest and latest temporal position
in a consistent schedule respectively. In the following, given an interval itv, vari-
ables earliestTime(start(itv)), latestTime(start(itv)), earliestTime(end (itv)),
and latestTime(end (itv)) are denoted more succinctly as est(itv), lst(itv),
eet(itv), and let(itv).

3.2 Specific Manipulation of Topological Orders

To implement a specific local search procedure, we need to control the main

decisions involved in the problem. These decisions first include the decompo-
sitions chosen for global transitions, which are directly represented by vari-
ables decompi,u . They also include the ordering between state intervals which
must not overlap. To represent such decisions, we explicitly manipulate in the
CBLS model a topological ordering topoOrder over intervals involved in non-
overlapping constraints (Constraints 13-14). Technically speaking, this topo-
logical ordering is a sequence which expresses that if two intervals itv, itv 
must not overlap and if itv is placed before itv  in the sequence, denoted by
before(itv , itv  , topoOrder) = 1, then end (itv) ≤ start(itv  ) must hold. Con-
straints noOverlap(itv, itv  ) appearing in Eq. 13-14 are then reformulated as:

(before(itv, itv  , topoOrder) = 1) → (end (itv) ≤ start(itv  )) (18)

(before(itv  , itv, topoOrder) = 1) → (end (itv  ) ≤ start(itv)) (19)

In the local search algorithm defined, some of the local moves are performed
by updating directly the topological ordering of conflicting intervals. To obtain
an efficient approach, it is crucial to avoid considering local moves which cre-
ate precedence cycles between intervals, where for instance an interval itv1 is
 Scheduling Running Modes of Satellite Instruments 713

requested to be placed both before and after another interval itv2 . For avoid-
ing this, as in Job Shop Scheduling, we exploit a mandatory precedence graph.
The latter captures all mandatory precedences over transitions and acquisitions,
including: (1) precedences between successive transition intervals involved in
the same sequence Deci,u,d (sequence associated with the dth possible decom-
position of the global transition realized before the uth use of instrument i),
and (2) precedences between the first (resp. last) transition intervals in Deci,u,d
and the acquisition that precedes (resp. follows) Deci,u,d . From this precedence
graph, when a transition interval itv must be moved just after another inter-
val itv  in topological order topoOrder, all mandatory successors of itv placed
before itv  in the order are also moved just after itv, to avoid the creation of a
precedence cycle.
Also, for each instrument i ∈ I and each use of this instrument u ∈
[1..Nacqsi ], we merge intervals itvHAi,u and itvAcqAcqi,u in the topological
order. Indeed, as transitions and acquisitions cannot overlap, a transition interval
itv which is in conflict with itvHAi,u either ends before itvHAi,u , or starts after
itvAcqAcqi,u , hence there is no need to dissociate itvHAi,u and itvAcqAcqi,u
in the order. For the same reason, intervals itvAHi,u and itvAcqAcqi,u are also
merged.

3.3 Specific Encoding of the Thermal Resource Constraint

Once a decomposition is selected for each global transition (decompi,u vari-

ables) and once an ordering between conflicting intervals is chosen (topoOrder
variable), the temporal consistency of the plan can be determined as well as
the value of the first criterion (number of times instruments are switched off).
The only remaining freedom concerns the choice of precise start and end times
of all intervals in the model. There is usually a huge freedom in the choice of
these precise times, because we must deal with a large possible set of time steps.
To avoid loosing time exploring all precise time values and to detect thermal
inconsistency earlier, there is a need to build an approximation of the impact of
decisions made so far on the thermal resource constraint given in Eq. 15 and on
the thermal consumption criterion given in Eq. 17.
To do this, we again use a specificity of the problem which is that in each
decomposition Deci,u,d , it is possible to identify one particular interval whose
thermal consumption is the lowest, and therefore one particular interval which
should be enlarged as much as possible. In Fig. 2, for each possible decom-
position, this particular interval to enlarge is represented with an underlined
label. From this observation, we build an approximation of the thermal con-
straint (Constraint 15) as follows: in Eq. 15, (1) if itv is an interval to enlarge,
start(itv) and end (itv) are replaced by est(itv) and let(itv) respectively (start
as early as possible, end as late as possible); (2) if itv is placed on the left of the
interval to enlarge in a decomposition, then start(itv) and end (itv) are replaced
by est(itv) and eet(itv) respectively (start and end as early as possible); (3) if
itv is placed on the right of the interval to enlarge in a decomposition, then
714 C. Pralet et al.

start(itv) and end (itv) are replaced by lst(itv) and let(itv) respectively (start
and end as late as possible). The schedule obtained with these dates may not be
feasible since it mixes earliest dates and latest dates, but the thermal consump-
tion obtained over each orbit for each instrument can be shown to be a lower
bound on the real consumption.1 The same rewriting is used for building a lower
approximation of the thermal criterion (Eq. 17).

3.4 Specific Management of Some Hard Constraints

As almost always in local search, there is a choice between handling constraints

as hard constraints, to be always satisfied by the current assignment, or as con-
straints returning a constraint violation degree. Here, we choose to make soft the
constraints enforcing due dates (Constraints 3 and 11) and the thermal resource
constraint (Constraint 15), and we keep the other constraints as hard. Such a
choice is made because finding a first plan satisfying Constraints 3, 11, 15 is
not straightforward, therefore we need to temporarily explore plans which vio-
late these constraints to progressively get a first consistent solution. On the other
side, for all other constraints, e.g. for all constraints enforcing contiguity between
successive activities, it is easy to produce a first consistent schedule.
More precisely, we transform Constraint 11 start(itvAcqa ) = TsAcqa into
one permanent constraint (Eq. 20), one constraint whose presence can be con-
trolled based on an additional variable named locka (Eq. 21), and one constraint
violation degree called delaya , defined as a CBLS invariant (Eq. 22). The latter
represents the gap between the earliest start time of a and its required start
time. Variable locka serves us to explore either consistent schedules or tem-
porarily inconsistent schedules when needed. All lock variables will be set to 1
at the end of the search, meaning that all acquisitions will be realized on time.

start(itvAcqa ) ≥ TsAcqa (20)

(locka = 1) → (start(itvAcqa ) ≤ TsAcqa ) (21)
delaya ← est(itvAcqa ) − TsAcqa (22)

The same transformation is used for Constraint 3, with a lock|A|+1 variable

and with a violation degree delay|A|+1 ← maxi∈I (eet(last(Seqi )) − Te). When
all lock variables are set to 0, provided that there is no precedence cycle, there
always exists a schedule which satisfies all active temporal constraints, essentially
because there is no maximum duration for off, standby, and heat intervals.
Last, given an instrument i and an orbit o, the violation degree thvi,o asso-
ciated with Constraint 15 is the maximum between 0 and the difference between
the thermal consumption obtained and the maximum thermal consumption
allowed:

1
Without going into further details, this result holds under the assumption that the
thermal consumption rate of the SS (resp. HH) state is equal to the thermal con-
sumption of the S (resp. H) state, which is true in practice.
 Scheduling Running Modes of Satellite Instruments 715

Fig. 3. Global view of the CBLS model


thvi,o ← max(0, pres(itv) · d(itv) · ThRatei,S(itv) − ThMaxi ) (23)
itv∈ItvOrb i,o

Fig. 3 gives a global view of the structure of the resulting CBLS model. The
figure does not represent all individual invariants of the model.

4 A Multi-phase Local Search Algorithm

We now describe how we actually solve the problem, using a specific local search
procedure. In the latter, the first step is to perform some preprocessing to restrict
the set of feasible decompositions of global transitions. This step is able to auto-
matically detect that for temporal reasons, the decomposition using the heat
mode is the only feasible one between successive acquisitions placed very close
to each other. It is also able to detect that for thermal reasons, only decomposi-
tions which come back to the off mode are feasible between successive acquisitions
placed very far from each other.
After this preprocessing step, the main difficulty is that the problem involves
antagonistic aspects, because for instance the best option regarding the on/off
criterion is to never come back to the off mode between two successive acquisi-
tions, whereas it is the worst option from a thermal point of view. This is why
the resolution process is divided into several search phases (see Fig. 4).

