Slides - Constraint Programming

Constraint Programming
Justin Pearson
Uppsala University
1st July 2016
Special thanks to Pierre Flener, Joseph Scott and Jun He

CP MiniZinc Strings Other Final
Outline
1 Introduction to Constraint Programming (CP)
2 Introduction to MiniZinc
3 Constraint Programming with Strings
4 Some recent, and not so Recent Advances
5 Final Thoughts
Uppsala University Justin Pearson Constraint Programming

Constraint Programming in a Nutshell
Slogan of CP
Constraint Program = Model [ + Search ]
CP provides:
high level declarative modelling abstractions,
a framework to separate search from from modelling.
Search proceeds by intelligent backtracking with intelligent
inference called propagation that is guided by high-level
modelling constructs called global constraints.

Outline

Modelling
Propagation
Systematic Search
History of CP
5 Final Thoughts

Constraint-Based Modelling
Example (Port tasting, PT)
Alva Dan Eva Jim Leo Mia Ulla
2011
2003
2000
1994
1977
1970
1966
Constraints to be satisfied:
Equal jury size: Every wine is evaluated by 3 judges.
Equal drinking load: Every judge evaluates 3 wines.
Fairness: Every port pair has 1 judge in common.

2011 3 3 3
2003 3 3 3
2000 3 3 3
1994 3 3 3
1977 3 3 3
1970 3 3 3
1966 3 3 3
Equal jury size: Every wine is evaluated by 3 judges.

2011 3 3 3
2003 3 3 3
2000 3 3 3
1994 3 3 3
1977 3 3 3
1970 3 3 3
1966 3 3 3
Equal jury size: Every port is evaluated by 3 judges.

2011 1 1 1 0 0 0 0
2003 1 0 0 1 1 0 0
2000 1 0 0 0 0 1 1
1994 0 1 0 1 0 1 0
1977 0 1 0 0 1 0 1
1970 0 0 1 1 0 0 1
1966 0 0 1 0 1 1 0

Example (PT integer model: 3 1 and 0)

1 int: NbrVint; int: NbrJudges;
2 set of int: Wines = 1..NbrVint;
3 set of int: Judges = 1..NbrJudges;
4 int: JurySize; int: Capacity; int: Fairness;
5 array[Wines,Judges] of var 0..1: PT;
6 solve satisfy;
7 forall(t in Wines)
8 (JurySize = sum(j in Judges)(PT[t,j]));
9 forall(j in Judges)
10 (Capacity = sum(t in Wines)(PT[t,j]));
11 forall(t,t in Wines where t<t)
12 (Fairness = sum(j in Judges)(PT[t,j]*PT[t,j]));
Example (Instance data for the SMT PT instance)

1 NbrVint = 7; NbrJudges = 7;
2 JurySize = 3; Capacity = 3; Fairness = 1;

Example (Idea for another PT model)

2011 {Alva, Dan, Eva }
2003 {Alva, Jim, Leo }
2000 {Alva, Mia, Ulla}
1994 { Dan, Jim, Mia }
1977 { Dan, Leo, Ulla}
1970 { Eva, Jim, Ulla}
1966 { Eva, Leo, Mia }

Example (PT set model: each port has a judge set )

1 ...
2 ...
3 ...
4 ...
5 array[Wines] of var set of Judges: PT;
6 ...
7 forall(t in Wines)
8 (JurySize = card(PT[t]));
9 forall(j in Judges)
10 (Capacity = sum(t in Wines)(bool2int(j in PT[t])));
11 forall(t,t in Wines where t<t)
12 (Fairness = card(PT[t] inter PT[t]));
Example (Instance data for the PT instance)

1 ...
2 ...

6 1 4 5
8 3 5 6
2 1
8 4 7 6
6 3
7 9 1 4
5 2
7 2 6 9
4 5 8 7
Example (Sudoku model)

1 array[1..9,1..9] of var 1..9: Sudoku;
2 ...
3 solve satisfy;
4 forall(r in 1..9)
5 (ALLDIFFERENT([Sudoku[r,c] | c in 1..9]));
6 forall(c in 1..9)
7 (ALLDIFFERENT([Sudoku[r,c] | r in 1..9]));
8 forall(i,j in {1,4,7})
9 (ALLDIFFERENT([Sudoku[r,c] | r in i..i+2, c in j..j+2]));
Global Constraints
Global constraints such as A LL D IFFERENT and S UM enable the

preservation of combinatorial sub-structures of a constraint
problem, both while modelling it and while solving it. Dozens of
n-ary constraints (Catalogue) have been identified and
encapsulate complex propagation algorithms declaratively.,
including
n-ary linear and non-linear arithmetic
Rostering under balancing & coverage constraints
Scheduling under resource & precedence constraints
Geometrical constraints between points, segments, . . .
There are many more.