Phase 1: First Heuristic Schedule. In the first phase, we build an initial schedule
from simple decision rules. For decomposing global transitions, we systematically
choose the option whose total minimum duration is the lowest. For choosing a
topological order between intervals, we use a dispatching rule called the best
date heuristic. In this heuristic, we compute, for each global transition and for
each of its possible decompositions dec, the optimal start time of each transition
716 C. Pralet et al.

Phase1 Phase2 Phase3 Phase4 Phase5

consistent
first temporal thermal thermal backtrack−free
heuristic repair consumption consumption schedule
schedule (acquisition reduction, reduction, extraction
delays) without off with off
Solution
inconsistent

Fig. 4. Multi-phase local search algorithm

interval itv involved in dec when non-overlapping constraints are discarded. For
intervals itv placed after the interval to enlarge in dec, this optimal start time
corresponds to the latest start time, denoted by lstAlone(itv), and for the others
it corresponds to the earliest start time, denoted by estAlone(itv). To get a
first schedule, transitions with the smallest best dates are inserted first into the
topological ordering. During this phase, all lock variables are set to 0.

Phase 2: Temporal Repair. The schedule obtained after Phase 1 may violate
temporal constraints on the realization dates of acquisitions. To repair such con-
straint violations, we consider acquisitions chronologically. When considering an
acquisition a, we try to enforce the realization date of a as long as the constraint
violation degree associated with it is not equal to 0 (Eq. 22). To do this, we build
the critical path explaining the late date obtained for a, we randomly select on
this path a critical arc itv → itv  relating two intervals associated with distinct
instruments, and we try to place itv just after itv  in the topological order. This
local move is accepted iff it strictly reduces the delay of acquisition a. When all
arcs in the critical path have been considered without improving the delay of a,
the algorithm returns an inconsistency, meaning that the acquisition plan is too
tight and should be revised. Otherwise, once the delay of acquisition a is equal to
0, variable locka is set to 1 before considering the next acquisition, meaning that
future moves are not allowed to delay the realization of a. After the traversal of
the acquisition plan, all acquisitions meet their due date. The same principle is
used when the last state interval ends after the end of the planning horizon.

Phase 3: Resource Optimization. During the third phase, we consider the

satisfaction of the thermal consumption constraint, by working on possible
decompositions of global transitions. More specifically, we traverse the schedule
chronologically, from the first acquisition to the last. At each step, we consider
an acquisition a and we lock the realization dates of all acquisitions except for
a (locka = 0). For each instrument involved in the realization of a, we try to
change the decomposition of the global transition leading to a. Only decompo-
sitions which do not use additional switch off operations are considered, and
decompositions which induce lower thermal consumption are considered first.
For assessing the impact of changing the current decomposition used dec for
another decomposition dec , we remove from the topological order all transition
intervals belonging to dec, and we insert into it all transition intervals belonging
to dec . To do this, we use an insertion heuristic which works as follows: the
transition intervals in dec which must be scheduled as early as possible (resp. as
 Scheduling Running Modes of Satellite Instruments 717

late as possible) are inserted from the first to the last (resp. from the last to the
first); the insertion of an interval itv is made just before the first (resp. after the
last) interval itv  in the topological order such that estAlone(itv) < start(itv  )
(resp. lstAlone(itv) > start(itv  )) and such that the insertion does not lead to
an inconsistency. Possible delays over the realization time of the acquisition con-
sidered are also repaired following the mechanism described in Phase 2. If delays
cannot be repaired, decomposition dec is rejected and we go on with the next
possible decomposition. Otherwise, dec is accepted as the new decomposition.

Phase 4: Resource Repair. After the third phase, the schedule may still be incon-
sistent from a thermal point of view. In this case, it is repaired by testing decom-
positions which add new off intervals during orbits where the thermal resource
constraint is violated. To do this, the plan is traversed chronologically, orbit by
orbit, and we work on an orbit until the thermal constraint of the CBLS model is
satisfied. For introducing off intervals, we consider first global transitions which
have the longest duration, because inserting off intervals during these global
transitions is more likely to better reduce the total thermal consumption. For
testing each decomposition, the same mechanism as in Phase 3 is used.

Phase 5: Solution Extraction. In the fifth phase, we use a backtrack-free pro-

cedure for determining good start and end times for intervals. This procedure
traverses the topological ordering and considers at each step the next interval
itv in the order. It sets the start time of itv to its latest consistent value when
itv appears on the right of the interval to enlarge in a decomposition, and to its
earliest consistent value otherwise. After setting the start time of itv, the proce-
dure propagates the effect of this decision in the temporal network before going
on to the next step. The schedule obtained at the end of the traversal of the
topological ordering is temporally consistent, essentially because of the decom-
posability property of Simple Temporal Networks [10]. If the schedule obtained
satisfies thermal constraints over each orbit, it is returned as a solution and the
algorithm stops. Otherwise, we come back to fourth phase to insert additional
off intervals. If no more off interval can be added, an inconsistency is returned.

5 Experiments
Three approaches are compared: (1) the local search algorithm described previ-
ously, built on top of the InCELL CBLS library [8], (2) a MIP approach based on
IBM ILOG CPLEX 12.5; in the MIP model, boolean variables before(itv , itv  ) are
introduced for every pair of intervals itv, itv  which must not overlap; (3) a CP
approach based on IBM ILOG CpOptimizer 12.5; in the CP model, CpOptimizer
constraints such as the alternative constraint or the noOverlap constraint are
used. From the definition of the formal model (Eq. 1-17) and from the definition
of data structures for representing the instances, obtaining a first implementation
with CpOptimizer took us less than one day, obtaining a first implementation
with CPLEX took us approximately one day, and obtaining the local search
718 C. Pralet et al.

Table 1. Results obtained on an Intel i5-520 1.2GHz, 4GBRAM for 8 realistic instances,
with a max. CPU time set to 2 minutes; #orbits gives the number of orbits which
contain acquisitions; #offs stands for the number of times instruments are switched off
and conso for the total thermal consumption; ’*’ indicates that the solution is optimal
CpOptimizer CPLEX Specific CBLS
#acqs #orbits time (s) #offs conso time (s) #offs conso time (s) #offs conso
sat1 6 1 120 6 1048 0.11 6* 1046.0* 0.43 6 1074
sat2 11 2 - - - 14.01 7* 774.38* 0.61 7 785.87
sat3 37 1 - - - 1.41 8* 2922.30* 0.95 8 2933.80
sat4 42 1 - - - 55.35 13* 3395.49* 0.95 13 3416.27
sat5 47 1 - - - 3.29 8* 2895.54* 1.10 8 2903.03
sat6 82 2 - - - 120 20 6413.07 1.76 19 6569.49
sat7 129 3 - - - - - - 2.25 17 8768.01
sat8 233 10 - - - - - - 3.86 39 17270.62

approach took us approximately one day for implementing the CBLS model and
one week for specifying and testing specific local search procedures.
Experiments were performed over realistic instances provided by the French
space agency, ranging from small instances involving 6 acquisitions to large
instances involving more than 200 acquisitions (see Table 1). To give an idea
of the problem sizes, the largest instance (sat8) leads to 61515 variables and
176842 constraints in the CpOptimizer model, 103428 rows and 62434 columns
in the CPLEX model, and 831074 variables and 258224 invariants in the InCELL
CBLS model. Globally, it can first be observed that the attempt made with
CpOptimizer is not efficient. We believe that the behavior observed is due to the
fact that the sum constraint involved in the computation of the total thermal
consumption does not allow to prune earliest/latest dates of activities that much
because of its duration-dependent nature. Another point is that the domain of
temporal variables may contain millions of values. As for the MIP approach, it
performs very well on small and medium instances, for which it is able to find
optimal solutions.2 However, for the largest instances, the MIP approach does
not find any solution in less than two minutes. Last, the CBLS approach scales
well and delivers good quality solutions in a few seconds, which is compatible
with the application needs. On the largest instance (sat8), the distribution of
computation times among local search phases is as follows: CBLS model cre-
ation: 2.5s, preprocessing step: 0.69s, Phase 1: 0.4s, Phase 2: 0.02s, Phase 3:
0.20s, Phase 4: 0.03s, Phase 5: 0.02s. Moreover, there is no come back to Phase 4
after Phase 5.

6 Conclusion
This paper presents a specific CBLS approach for tackling a scheduling problem
from the space domain. This problem raises several questions for CP. Indeed,
2
Without further details, we relax the requirement of having integer dates, essentially
because an integer solution can almost often be directly extracted from the solution
found by the MIP.
 Scheduling Running Modes of Satellite Instruments 719

whereas scheduling is known to be one of the most successful application area of

CP, one lesson learned is that it can be difficult for standard CP solvers to deal
with scheduling problems where the duration of some tasks is not fixed initially
and where the duration of these tasks has an impact on resource consumption.
Also, we believe that alternative choices between task decompositions make the
propagation of the resource constraint even harder. On the opposite, it is easy
to enlarge the right state intervals for the specific local search procedure or for
the MIP approach. Due to the good performances of MIP on small and medium
instances, it would probably be useful either to run CBLS and MIP in parallel,
or to use the MIP model for realizing Large Neighborhood Search (LNS [11]). In
another direction, the treatment of this application will also make us add new
elements in our CBLS solver, for representing abstract tasks decomposable into
basic activities following a set of available decomposition methods.

References
1. Baptiste, P., Pape, C.L., Nuijten, W.: Constraint-based Scheduling: Applying Con-
straint Programming to Scheduling Problems. Kluwer Academic Publishers (2001)
2. Pinedo, M.: Scheduling: Theory, Algorithms, and Systems. Springer (2012)
3. Erol, K., Hendler, J., Nau, D.S.: HTN Planning: Complexity and Expressivity.
In: Proc. of the 12th National Conference on Artificial Intelligence (AAAI 1994),
pp. 1123–1128 (1994)
4. Laborie, P., Rogerie, J.: Reasoning with Conditional Time-Intervals. In: FLAIRS
Conference, pp. 555–560 (2008)
5. Pesant, G.: A Regular Language Membership Constraint for Finite Sequences of
Variables. In: Proc. of the 10th International Conference on Principles and Practice
of Constraint Programming (CP 2004), pp. 482–495 (2004)
6. Nemhauser, G., Wolsey, L.: Integer and Combinatorial Optimization. John Wiley
& Sons (1988)
7. Hentenryck, P.V., Michel, L.: Constraint-Based Local Search. The MIT Press
(2005)
8. Pralet, C., Verfaillie, G.: Dynamic Online Planning and Scheduling Using a Static
Invariant-based Evaluation Model. In: Proc. of the 23th International Conference
on Automated Planning and Scheduling (ICAPS 2013) (2013)
9. Pralet, C., Verfaillie, G.: Time-dependent Simple Temporal Networks: Properties
and Algorithms. RAIRO Operations Research (2013)
10. Dechter, R., Meiri, I., Pearl, J.: Temporal Constraint Networks. Artificial Intelli-
gence 49, 61–95 (1991)
11. Shaw, P.: Using Constraint Programming and Local Search Methods to Solve Vehi-
cle Routing Problems. In: Proc. of the 4th International Conference on Principles
and Practice of Constraint Programming (CP 1998), 417–431 (1998)
 Abstracts of Papers Fast Tracked
to Constraints Journal

Nicolas Beldiceanu1, Mats Carlsson1, Rémi Douence1, Helmut Simonis1

1
TASC-ASCOLA (CNRS/INRIA), Mines Nantes44307NantesFrancee-mail:
Nicolas.Beldiceanu@mines-nantes.fr e-mail: Remi.Douence@mines-nantes.fr
{Nicolas.Beldiceanu,Remi.Douence}@mines-nantes.fr

2
SICSP.O. Box 1263164 29KistaSwedene-mail: Mats.Carlsson@sics.se Mats.
Carlsson@sics.se
3
Insight Centre for Data AnalyticsUniversity College CorkCorkIrelande-mail:
Helmut.Simonis@insight-centre.org Helmut.Simonis@insight-centre.org
 Using Finite Transducers for Describing and
Synthesising Structural Time-Series Constraints

Nicolas Beldiceanu1 , Mats Carlsson2 , Rémi Douence1 , and Helmut Simonis3

1
TASC-ASCOLA (CNRS/INRIA), Mines Nantes, FR – 44307 Nantes, France
FirstName.LastName@mines-nantes.fr
2
SICS, P.O. Box 1263, SE – 164 29 Kista, Sweden
Mats.Carlsson@sics.se
3
Insight Centre for Data Analytics, University College Cork, Ireland
Helmut.Simonis@insight-centre.org

Abstract. We describe a large family of constraints for structural time

series by means of function composition. These constraints are on aggre-
gations of features of patterns that occur in a time series, such as the
number of its peaks, or the range of its steepest ascent. The patterns
and features are usually linked to physical properties of the time series
generator, which are important to capture in a constraint model of the
system, i.e. a conjunction of constraints that produces similar time series.
We formalise the patterns using finite transducers, whose output alpha-
bet corresponds to semantic values that precisely describe the steps for
identifying the occurrences of a pattern. Based on that description, we
automatically synthesise automata with accumulators, which can be used
as constraint propagators, as well as constraint checkers. We show the
correctness of the synthesis method using a parametrised formalisation.
The description scheme not only unifies the structure of the existing
30 time-series constraints in the Global Constraint Catalogue, but also
leads to over 600 new constraints, with more than 100,000 lines of syn-
thesised code. Examples of initial use-cases based on the new constraints
are given.1

Acknowledgements. The first author was partially supported by the Gaspard-Monge

program that initially motivated this research. The first and third authors were partially
supported by the European H2020 FETPROACT-2014 project “GRACeFUL”. The
last author was partially supported by a senior researcher chair offered by the Région
Pays de la Loire. He was also supported by EU FET grant ICON (project number
284715).

1
A full description of this work will appear in Constraints 21(1).

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 723, 2015.
DOI: 10.1007/978-3-319-23219-5
 Projection, Consistency, and George Boole

J.N. Hooker

Carnegie Mellon University, Pittsburgh, USA

jh38@andrew.cmu.edu

Abstract. Although George Boole is best known for his work in sym-
bolic logic, he made a strikingly original contribution to the logic of
probabilities. He formulated the probabilistic inference problem as an
optimization problem we now call linear programming. He solved the
problem with a projection method we now call Fourier-Motzkin elimina-
tion. In a single stroke, he linked the concepts of projection, optimization,
and logical inference.

Boole’s insight extends to constraint programming as well, because consistency

maintenance is a form of projection. Projection is, in fact, the root idea that
unites the other three concepts. Optimization is projection of the feasible set
onto a variable that represents the objective value. Inference can be understood
as projection onto a desired subset of variables. Consistency maintenance is a
form of inference that is likewise equivalent to projection.
We propose to exploit this unifying vision in constraint programming by
addressing consistency maintenance explicitly as a projection problem. Existing
types of consistency are already forms of projection, but viewing them in this
light suggests a particularly simple type of consistency that has apparently not
seen application. We call it J-consistency, which is achieved by projecting the
problem’s solution set onto a subset J of variables.
After a review of Boole’s contribution, we first present a projection method
for inference in propositional logic, based on logic-based Benders decomposition.
We show how conflict clauses generated during solution of the satisfiability
problem can deliver Benders cuts that describe the projection.
We next observe that domain consistency is a particularly simple form of
projection, while k-consistency is a less obvious form. We then indicate how
achieving J-consistency can reduce backtracking when the solver propagates
through a richer structure than a domain store, such as a relaxed decision dia-
gram. We show how to compute projections for a few popular global constraints,
including among, sequence, regular, and all-different constraints. These results
suggest that achieving J-consistency could have practical application in a solver.
We conclude by proposing a generalization of bounds consistency as pro-
jection of the convex hull of the solution set. This implies that computing
generalized bounds consistency is related to the the project of identifying valid
cutting planes, which has long been pursued in mathematical programming.