CP Solving = Search + Propagation
A CP solver conducts search interleaved with propagation:

Familiar idea as in SAT solvers.
Propagate until fix point.
Make a choice.
Backtrack on failure.
Because we have global constraints we can often do more
propagation than unit-propagation of clauses at each step.

The A LL D IFFERENT constraint
Consider the n-ary constraint A LL D IFFERENT, with n = 4:
A LL D IFFERENT([a, b, c, d ]) (1)
n(n1)
Modelling: (1) is equivalent to 2 binary constraints:
a 6= b a 6= c a 6= d b 6= c b 6= d c 6= d (2)
Inference: (1) propagates much better than (2). Example:
a {4, 5}, b {4, 5}, c {3, 4}, d {1, 2, 3, 4, 5}
No domain pruning by (2).

n(n1)
a 6= b a 6= c a 6= d b 6= c b 6= d c 6= d (2)
a {4, 5}, b {4, 5}, c {3, 4}, d {1, 2, 3, 4, 5}

n(n1)
a 6= b a 6= c a 6= d b 6= c b 6= d c 6= d (2)
a {4, 5}, b {4, 5}, c {3, 4}, d {1, 2, 3, 4, 5}

n(n1)
a 6= b a 6= c a 6= d b 6= c b 6= d c 6= d (2)
a {4, 5}, b {4, 5}, c {3, 4}, d {1, 2, 3, 4, 5}
No domain pruning by (2). But perfect propagation by (1)

Propagator for A LL D IFFERENT

Solutions to A LL D IFFERENT([u, v , w, x, y , z]) correspond to
maximum matchings in a bipartite graph for the domains:
u {0, 1} u 0
v {1, 2} v 1
w {0, 2} w 2
x {1, 3} x 3
y {2, 3, 4, 5} y 4
z {5, 6} z 5


Mark all edges of some
maximum matching; Hopcroft-Karp
algorithm takes O(m n) time for n variables and m values:
u {0, 1} u 0
v {1, 2} v 1
w {0, 2} w 2
x {1, 3} x 3
y {2, 3, 4, 5} y 4
z {5, 6} z 5


Mark all edges of some
maximum matching; Hopcroft-Karp
algorithm takes O(m n) time for n variables and m values:
u {0, 1} u 0
v {1, 2} v 1
w {0, 2} w 2
x {1, 3} x 3
y {2, 3, 4, 5} y 4
z {5, 6} z 5


Mark all other edges in all other maximum matchings.
u {0, 1} u 0
v {1, 2} v 1
w {0, 2} w 2
x {1, 3} x 3
y {2, 3, 4, 5} y 4
z {5, 6} z 5


Mark all other edges in all other maximum matchings.
u {0, 1} u 0
v {1, 2} v 1
w {0, 2} w 2
x {1, 3} x 3
y {2, 3, 4, 5} y 4
z {5, 6} z 5


Every still unmarked edge is in no maximum matching.
Propagate accordingly within the current domains:
u {0, 1} u 0
v {1, 2} v 1
w {0, 2} w 2
x {1, 3} x 3
y {2, 3, 4, 5} y 4
z {5, 6} z 5


Every still unmarked edge is in no maximum matching.
Propagate accordingly within the current domains:
u {0, 1} u 0
v {1, 2} v 1
w {0, 2} w 2
x {1, 3} x 3
y {2, 3, 4, 5} y 4
z {5, 6} z 5

Search + Propagation in the PT example

2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 ?
1970
1966


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 ?
1970
1966
Alva cannot be a judge of 1977 as that would violate the

second constraint (every judge evaluates 3 wines).


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7
1970
1966

second constraint (every judge evaluates 3 wines).