Full paper to be published in Constraints 21(1).


c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 724, 2015.
DOI: 10.1007/978-3-319-23219-5
 On Computing Minimal Independent Support
and Its Applications to Sampling and Counting
(Extended Abstract)

Alexander Ivrii1 , Sharad Malik2 , Kuldeep S. Meel3 , and Moshe Y. Vardi3

1
IBM Research Lab, Haifa, Israel
2
Princeton University
3
Department of Computer Science, Rice University

Abstract. Constrained sampling and counting are two fundamental

problems arising in domains ranging from artificial intelligence and secu-
rity, to hardware and software testing. Recent approaches to approximate
solutions for these problems rely on employing combinatorial (SAT/SMT)
solvers and universal hash functions that are typically encoded as XOR
constraints of length n/2 for an input formula with n variables. It has
been observed that lower density XORs are easy to reason in practice and
the runtime performance of solvers greatly improves with the decrease
in the length of XOR-constraints. Therefore, recent research effort has
focused on reduction of length of XOR constraints. Consequently, a
notion of Independent Support was proposed, and it was shown that
constructing XORs over independent support (if known) can lead to a
significant reduction in the length of XOR constraints without losing the
theoretical guarantees of sampling and counting algorithms [1].

In this paper, we present the first algorithmic procedure (and a correspond-

ing tool, called MIS) to determine minimal independent support for a given
CNF formula by employing a reduction to group minimal unsatisfiable subsets
(GMUS). Extensive experiments demonstrate that MIS scales to large formulas
with tens of thousands of variables, and the minimal independent support com-
puted by MIS is typically of 1/10 to 1/100th size of the support of the formulas.
By utilizing minimal independent supports computed by MIS, we provide new
tight theoretical bounds on the size of XOR constraints for constrained counting
and sampling – in some cases, even better than previously observed empirical
bounds. Furthermore, the universal hash functions constructed from independent
supports computed by MIS provide one to two orders of magnitude performance
improvement in state-of-the-art constrained sampling and counting tools, while
still retaining theoretical guarantees.

References
1. S. Chakraborty, K. S. Meel, and M. Y. Vardi. Balancing scalability and uniformity
in SAT witness generator. In Proc. of DAC, pages 1–6, 2014.
Authors names are ordered alphabetically by last name and does not indicate
contribution.
The full version appears in Constraints 21(1).

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 725, 2015.
DOI: 10.1007/978-3-319-23219-5
 General Game Playing with Stochastic CSP

Frédéric Koriche, Sylvain Lagrue, Éric Piette, and Sébastien Tabary

Université Lille-Nord de France

CRIL - CNRS UMR 8188 Artois, F-62307 Lens, France
{koriche,lagrue,epiette,tabary}@cril.fr
http://www.cril.univ-artois.fr/∼koriche,lagrue,epiette,tabary

Abstract. 1 The aim of General Game Playing (GGP) is to devise game

playing algorithms which are not dedicated to a particular strategic
game, but are general enough to effectively play a wide variety of games.
A tournament is held every year by AAAI, in which artificial game play-
ers are supplied the rules of arbitrary new games and, without human
intervention, have to play these game optimally. Games rules are described
in a declarative representation language, called GDL for Game Descrip-
tion Language. The lastest version of this language is expressive enough
to describe finite multi-player games with uncertain and incomplete infor-
mation. GGP algorithms include, among others, answer set programming
methods, automated construction of evaluation functions, and Monte-
Carlo methods such as Upper Confidence bounds for Trees (UCT). Beyond
its play value, GGP offers a rigorous setting for modeling and analyzing
sequential decision-making algorithms in multi-agent environments.

By providing a declarative approach for representing and solving combinatorial

problems, Constraint Programming appears as a promising technology to address
the GGP challenge. Currently, several constraint-based formalisms have already
been proposed to model and solve games; they include notably Quantified CSP,
Strategic CSP and Constraint Games. Most of these formalisms are, however,
restricted to deterministic, perfect information games: during each round of the
game, players have full access to the current state and their actions have deter-
ministic effects. This paper focuses on stochastic games, with chance events,
using the framework of stochastic constraint networks.
More precisely, we study a fragment of the Stochastic Constraint Satisfaction
Problem (SCSP), that captures GDL games with uncertain (but complete) infor-
mation. Interestingly, the SCSP for this class of games can be decomposed into a
sequence of µSCSPs (a.k.a one-stage stochastic constraint satisfaction problems).
Based on this decomposition, we propose a sequential decision-making algorithm,
MAC-UCB, that combines the MAC algorithm (Maintaining Arc Consistency) for
solving each µSCSP, and the multi-armed bandits Upper Confidence Bound (UCB)
method for approximating the expected utility of strategies. We show that in prac-
tice MAC-UCB significantly outperforms (the multi-player version of) UCT, which
1
This paper has been published in Constraints 21(1), the Journal Fast Track issue of
CP’15.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 726–727, 2015.
DOI: 10.1007/978-3-319-23219-5
 General Game Playing with Stochastic CSP 727

is the state-of-the-art GGP algorithm for stochastic games. MAC-UCB also domi-
nates FC-UCB, a variant where MAC is replaced with the Forward Checking (FC)
algorithm. Such conclusions are drawn from comparing the performance of these
algorithms, using extensive experiments (about 1, 350, 000 face-offs) over a wide
range of GDL games.
 Visual Search Tree Profiling

Maxim Shishmarev1 Christopher Mears1 ,

Guido Tack2 , and Maria Garcia de la Banda2
1
Faculty of IT, Monash University, Australia
2
National ICT Australia (NICTA) Victoria
{maxim.shishmarev,chris.mears,guido.tack,maria.garciadelabanda}@monash.edu

Finding the best combination of model, solver and search strategy for solving a
given combinatorial problem is a challenging, iterative process with three classi-
cal steps: (1) observing the behaviour of the program; (2) developing a hypothesis
about why certain unwanted behaviour occurs; (3) modifying the program to test
the hypothesis by observing a change in behaviour.
Performance profiling tools are designed to support and speed up this process
in a variety of ways. For constraint programs based on tree search, the existing
profiling tools mostly focus on visualising the search tree and related informa-
tion or gathering crude aggregate measures. While these tools can provide very
useful information, they are also limited by requiring close coupling between
the profiling tool and the solver or search strategy. This makes it difficult for
the profiling tool to integrate new solvers and search strategies, or to compare
different solvers. Also, they often leave the user to explore the search tree in a
relatively unguided way, which is particularly problematic for large trees. Finally,
and most importantly, almost none of these tools support the comparison of dif-
ferent executions of the same problem, where one or more aspects of the model,
solver, or search strategy have been modified.
We present an architecture and tool set specifically designed to support pro-
grammers during the performance profiling process. In particular, the profiling
tool enables programmers to extract the information they need to build hypothe-
ses as to why the search is behaving in a particular way. Then it helps to validate
their hypotheses by identifying and visualising the effect of changes in the program
(be they changes in the model, solver, or search strategy). The former is achieved
by providing different views and analyses of the execution as well as efficient nav-
igation tools to help users focus their attention on the parts of the execution that
might be worth modifying. The latter is achieved by providing two tools: one that
can replay searches using a different model or solver, so that the user can isolate
the effect of a change; and a second tool that can visually “merge” the common
parts of two executions and allows users to explore the parts that differ.
The architecture is designed to easily accommodate new solvers and search
strategies, and it allows for solvers to be easily integrated without imposing
an unreasonable overhead on the program’s execution. The modularity of the
design supports the combination of many different visualisation and analysis
components. Our implementation is available at https://github.com/cp-profiler.

This paper is published in Constraints 21(1).


c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 728, 2015.
DOI: 10.1007/978-3-319-23219-5
Long-Haul Fleet Mix and Routing Optimisation
with Constraint Programming and Large
Neighbourhood Search

Philip Kilby and Tommaso Urli

Optimisation Research Group, NICTA Canberra Research Lab, 7 London Circuit,

Canberra, 2601 ACT, Australia
{philip.kilby,tommaso.urli}@nicta.com.au

Abstract. We present an original approach to compute efficient mid-

term fleet configurations, at the request of a Queensland-based long-haul
trucking carrier. Our approach considers one year’s worth of demand
data, and employs a constraint programming (CP) model and an adap-
tive large neighbourhood search (LNS) scheme to solve the underly-
ing multi-day multi-commodity split delivery capacitated vehicle routing
problem. A Pareto-based pre-processing step allows to tackle the large
amount of data by extracting a set of representative days from the full
horizon. Our solver is able to provide the decision maker with a set of
Pareto-equivalent fleet setups trading off fleet efficiency against the likeli-
hood of requiring on-hire vehicles and drivers. Moreover, the same solver
can be used to solve the daily loading and routing problem. We carry
out an extensive experimental analysis based on real-world data, and we
discuss the properties of the generated fleets. We show that our app-
roach is a sound methodology to provide decision support for the mid-
and short-term decisions of a long-haul carrier1 .

1
An extended version of this paper, with the title Fleet Design Optimisation From
Historical Data Using Constraint Programming and Large Neighbourhood Search,
will be published in Constraints 21(1).

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 729, 2015.
DOI: 10.1007/978-3-319-23219-5
Abstracts of Published Journal Track Papers
 On the Reification of Global Constraints
(Abstract)

Nicolas Beldiceanu1 , Mats Carlsson2 , Pierre Flener3 , and Justin Pearson3

1
TASC (CNRS/INRIA), Mines Nantes, FR – 44307 Nantes, France
2
SICS, P.O. Box 1263, SE – 164 29 Kista, Sweden
3
Uppsala University, Dept of Information Technology, SE – 751 05 Uppsala, Sweden

Conventional wisdom has it that many global constraints cannot be easily reified,
i.e., augmented with a 0-1 variable reflecting whether the constraint is satisfied
(value 1) or not (value 0). Reified constraints are useful for expressing proposi-
tional formulas over constraints, and for expressing that a number of constraints
hold (e.g., a cardinality operator). The negation of constraints is important for
inferring constraints from negative examples and for proving the equivalence of
models. Many early CP solvers, such as CHIP, GNU Prolog, Ilog Solver, and
SICStus Prolog, provide reification for arithmetic constraints. However, when
global constraints started to get introduced (e.g., alldifferent and cumu-
lative), reification was not available for them. It was believed that reification
could only be obtained by modifying the filtering algorithms attached to each
constraint. We present a reification method that works in all CP solvers.
A global constraint G(A) can be defined by restrictions R on its arguments A,
say on their bounds, and by a condition C on A, i.e., G(A) ≡ R(A) ∧ C(A).
We define the reified version of G(A) as R(A) ∧ (C(A) ⇔ b), where b is a 0-1
variable reflecting whether G(A) holds or not. We require the negation of G(A)
to satisfy the same restrictions R(A). It turns out that the condition C(A)
can often be reformulated as a conjunction F (A, V) ∧ N (A, V) of constraints,
where A and V are disjoint sets of variables, R(A) implies that F (A, V) is
a total-function constraint uniquely determining V from any A, and N (A, V)
is a Boolean combination of linear arithmetic (in)equalities and 0-1 variables;
we assume such Boolean combinations are already reifiable, as is the case in
all CP solvers we are aware of. The reified version of G(A) is then defined as
R(A) ∧ F (A, V) ∧ (N (A, V) ⇔ b). For example, alldifferent(v1 , . . . , vn ) is
reified by sort(v1 , . . . , vn  , w1 , . . . , wn ) ∧ ((w1 < w2 ∧ · · · ∧ wn−1 < wn ) ⇔ b),
where v1 , . . . , vn  uniquely determine w1 , . . . , wn  in the total-function con-
straint sort(v1 , . . . , vn  , w1 , . . . , wn ).
Surprisingly, this insight that many constraints can naturally be defined by
a determine and test scheme, where the determine part is associated to total-
function constraints that determine additional variables, and the test part to
a reifiable constraint on these variables, allows us to reify at least 313 of the
381 (i.e., 82%) constraints of the Global Constraint Catalogue. Most of the
constraints not covered are graph constraints involving set variables.
Beldiceanu, N., Carlsson, M., Flener, P., Pearson, J.: On the reification of global
constraints. Constraints 18(1), 1–6 (January 2013), http://dx.doi.org/10.1007/
s10601-012-9132-0, full details at http://soda.swedish-ict.se/5194

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 733, 2015.
DOI: 10.1007/978-3-319-23219-5
 MDD Propagation for Sequence Constraints

David Bergman1 , Andre A. Cire2 , and Willem-Jan van Hoeve3

1
School of Business, University of Connecticut
2
Department of Management, University of Toronto Scarborough
3
Tepper School of Business, Carnegie Mellon University
david.bergman@business.uconn.edu,acire@utsc.utoronto.ca,
vanhoeve@andrew.cmu.edu

We study MDD propagation for the well-known Sequence constraint, which

finds applications in, e.g., car sequencing and employee scheduling problems. It
is defined as a specific conjunction of Among constraints, where an Among con-
straint restricts the occurrence of a set of values for a sequence of variables to be
within a lower and upper bound. It is known that conventional domain consis-
tency can be established for Sequence in polynomial time, and that MDD con-
sistency can be established for Among constraints in polynomial time. However,
it remained an open question whether or not MDD consistency for Sequence
can be established in polynomial time as well.
In this work, we answer that question negatively and our first contribution is
showing that establishing MDD consistency on the Sequence constraint is NP-
hard. This is an important result from the perspective of MDD-based constraint
programming. Namely, of all global constraints, the Sequence constraint has
perhaps the most suitable combinatorial structure for an MDD approach; it has
a prescribed variable ordering, it combines sub-constraints on contiguous vari-
ables, and existing approaches can handle this constraint fully by using bounds
reasoning only. As our second contribution, we show that establishing MDD con-
sistency on the Sequence constraint is fixed parameter tractable with respect
to the lengths of the sub-sequences (the Among constraints), provided that the
MDD follows the order of the Sequence constraint. The proof is constructive,
and follows from a generic algorithm to filter one MDD with another. The third
contribution is a partial MDD propagation algorithm for Sequence, that does
not necessarily establish MDD consistency. It relies on the decomposition of
Sequence into ‘cumulative sums’, and a new extension of MDD filtering to the
information that is stored at its nodes.
Our last contribution is an experimental evaluation of our proposed partial
MDD propagation algorithm. We compare its performance with existing domain
propagators for Sequence, and with the currently best known MDD approach
that uses the natural decomposition of Sequence into Among constraints. Our
experiments demonstrate that our algorithm can reduce the search tree size,
and solving time, by several orders of magnitude in both cases. Our results thus
provide further evidence for the power of MDD propagation in the context of
constraint programming.

This is a summary of the paper “D. Bergman, A. A. Cire, and W.-J. van Hoeve. MDD
Propagation for Sequence Constraints. JAIR, Volume 50, pages 697-722, 2014”.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 734, 2015.
DOI: 10.1007/978-3-319-23219-5
 Discrete Optimization with Decision Diagrams

David Bergman1 , Andre A. Cire2 , Willem-Jan van Hoeve3 , and John Hooker3
1
School of Business, University of Connecticut
2
Department of Management, University of Toronto Scarborough
3
Tepper School of Business, Carnegie Mellon University
david.bergman@business.uconn.edu,acire@utsc.utoronto.ca,
vanhoeve@andrew.cmu.edu, jh38@andrew.cmu.edu

Decision diagrams (DDs) are compact graphical representations of Boolean func-

tions originally introduced for applications in circuit design and formal verifica-
tion. Recently, DDs have also been used for a variety of purposes in optimization
and operations research. These include facet enumeration in integer program-
ming, maximum flow computation in large-scale networks, solution counting in
combinatorics, and learning in genetic programming.
In this paper we provide a number of contributions for the application of
decision diagrams to discrete optimization problems. First, we propose a novel
branch-and-bound algorithm for discrete optimization in which binary decision
diagrams (BDDs) play the role of the traditional linear programming relaxation.
In particular, relaxed BDD representations of the problem provide bounds and
guidance for branching, while restricted BDDs supply a primal heuristic. We
show how to compile relaxed and restricted BDDs given a dynamic programming
(DP) model of the problem. This compilation process is based on augment-
ing the DP model with problem-specific relaxation and restriction operators,
which are applied on the state transition graph of the model to derived relaxed
and restricted BDDs, respectively. Finally, our novel search scheme branches
within relaxed BDDs rather than on variable values, eliminating symmetry
during search.
We report computational results of our branch-and-bound algorithm on a
previously introduced DP model of the maximum weighted independent set
problem, as well as on new DP formulations for the maximum cut problem and
the maximum 2-SAT problem. The results are compared with those obtained
from generic state-of-the-art solvers, in particular an integer programming and
a semidefinite programming solver. We show that a rudimentary BDD-based
solver is already competitive with or superior to existing technology. Specific to
the maximum cut problem, we tested the BDD-based solver on a classical bench-
mark set and identified tighter relaxation bounds than have ever been found
by any technique, nearly closing the entire optimality gap on four large-scale
instances.