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7
1970
1966

second constraint (every judge evaluates 3 wines). Actually,
Alva cannot be a judge of 1994, 1970, 1966 either, for the same
reason, and this was already propagated when trying the
search guess that Alva be a judge of 2000!


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7
1970 7
1966 7

second constraint (every judge evaluates 3 wines). Actually,
Alva cannot be a judge of 1994, 1970, 1966 either, for the same
reason, and this was already propagated when trying the
search guess that Alva be a judge of 2000!

PT: partial assignment

2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 ?
1970 7
1966 7


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 ?
1970 7
1966 7
Search guess: Dan is a judge of 1977. (3 guesses first.)


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 3
1970 7
1966 7
Search guess: Dan is a judge of 1977. (3 guesses first.)


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 3
1970 7
1966 7
Propagation: Dan cannot be a judge of 1970 and 1966 as

otherwise the second constraint (every judge evaluates 3
wines) would be violated for Dan.


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 3
1970 7 7
1966 7 7
Propagation: Dan cannot be a judge of 1970 and 1966 as

wines) would be violated for Dan.


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 3
1970 7 7
1966 7 7
Propagation: Eva, Jim, and Mia cannot be judges of 1977 as

otherwise the third constraint (every port pair has 1 judge in
common) would be violated.


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 3 7 7 7
1970 7 7
1966 7 7
Propagation: Eva, Jim, and Mia cannot be judges of 1977 as

otherwise the third constraint (every port pair has 1 judge in
common) would be violated.


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 3 7 7 7
1970 7 7
1966 7 7
Propagation: Leo and Ulla must be judges of 1977 as

otherwise the first constraint (every port is evaluated by
3 judges) would be violated for 1977.


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 3 7 7 3 7 3
1970 7 7
1966 7 7
Propagation: Leo and Ulla must be judges of 1977 as

otherwise the first constraint (every port is evaluated by
3 judges) would be violated for 1977.


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 3 7 7 3 7 3
1970 7 7
1966 7 7
Propagation: Eva must be a judge of 1970 and 1966 as

wines) would be violated for Eva.


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 3 7 7 3 7 3
1970 7 7 3
1966 7 7 3
Propagation: Eva must be a judge of 1970 and 1966 as

wines) would be violated for Eva.


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 3 7 7 3 7 3
1970 7 7 3
1966 7 7 3
Common fixpoint reached: No more propagation possible.


2011 3 3 3 7 7 7 7
2003 3 7 7 3 3 7 7
2000 3 7 7 7 7 3 3
1994 7 3 7 3 7 3 7
1977 7 3 7 7 3 7 3
1970 7 7 3 3
1966 7 7 3
Search guess: Jim is a judge of 1970. (3 guesses first.)
Propagation: etc.

Systematic Search Algorithm, with Propagation

1: post all given constraints, and propagate
2: while there is at least one suspended constraint do
3: pick a variable x with |domain(x)| 2
4: pick some value d domain(x)
5: try one of mutually exclusive guesses (constraints),
say x = d & x 6= d, or x > d & x d, and propagate
Heuristics
Line 3: variable ordering heuristic: smallest domain . . .
Lines 4 & 5: value & guess ordering heuristic: max, . . .
Tree exploration: depth-first, . . . , with backtracking
Example (PT search)

1 solve :: int_search(PT,input_order,indomain_max,complete)
2 satisfy;
History of CP
Stand-alone languages:
ALICE by Jean-Louis Lauriere, France, 1976
CHIP at ECRC, Germany, 1987 1990,
then marketed by Cosytec, France
OPL, by P. Van Hentenryck, USA, and ILOG, France:
front-end to both ILOG CP Optimizer and ILOG CPLEX
Comet, by P. Van Hentenryck and L. Michel, USA
MiniZinc, by the ORG group at CSIRO/NICTA, Australia
Libraries (the ones listed before ; are open-source):
Prolog: ECLiPSe, . . . ; SICStus Prolog, . . .
C++: Gecode, Google or-tools; IBM CP Optimizer, CHIP
Java: Choco, Google or-tools, JaCoP; . . .
Scala: OscaR; . . .

Outline
Features
Backends
Internals
5 Final Thoughts

Overview
MiniZinc is a declarative language to model combinatorial

problems using constraints.
Mainly developed at Nicta (Now centre 61), Australia

Introduced in 2007
Version 2.0 in 2014, with ongoing development.
Annual solver competition since 2008
http://www.minizinc.org

MiniZinc Features
Declarative language
High-level syntax
Solving technology independent
Solver independent
Separation of model and data
Open-source toolchain
Library of many predefined constraint predicates
Support for user-defined predicates and functions
Support for annotations

MiniZinc Backends
Constraint programming CP: Gecode, SICStus Prolog,

Choco, Google or-Tools, JaCoP, Mistral, . . .

MiniZinc Backends

Lazy-clause generation LCG: Opturion CPX, Chuffed, . . .

MiniZinc Backends

Mixed Integer programming MIP: G12/MIP, CPLEX,
Gurobi, . . .

MiniZinc Backends

Gurobi, . . .
Local search and metaheuristics LS: OscaR/CBLS,
EasyLocal++, YACS, . . .

MiniZinc Backends

Gurobi, . . .
SAT: G12/SAT, . . .

MiniZinc Backends

Gurobi, . . .
SAT: G12/SAT, . . .
SAT modulo theories SMT: fzn2smt with Yices 2,. . .

MiniZinc Backends

Gurobi, . . .
SAT: G12/SAT, . . .
Hybrids: iZplus, Minisat(ID), . . .

MiniZinc Backends

Gurobi, . . .
SAT: G12/SAT, . . .
Hybrids: iZplus, Minisat(ID), . . .
Portfolios of solvers: sunny-cp, . . .

MiniZinc Challenge 2015: Some Results

Winners per model (no portfolio or parallel categories):
Model Winner Technology
costas array Mistral CP
capacited VRP iZplus hybrid
gfd schedule Chuffed LCG
grid colouring MinisatID hybrid
instruction scheduling Chuffed LCG
large scheduling OR-Tools CP
application mapping JaCoP CP
multi-knapsack CPLEX MIP
portfolio design OscaR/CBLS LS
open stacks Chuffed LCG
project planning Chuffed LCG
radiation Gurobi MIP
satellite management Gurobi MIP
time dependent TSP G12/FD CP
zephyrus configuration CPLEX MIP
Solving a MiniZinc Model
Two phases:
1 Compile (or flatten) the model into a FlatZinc model.
2 Interpret the FlatZinc model with a solver.
FlatZinc is a low-level subset of MiniZinc:

No quantifiers or array comprehensions
Only built-in constraint predicates of the targeted solver
Only variables of types supported by the targeted solver
Constraints only on (arrays of) variables and parameters

Flattening
Unroll all quantifiers and array comprehensions
Inline all function and predicate calls
Evaluate all expressions not depending on the value of a
variable
Replace unsupported variable types
Decompose complex expressions,
introducing new variables
Flattening produces solver-specific models:

Not all solvers have the same set of built-in constraint
predicates
Different solvers may use different function and constraint
predicate definitions, also called decompositions

Example: A LL D IFFERENT
We have all ready seen that of a CP solver it is better to write:
Instead of writing:
constraint alldifferent(x);
instead of:
constraint forall(i in 1..n, j in i+1..n)
(x[i] != x[j]);
Even if your (constraint)-solver does not support

A LL D IFFERENTThere are strong advantages to using the
former:
More concise and clear model
Better solving, even for non-CP solvers

Decompositions of A LL D IFFERENT
Decompositions can be provided in the model or each backend

can provide its own decompositions.
Default Decomposition
predicate alldifferent(array[int] of var int: x) =
forall(i,j in index_set(x) where i < j)
( x[i] != x[j] );
Used, e.g., by the SMT backend fzn2smt
Built-in Predicate
predicate alldifferent(array[int] of var int: x);
Most CP solvers use a dedicated propagator.

Decompositions
Linear Decomposition for MIP

let {
array[int,int] of var 0..1: y = eq_encode(x)
} in forall(j in index_set_2of2(y))
( sum(i in index_set(x))(y[i,j]) <= 1 );
[...]
predicate equality_encoding(var int: xi,

array[int] of var 0..1: yi) =
sum(j in index_set(yi))(yi[j]) = 1
/\
sum(j in index_set(yi))(j * yi[j]) = xi;

Decompositions
Boolean Decomposition for SAT: Ladder Encoding

let {
array[int,int] of var bool: y = int2bools(x);
array[...,...] of var bool: a;
} in forall(i in ..., j in ...)
((a[i-1,j] -> a[i,j])
/\
(y[i,j] <-> (not a[i-1,j] /\ a[i,j])));
function array[int,int] of var bool: int2bools

(array[int] of var int: x) =
[...]