This is a summary of the paper “D. Bergman, A. A. Cire, W.-J. van Hoeve, and
J. Hooker. Discrete Optimization with Decision Diagrams. INFORMS Journal on
Computing, forthcoming 2015.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 735, 2015.
DOI: 10.1007/978-3-319-23219-5
 A Hybrid Approach Combining Local Search
and Constraint Programming for a Large Scale
Energy Management Problem

Haris Gavranović1 and Mirsad Buljubašić2

1
International University of Sarajevo, Bosnia and Herzegovina
haris.gavranovic@gmail.com
2
Ecole des Mines d’Ales, LGI2P Research Center, Nimes, France
mirsad.buljubasic@mines-ales.fr

In this work we take the perspective of a large utility company, tackling their
problems in modeling and planning production assets, i.e., a multitude of power
plants. The goal is to fulfill the respective demand of energy over a time horizon
of several years, while minimizing a total operation cost.
The proposed problem consists of modeling the production assets and finding
an optimal outage schedule that includes two mutually dependent and related
subproblems:
1. Determining the schedule of plant outages. This schedule must satisfy a cer-
tain number of constraints in order to comply with limitations on resources,
which are necessary to perform refueling and maintenance operations.
2. Determining an optimal production plan to satisfy demand. Production must
also satisfy some technical constraints.
The work presents a heuristic approach combining constraint satisfaction,
local search and a constructive optimization algorithm for a large-scale energy
management and maintenance scheduling problem. The methodology shows how
to successfully combine and orchestrate different types of algorithms and to
produce competitive results. The numerical results obtained on the close-to-
real EDF instances testify about the quality of the approach. In particular, the
method achieves 3 out of 15 possible best results.
It is possible to use different approaches, such as local search, constraint pro-
gramming or constructive procedure, to determine a feasible outage schedule.
Constraint programming with a performant solver emerged to be the most reli-
able, and probably the most elegant approach, with very good numerical results.
The set of additional, artificial, constraints are added to the model to improve
the diversity and the feasibility of the remaining production assignment problem.
Finally, the feasible solution found in this way is then improved using a series of
local improvements.

References
1. H. Gavranović, and M. Buljubaić, A Hybrid Approach Combining Local Search and
Constraint Programming for a Large Scale Energy Management Problem, RAIRO -
Operations Research, Volume 47,Issue 04,2013,pp 481-500,DOI 10.1051/ro/2013053,

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 736, 2015.
DOI: 10.1007/978-3-319-23219-5
 Representing and Solving Finite-Domain
Constraint Problems using Systems of
Polynomials (Extended Abstract)

Chris Jefferson1 , Peter Jeavons2 , Martin J. Green3 , and M.R.C. van Dongen4
1
Department of Computer Science, University of St Andrews, UK
2
Department of Computer Science, University of Oxford, UK
3
Department of Computer Science, Royal Holloway, University of London, UK
4
Department of Computer Science, University College Cork, Ireland
In this paper [1] we investigate the use of a system of multivariate polynomials
to represent the restrictions imposed by a collection of constraints. A system
of polynomials is said to allow a particular combination of values for a set of
variables if the simultaneous assignment of those values to the variables makes
all of the polynomials in the system evaluate to zero.
The use of systems of polynomials has been considered a number of times
in the constraints literature, but is typically used to represent constraints on
continuous variables. Here we focus on the use of polynomials to represent finite
domain constraints. One advantage of representing such constraints using poly-
nomials is that they can then be treated in a uniform way along with continuous
constraints, allowing the development of very general constraint-solving tech-
niques. Systems of polynomials can be processed by standard computer algebra
packages such as Mathematica and REDUCE, so our approach helps to unify
constraint programming with other forms of mathematical programming.
Systems of polynomials have been widely studied, and a number of general
techniques have been developed, including algorithms that generate an equiva-
lent system with certain desirable properties, called a Gröbner Basis. Given a
Gröbner Basis, it is possible to obtain the solutions to a system of polynomials
very easily (or determine that it has no solutions). A Gröbner Basis provides
a convenient representation for the whole set of solutions which can be used to
answer a wide range of questions about them, such as the correlations between
individual variables, or the total number of solutions.
In general, the complexity of computing a Gröbner Basis for a system of
polynomials is doubly-exponential in the size of the system. However, we observe
that with the systems we use for constraints over finite domains this complexity
is only singly exponential, and so comparable with other search techniques for
constraints. Our main contributions are a family of polynomial-time algorithms,
related to the construction of Gröbner Bases, which simulate the effect of the
local-consistency algorithms used in constraint programming. Finally, we discuss
the use of adaptive consistency techniques for systems of polynomials.
References
1. C. Jefferson, P. Jeavons, M. J. Green, and M. R. C. van Dongen. Representing and
solving finite-domain constraint problems using systems of polynomials. Annals of
Mathematics and Artificial Intelligence, 67(3-4):359–382, 2013.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 737, 2015.
DOI: 10.1007/978-3-319-23219-5
 A Quadratic Extended Edge-Finding Filtering
Algorithm for Cumulative Resource Constraints

Roger Kameugne1,2 , Laure Pauline Fotso3 , and Joseph Scott4

1
Univ. of Maroua, Dept. of Mathematics, Cameroon
rkameugne@yahoo.fr
2
Univ. of Yaoundé I, Dept. of Mathematics, Yaoundé, Cameroon
3
Univ of Yaoundé I, Dept. of Computer Sciences, Cameroon
laurepfotso@yahoo.com
4
Uppsala Univ., Dept. of Information Technology, Sweden
joseph.scott@it.uu.se

Extended edge-finding is a well-known filtering rule used in constraint-based

scheduling problems for the propagation of constraints over both disjunctive
and cumulative resources. A cumulative resource is renewable resource of fixed
capacity, which may execute multiple tasks simultaneously so long as the resource
requirement of tasks scheduled together never exceeds the capacity. In constraint-
based scheduling, a single cumulative resource is typically modelled by means
of the global constraint cumulative, which enforces the resource constraint for
a set of tasks. While complete filtering for cumulative is NP-complete there
are known polynomial time algorithms for solving specific relaxations of the
constraint, including timetabling, edge-finding variants (i.e., standard, extended,
and timetable), not-first/not-last, and energetic reasoning. These relaxations
may lead to different pruning, so cumulative is commonly implemented as a
conjunction of relaxations, yielding stronger fixpoints and smaller search trees.
Extended edge-finding is similar to the better known edge-finding algorithm:
both reduce the range of possible start (resp. end) times of tasks through the
deduction of new ordering relations. For a task i ∈ T , an extended edge-finder
searches for a set of tasks that must end before the end (resp. start before the
start) of i. Based on this newly detected precedence, the earliest start time (resp.
latest completion time) of i is updated. The best previously known complexity for
a cumulative extended edge-finding algorithm was O(kn2 ), where n is the number
of tasks on the resource, and k is the number of distinct capacity requirements.
We propose an O(n2 ) extended edge-finding algorithm for cumulative resources.
Similar to previous work on edge-finding, this new algorithm uses the notion of
minimum slack to detect when extended edge-finding may strengthen a domain.
The algorithm is more efficacious on domains already at the fix point of edge-
finding. Experimental evaluation on a standard benchmark suite confirms that
our algorithm outperforms previous extended edge-finding algorithms. We find
This is a summary of the paper: R. Kameugne, L.P. Fotso, and J. Scott. A Quadratic
Extended Edge-Finding Filtering Algorithm for Cumulative Resource Constraints.
International Journal of Planning and Scheduling 1(4):264–284, 2013.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, pp. 738–739, 2015.
DOI: 10.1007/978-3-319-23219-5
 A Quadratic Extended Edge-Finding Filtering Algorithm 739

that the conjunction of edge-finding and extended edge-finding leads to a slight

decrease in efficiency in most cases when compared to edge-finding alone; how-
ever, on some instances the stronger pruning of the combined algorithm results
in substantial performance gains. Furthermore, we show that our method is com-
petitive with the current state-of-the-art in edge-finding based algorithms.
 Achieving Domain Consistency and Counting
Solutions for Dispersion Constraints