Decompositions
A LL D IFFERENT is merely one example of how a

(technology-specific) modelling idiom can be encapsulated
by a constraint predicate.

Decompositions

Solver-specific decompositions exist for many other
predefined constraint predicates.

Decompositions

Solver developers and researchers can add their own
decompositions for existing and new constraint predicates.

Decompositions

Solver developers and researchers can add their own
decompositions for existing and new constraint predicates.
Decompositions are transparent to the modeller.

Recent and Current Research
MiniZinc and its toolchain are continuously extended,

both by the core team at Monash University and the University
of Melbourne, and by other researchers around the world.
One can divide into three categories:

Tool support
Modelling extensions
Solving extensions

Tool Support
Interface from general-purpose programming languages

(coming soon)
Integrated development environment (since 2014)
Visualisation: CP search tree visualisation (CP 2015)
Globaliser: suggests global constraints to replace parts of
a given model (CP 2013)

Modelling Extensions
User-defined functions (CP 2013)

MiningZinc: extension targeted at itemset mining (CP
2013)
Modelling with option types (CPAIOR 2014)
Nested constraint programs (CP 2014)
Stochastic programming support (CP 2014)
CLPZinc: Prolog-based extension (PPDP 2015)
Auto pre-solving of predicates.
Ongoing work, strings.

Solving Extensions
OscaR/CBLS, EasyLocal++, YACS: local search backends

expand the solving capabilities (CPAIOR 2015, MIC 2015)
sunny-cp: MiniZinc makes it easier to design powerful
portfolio solvers (ICLP 2014, IJCAI 2015)
New decompositions: linearisation, Boolean encoding, set
variable elimination (CP 2015 and current research)
Search specification and scripting languages: search
combinators (CP 2011), MiniSearch (CP 2015),
annotations for local search (current research)
Ongoing, native string type of Gecode.

Outline

Motivation
Decision Variables
Constraints
Bounded-length Strings in CP
Native
Experimental Evaluation
5 Final Thoughts
String Constraint Problems

Constraints on strings occur in a wide variety of real-world
application areas, such as:
Web security: SQL code injection & XSS vulnerabilities
Program testing: program test-case generation
Program analysis
Model checking
Database testing
Interactive configuration: of bikes or radar systems, say
Biology: stem loops & pseudo-knots in bio-sequences
Image processing: scene analysis
Natural language processing: morphological analysis
Most of the work in the work is taken from [Joesph Scotts Ph.D.
Thesis].

Decision Variables
There are essentially three types of strings that are considered

in CP and in verification:
Fixed length strings,
Unknown but bounded length string,
Unbounded but finite length.
In the first two cases the domain of a string variable is a finite

set of strings. The last case requires decision procedures.
Even when the sets are finite we have to make approximations.

Constraints
Some of the constraints that we have considered, sj are string
variables, cj are character variables and ij are integer variables.
E QUAL(s1 , s2 ) if s1 and s2 are equal, that is s1 = s2

R EVERSE(s1 , s2 ) if s1 = c1 c2 cn and s2 = cn c2 c1
C ONCAT(s1 , s2 , s) if s1 s2 = s, with concatenation
S UB S TRING(s1 , i1 , i2 , s) if s1 [i1 : i2 ] = s
C HARACTER AT(s, i, c) if S UB S TRING(s, i, i, c )
L ENGTH(s, i ) if s has i characters, that is |s| = i
R EGULAR(s, R) if s is a word of a regular language R,
given by a regular expression or a finite automaton
C ONTEXT F REE(s, F) if s is a word of a context-free
language F, given by a context-free grammar
C OUNT(s, [c1 , ..., cn ], [i1 , ..., in ]) if in s all cj occur ij times
Constraints and Decision Variables
Note that the decision variables and the strings live in a

constraint solver not as a specialised solver but as first class
citizens. Therefore you get all the reasoning on integers for
free.