Gilles Pesant

École Polytechnique de Montréal, Montreal, Canada

CIRRELT, Université de Montréal, Montreal, Canada
gilles.pesant@polymtl.ca

Many combinatorial problems require of their solutions that they achieve a cer-
tain balance of given features. For example in assembly line balancing the work-
load of line operators must be balanced. In rostering we usually talk of fairness
instead of balance, because of the human factor — here we want a fair distri-
bution of the workload or of certain features such as weekends off and night
shifts. In the Balanced Academic Curriculum Problem, courses are assigned to
semesters so as to balance the academic load between semesters. Because of
additional constraints (prerequisite courses, minimum and maximum number of
courses per semester) and a varying number of credits per course, reaching per-
fect balance is generally impossible. A common way of encouraging balance at
the modeling level is to set reasonable bounds on each load, tolerating a certain
deviation from the ideal value, but it has the disadvantage of putting on an
equal footing solutions with quite different balance. Another way is to minimize
some combination of the individual deviations from the ideal value, but if other
criteria to optimize are already present one must come up with suitable weights
for the different terms of the objective function. Yet another way is to bound
the sum of individual deviations.
In Constraint Programming the spread and deviation constraints have been
proposed to express balance among a set of variables by constraining their mean
and their overall deviation from the mean. Currently the only practical filtering
algorithms known for these constraints achieve bounds consistency. In [1] we
propose a more general constraint, dispersion, which includes both previous
constraints as special cases and other variants of interest. We improve filtering
by presenting an efficient domain consistency algorithm and also extend it to
count solutions so that it can be used in counting-based search, a generic and
effective family of branching heuristics that free the user from having to write
problem-specific search heuristics.
We provide empirical evidence that these can lead to significant practical
improvements in combinatorial problem solving. In particular we improve the
state of the art on the Nurse to Patient Assignment Problem and on the Balanced
Academic Curriculum Problem.

References
1. Gilles Pesant. Achieving Domain Consistency and Counting Solutions for Dispersion
Constraints. INFORMS Journal on Computing, forthcoming. (Also available as
https://www.cirrelt.ca/DocumentsTravail/CIRRELT-2015-08.pdf)

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 740, 2015.
DOI: 10.1007/978-3-319-23219-5
 meSAT: Multiple Encodings of CSP to SAT

Mirko Stojadinović and Filip Marić

Faculty of Mathematics, University of Belgrade, Serbia

{mirkos,filip}@matf.bg.ac.rs

Abstract. One approach for solving Constraint Satisfaction Problems

(CSP) (and related Constraint Optimization Problems (COP)) involving
integer and Boolean variables is reduction to propositional satisfiability
problem (SAT). A number of encodings (e.g., direct, log, support, order,
compact-order) for this purpose exist as well as specific encodings for
some constraints that are often encountered (e.g., cardinality constraints,
global constraints). However, there is no single encoding that performs
well on all classes of problems and there is a need for a system that
supports multiple encodings. In our paper meSAT: Multiple encodings
of CSP to SAT we describe a system that translates specifications of
finite linear CSP problems into SAT instances using several well-known
encodings, and their combinations. The system also provides choice of
using different SAT encodings of cardinality constraints. The experi-
ments show that efficiency of direct, support and order encodings can be
improved by their combinations, namely direct-support and direct-order
encodings. Description of efficient encodings of global constraints within
direct encoding and the correctness proof of a variant of order encod-
ing are also provided. We present a methodology for selecting a suitable
encoding based on simple syntactic features of the input CSP instance.
The selection method is a modification of approach ArgoSmArT-kNN
that uses k-nearest neighbors algorithm and was originally developed for
SAT. Thorough evaluation has been performed on large publicly available
corpora and our encoding selection method improves upon the efficiency
of existing encodings and state-of-the-art tools used in comparison. We
also show that if families of syntactically similar instances can be formed,
then the selection can be trained only on very easy instances thus signif-
icantly reducing the training time. The idea is based on the observation
that very often the solver that is efficient in solving easier instances of
some problem will be also efficient in solving harder instances of the very
same problem. The obtained results are very good, despite using short
training time.

This is extended abstract of the paper: M. Stojadinović, F. Marić, mesat: Multiple

encodings of csp to sat Constraints, Volume 19(4), pp. 380-403, (2014.)
This work was partially supported by the Serbian Ministry of Science grant 174021.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 741, 2015.
DOI: 10.1007/978-3-319-23219-5
 Constraint programming for LNG ship
scheduling and inventory management
(Abstract)

Willem-Jan van Hoeve

Tepper School of Business, Carnegie Mellon University

5000 Forbes Avenue, Pittsburgh, PA 15213, USA
vanhoeve@andrew.cmu.edu

This work considers a central operational issue in the liquefied natural gas (LNG)
industry: designing schedules for the ships to deliver LNG from the production
(liquefaction) terminals to the demand (regasification) terminals. Operationally,
a delivery schedule for each customer in the LNG supply chain is negotiated on an
annual basis. The supply and demand requirements of the sellers and purchasers
are accommodated through the development of this annual delivery schedule,
which is a detailed schedule of LNG loadings, deliveries and ship movements for
the planning year. Developing and updating this schedule is a complex problem
because it combines ship routing with inventory management at the terminals.
A specific application of our work is to support supply chain design analysis
for new projects. Typically, new LNG projects require building not only the
natural gas production and treatment facilities, but also constructing the ships
and the production-side terminals, berths and storage tanks. Goel et al. (Journal
of Heuristics, 2012) first introduced an LNG inventory routing problem for this
purpose. They proposed a mixed integer programming (MIP) model as well
as a MIP-based heuristic for finding good solutions to this problem. Yet, both
the MIP model and the MIP heuristic suffered from scalability issues for larger
instances.
We propose two constraint programming (CP) models, both of which are
based on a disjunctive scheduling representation. Visits of ships to facilities
are modeled as optional activities, and distances are represented as sequence-
dependent setup times for the disjunctive resource that represents each ship’s
schedule. In addition, we model the inventory at each location as a cumulative
resource. Even if the number of (optional) activities is proportional to the num-
ber of time points, our CP models are still much smaller than the corresponding
MIP models and much more scalable. We also propose an iterative search heuris-
tic to generate good feasible solutions for these models. Computational results
on a set of large-scale test instances demonstrate that our approach can find
better solutions than existing approaches based on mixed integer programming,
while being 4 to 10 times faster on average.

This is a summary of the paper “V. Goel, M. Slusky, W.-J. van Hoeve, K. C. Fur-
man, and Y. Shao. Constraint Programming for LNG Ship Scheduling and Inventory
Management. European Journal of Operational Research 241(3): 662-673, 2015.”.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 742, 2015.
DOI: 10.1007/978-3-319-23219-5
 Revisiting the Limits of MAP Inference
by MWSS on Perfect Graphs

Adrian Weller

University of Cambridge
aw665@cam.ac.uk

We examine Boolean binary weighted constraint satisfaction problems without

hard constraints, and explore the limits of a promising recent approach that is
able to solve the problem exactly in polynomial time in a range of cases.
Specifically, we are interested in the problem of finding a configuration of
variables x = (x1 , . . . , xn ) ∈ {0, 1}n thatmaximizes a score
 function, defined
n
by unary and pairwise real terms f (x) = i=1 ψi (xi ) + (i,j)∈E ψij (xi , xj ). In
the machine learning community, this is typically known as MAP (or MPE )
inference, yielding a configuration of variables with maximum probability.
[2] showed that this problem (for discrete models of any arity) may be reduced
to finding a maximum weight stable set (MWSS) in a derived weighted graph
called an NMRF (closely related to the microstructure complement of the dual
representation [1,3,4]), and introduced the technique of pruning certain nodes.
Different reparameterizations can lead to different pruned NMRFs. A perfect
pruned NMRF allows the initial problem to be solved in polynomial time.
Here we extend [5] to consider all possible reparameterizations of Boolean
binary models. This yields a simple, exact characterization of the new, enlarged
set of such models tractable with this approach. We also show how these may
be efficiently identified, thus settling the power of the approach on this class.
The main result is that the method works for all valid cost functions if and
only if each block of the original model is almost balanced.1 This hybrid con-
dition combines restrictions on both the structure and the types of pairwise
constraints/score terms (no restrictions on unary terms).
References
1. M. Cooper and S. Živný. Hybrid tractability of valued constraint problems. Artificial
Intelligence, 175(9):1555–1569, 2011.
2. T. Jebara. MAP estimation, message passing, and perfect graphs. In Uncertainty
in Artificial Intelligence, 2009.
3. P. Jégou. Decomposition of domains based on the micro-structure of finite
constraint-satisfaction problems. In AAAI, pages 731–736, 1993.
4. J. Larrosa and R. Dechter. On the dual representation of non-binary semiring-based
CSPs. In CP2000 workshop on soft constraints, 2000.
5. A. Weller and T. Jebara. On MAP inference by MWSS on perfect graphs. In
Uncertainty in Artificial Intelligence (UAI), 2013.

This is a summary of the paper “A.Weller. Revisiting the Limits of MAP Inference
by MWSS on Perfect Graphs. JMLR, Volume 38, pages 1061-1069, 2015”.
1
A block is a maximal 2-connected subgraph. A balanced block contains no frustrated
cycle. An almost balanced block may be rendered balanced by removing one node.

c Springer International Publishing Switzerland 2015
G. Pesant (Ed.): CP 2015, LNCS 9255, p. 743, 2015.
DOI: 10.1007/978-3-319-23219-5
 Author Index

Abío, Ignasi 3 Escamocher, Guillaume 58

Akgün, Özgür 480 Even, Caroline 574
Aldanondo, M. 508
Alliot, Jean-Marc 446 Fages, Jean-Guillaume 105, 508
Allouche, David 12 Fioretto, Ferdinando 121
Arbelaez, Alejandro 499 Flener, Pierre 313, 733
Artigues, Christian 393 Fotso, Laure Pauline 738

Bagley, Claire 627 Gaborit, P. 508

Barco, A.F. 508 Gavranović, Haris 736
Beldiceanu, Nicolas 723, 729 Gay, Steven 140, 149
Benini, Luca 524 Gotteland, Jean-Baptiste 446
Bergman, David 30, 734, 735 Green, Martin J. 737
Bessiere, Christian 463, 541 Guidotti, Riccardo 638
Blindell, Gabriel Hjort 609 Guns, Tias 376
Boizumault, Patrice 226
Bonaventure, Olivier 592 Hà, Minh Hoàng 158
Bonfietti, Alessio 279 Hartert, Renaud 140, 149, 592
Borghesi, Andrea 524 Hebrard, Emmanuel 393
Briot, Nicolas 541 Hijazi, Hassan L. 39
Brown, Kenneth N. 463, 687 Hoffmann, Hella-Franziska 429
Buljubašić, Mirsad 736 Hooker, John 724, 735
Huberman, Steven 627
Carlsson, Mats 609, 723, 733
Charnois, Thierry 226 Ignatiev, Alexey 173
Cire, Andre A. 30, 734, 735 Ivrii, Alexander 725
Coffrin, Carleton 39
Colena, Mike 627 Jaubert, Jean 704
Collina, Francesca 524 Jeavons, Peter 737
Cooper, Martin C. 58, 74 Jefferson, Chris 737
Jégou, Philippe 74
Dao, Thi-Bich-Hanh 557 Jonsson, Peter 183
de Givry, Simon 12 Joshi, Saurabh 200
de la Banda, Maria Garcia 728
Dejemeppe, Cyrille 89 Kadioglu, Serdar 627
Derrien, Alban 105 Kameugne, Roger 738
Desharnais, Josée 403 Karpiński, Michał 210
Douence, Rémi 723 Katsirelos, George 12
Duchein, Aymeric 58 Kemmar, Amina 226
Duong, Khanh-Chuong 557 Kilby, Philip 654, 729
Durand, Nicolas 446 Kong, Shufeng 244
746 Author Index

Koriche, Frédéric 726 Pearson, Justin 313, 733

Kotthoff, Lars 638 Pesant, Gilles 740
Kreter, Stefan 262 Petit, Thierry 105
Piette, Éric 726
Lagerkvist, Victor 183 Piotrów, Marek 210
Lagrue, Sylvain 726 Pontelli, Enrico 121
Lam, Edward 654 Pralet, Cédric 704
Laviolette, François 403 Prestwich, Steven D. 351
Le, Tiep 121 Previti, Alessandro 173
Lebbah, Yahia 226 Prosser, Patrick 295, 671
Lecoutre, Christophe 140 Prud’homme, Charles 105
Lemai-Chenevier, Solange 704
Li, Sanjiang 244 Quimper, Claude-Guy 158, 403
Li, Yongming 244
Liffiton, Mark 173
Lombardi, Michele 279, 524 Razgon, Igor 367
Long, Zhiguo 244 Rendl, Andrea 376
Loudni, Samir 226 Rossi, Roberto 351
Lozano, Roberto Castañeda 609 Rousseau, Louis-Martin 158
Roy, Pierre 341
Macdonald, Craig 671 Sakellariou, Jason 341
Malik, Sharad 725 Schaus, Pierre 89, 140, 149, 592
Manquinho, Vasco 200 Schiex, Thomas 12
Manzano, Óscar 687 Schulte, Christian 609
Marić, Filip 741 Schutt, Andreas 262, 574
Marques-Silva, Joao 173 Scott, Joseph 738
Martins, Ruben 200 Shishmarev, Maxim 728
Mayer-Eichberger, Valentin 3 Siala, Mohamed 393
McCreesh, Ciaran 295, 671 Simard, Frédéric 403
Mears, Christopher 728 Simonis, Helmut 723
Mechqrane, Younes 463 Son, Tran Cao 121
Meel, Kuldeep S. 725 Spracklen, Patrick 330
Mehta, Deepak 499 Stergiou, Kostas 419
Miguel, Ian 330, 480 Stojadinović, Mirko 741
Milano, Michela 279, 524 Stuckey, Peter J. 3, 262, 376
Miller, Alice 671
Monette, Jean-Noël 313 Tabary, Sébastien 726
Morin, Michael 403 Tack, Guido 376, 728
Murphy, Seán Óg 687 Tarim, S. Aramagan 351
Terrioux, Cyril 74
Nanni, Mirco 638
Nightingale, Peter 330 Urli, Tommaso 729

O’Sullivan, Barry 499, 638 van Beek, Peter 429

Van Cauwelaert, Sascha 89
Pachet, François 341 van Dongen, M.R.C. 737
Papadopoulos, Alexandre 341 Van Hentenryck, Pascal 39, 574, 654
 Author Index 747

van Hoeve, Willem-Jan 30, 734, 735, 742 Wahbi, Mohamed 463
Vanaret, Charlie 446 Weller, Adrian 743
Vareilles, E. 508 Wetter, James 480
Vardi, Moshe Y. 725
Vismara, Philippe 541 Yeoh, William 121
Vissicchio, Stefano 592
Vrain, Christel 557 Zytnicki, Matthias 12