Bounded-length Strings
A bounded length is a string of a (possibly)-unknown that is

bounded from above by some implementation specific constant.
Possible implementations
Decompose are arrays of characters with a length variable
and a padding character (padded).
Implement special propagators to work with the padding
approach approach (aggregate)
Implement a bespoke variable type inside a constraint
solver (native).
We need padding characters because when a domain
becomes empty a CP solver will fail at that node and backtrack.

New data-types in CP
Implementing a new data-type can be quite difficult.

Choice of representation.
How to interact with the propagation loop
Changes in domains are signalled by events that form a
lattice.
A propagator subscribes to events to control how much
information and how often the propagator is woken up.
What exactly should we propagate?
A representation is an approximation of the mathematical
reality.
We have a galois-based framework for specifying
propagators and deriving what propagation should and can
be done in different representations.

Bounded-length Strings in CP
bounded-length sequence representation
A b-length sequence over-approximates a set of strings of

length b.
hhA[1], . . . , A[b]i, N i
Each A[i] is a set of characters. N in interval giving the lower
and upper bound of the string length.
Note that this is already an over approximation of a set of
strings.

Implementation 1: Aggregate Strings
Idea: represent the components as an aggregation of

pre-existing variable types (e.g., integers, sets, Booleans,
etc)
For example, a b-length sequence can be represented by:
hhA[1], . . . , A[b]i, N i ' hhD (X1 ) , . . . , D (Xb )i, D (N)i
Complications
undefinedness: empty variable domains failure!
representation invariant: implemented as an additional
constraint

Implementation 1: Aggregate Strings

Aggregate representation easy to implement in existing CP
solvers
logical derive formal
constraint propagator
semantics function
h, i-consistency
translate
translate
bounds(Z) consistency
domain consistency
indexical indexical solver-specific

constraint propagator propagator
checker synthesize description generate algorithm
Code generation and propagator synthesis:

[Monette et al @ CP12] and [Monette et al @ CP15]

Implementation 2: Native Strings
Three tightly related choices:
data structure
candidate lengths N: range sequence, bitset, interval
candidate characters N: range sequence, bitset, interval
sequence: array, list, list of arrays, etc
restriction operations must consider undefinedness

work by removing values from components
result is to remove strings from the domain
propagation events
representation invariant: many promising looking event
systems are not monotonic

Implementation 2: Native Strings
Three tightly related choices:
data structure
candidate lengths N: range sequence, bitset, interval
candidate characters N: range sequence, bitset, interval
sequence: array, list, list of arrays, etc
restriction operations must consider undefinedness

work by removing values from components
result is to remove strings from the domain
propagation events
representation invariant: many promising looking event
systems are not monotonic

Results
Three CP-based methods investigated
padded
fixed-length, pessimistically large array of integer variables
multiple occurrences of a padding character allowed at end
aggregate
native
2 benchmark sets from software verification:
Kaluza
Norn
even the padded method beats Kaluza [He et al @ CP 13]
Results
Aggregate implementation: 3 to 4 times speed-up over
padded
Native implementation: roughly 10 times speed-up over
padded

Outline

Lazy Clause Generation
Hybrids with MIP
5 Final Thoughts

Lazy Clause Advantages
CP solvers are good at propagation, SAT solvers are good at

clause learning. How can we combine them?
Not such a good idea: Compile global constraints into SAT
clauses. Resulting SAT problems are often too big.
Better idea: Generate clauses on the fly during search.
Even better idea: Have global constraints that also add
clauses during search that encode propagation.
State of the Art: Also have the global constraint give
higher-level reasons for failure, referred to as explanations
in the literature.
See the work (Chuffed) of [Peter Stuckey et. al. in Melborne]

Hybrids with MIP
Split a problem into sub-problems and solve some parts

with CP and some parts with Mixed integer programming
(MIP). With some care very good results can be obtained.
[Logic based Benders Decomposition. John Hooker]
Sometimes CP is not so good at propagating information

from cost functions. Use ideas similar to Lagrangian
Relaxation in CP and move problem constraints into the
cost function.

Final Thoughts
Please dont take this slide so seriously.

In CP we are less interested in decision procedures and
more interested in the actual solutions.
We are much better at finding solutions than proving
unsatisfiability.
We almost always consider finite domains this makes
integration with SMT much harder.
Because we consider only finite domains, decidability is
not a problem. We are not afraid of CFG intersection.
Sometimes in SMT equality propagation really pays off.
We have no real way of doing equality propagation.

Thank you
Questions?

References
To read about Strings and Galois Connections see:
J. Scott. Other Things Besides Number: Abstraction,
Constraint Propagation, and String Variable Types.
PhD thesis, Department of Information Technology,
Uppsala University, Sweden, March 2016.
http: // urn. kb. se/ resolve? urn= urn: nbn: se:
uu: diva-273311
The CFG constraint:

J. He. Constraints for Membership in Formal
Languages under Systematic Search and Stochastic
Local Search. PhD thesis, Department of Information
Technology, Uppsala University, Sweden, 2013
http: // urn. kb. se/ resolve? urn= urn: nbn: se:
uu: diva-196347

References
The work on propagator synthesis by Monette et.al. among

other things can be found at http:
//www.it.uu.se/research/group/astra/publications
To learn about Lazy Clause Generation look at the papers
of Peter Stuckey et.al. at http:
//people.eng.unimelb.edu.au/pstuckey/papers.html

Slides - Constraint Programming

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Slides - Constraint Programming

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Slides - Constraint Programming

Uploaded by

Copyright:

Available Formats

Constraint Programming

1st July 2016

Special thanks to Pierre Flener, Joseph Scott and Jun He

1 Introduction to Constraint Programming (CP)

3 Constraint Programming with Strings

4 Some recent, and not so Recent Advances

Uppsala University Justin Pearson Constraint Programming

Constraint Programming in a Nutshell

Uppsala University Justin Pearson Constraint Programming

1 Introduction to Constraint Programming (CP)

3 Constraint Programming with Strings

4 Some recent, and not so Recent Advances

Uppsala University Justin Pearson Constraint Programming

Uppsala University Justin Pearson Constraint Programming

Uppsala University Justin Pearson Constraint Programming

Uppsala University Justin Pearson Constraint Programming

Uppsala University Justin Pearson Constraint Programming

Example (PT integer model: 3 1 and 0)

Example (Instance data for the SMT PT instance)

Uppsala University Justin Pearson Constraint Programming

Example (Idea for another PT model)

Uppsala University Justin Pearson Constraint Programming

Example (PT set model: each port has a judge set )

Example (Instance data for the PT instance)

Uppsala University Justin Pearson Constraint Programming

Example (Sudoku model)

Global constraints such as A LL D IFFERENT and S UM enable the

Uppsala University Justin Pearson Constraint Programming

CP Solving = Search + Propagation

A CP solver conducts search interleaved with propagation:

Uppsala University Justin Pearson Constraint Programming

The A LL D IFFERENT constraint

Consider the n-ary constraint A LL D IFFERENT, with n = 4:

Inference: (1) propagates much better than (2). Example:

a {4, 5}, b {4, 5}, c {3, 4}, d {1, 2, 3, 4, 5}

No domain pruning by (2).

Uppsala University Justin Pearson Constraint Programming

The A LL D IFFERENT constraint

Consider the n-ary constraint A LL D IFFERENT, with n = 4:

Inference: (1) propagates much better than (2). Example:

a {4, 5}, b {4, 5}, c {3, 4}, d {1, 2, 3, 4, 5}

No domain pruning by (2).

Uppsala University Justin Pearson Constraint Programming

The A LL D IFFERENT constraint

Consider the n-ary constraint A LL D IFFERENT, with n = 4:

Inference: (1) propagates much better than (2). Example:

a {4, 5}, b {4, 5}, c {3, 4}, d {1, 2, 3, 4, 5}

No domain pruning by (2).

Uppsala University Justin Pearson Constraint Programming

The A LL D IFFERENT constraint

Consider the n-ary constraint A LL D IFFERENT, with n = 4:

Inference: (1) propagates much better than (2). Example:

a {4, 5}, b {4, 5}, c {3, 4}, d {1, 2, 3, 4, 5}

No domain pruning by (2). But perfect propagation by (1)

Uppsala University Justin Pearson Constraint Programming

Propagator for A LL D IFFERENT

Uppsala University Justin Pearson Constraint Programming

Propagator for A LL D IFFERENT