POMS AnliCaramanisPaschalidis

Journal of Manufacturing Systems 26 (2007) 116134
Contents lists available at ScienceDirect

Journal of Manufacturing Systems
journal homepage: www.elsevier.com/locate/jmansys
Technical paper
Tractable supply chain production planning, modeling nonlinear lead time and
quality of service constraints
Osman Murat Anli
a,
, Michael C. Caramanis
b,c
, Ioannis Ch. Paschalidis
b,c
a
Industrial Engineering Department, Isik University, Sile, Istanbul 34980, Turkey
b
Center for Information and Systems Engineering (CISE), Boston University, Boston, MA, USA
c
Department of Manufacturing Engineering, Boston University, Boston, MA, USA
a r t i c l e i n f o
Article history:
Received 15 November 2006
Received in revised form
29 January 2008
Accepted 2 May 2008
a b s t r a c t
This paper addresses the task of coordinated planning of a supply chain (SC). Work in process (WIP)
in each facility participating in the SC, finished goods inventory, and backlogged demand costs are
minimized over the planning horizon. In addition to the usual modeling of linear material flow balance
equations, variable lead time (LT) requirements, resulting from the increasing incremental WIP as a
facilitys utilization increases, are also modeled. In recognition of the emerging significance of quality
of service (QoS), that is, control of stockout probability to meet demand on time, maximum stockout
probability constraints are also modeled explicitly. Lead time and QoS modeling require incorporation
of nonlinear constraints in the production planning optimization process. The quantification of these
nonlinear constraints must capture statistics of the stochastic behavior of production facilities revealed
during a time scale far shorter than the customary weekly time scale of the planning process. The
apparent computational complexity of planning production against variable LT and QoS constraints
has long resulted in MRP-based scheduling practices that ignore the LT and QoS impact to the plans
detriment. The computational complexity challenge was overcome by proposing and adopting a time-
scale decomposition approach to production planning, where short-time-scale stochastic dynamics are
modeled in multiple facility-specific subproblems that receive tentative targets from a deterministic
master problem and return statistics to it. A converging and scalable iterative methodology is
implemented, providing evidence that significantly lower cost production plans are achievable in a
computationally tractable manner.
2008 The Society of Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.
1. Introduction
1.1. Motivation and objectives
Modern manufacturing enterprises are becoming more global
than ever. They encompass owned or contract manufacturing
and transportation facilities, suppliers, distributors, and customer
service centers scattered over the globe. Manufacturers are no
longer the sole drivers of the supply chain (SC). A shift from
a push to a pull environment is well on its way. Customer
needs and preferences influence the SCs inner workings: product
functionality, quality, speed of production, timeliness of deliver-
ies, flexibility in adjusting to demand changes. In todays highly
competitive marketplace, companies are challenged with achiev-
ing shorter order-to-delivery times while allowing customers to
Corresponding author.
E-mail addresses: omanli@isikun.edu.tr (O.M. Anli), mcaraman@bu.edu
(M.C. Caramanis), yannisp@bu.edu (I.Ch. Paschalidis).
customize their orders. Manufacturers recognize the significance
of short lead times (LT) and high quality of service (QoS) pro-
visioning for control of stockout probability. Furthermore, time-
based competition has had a significant impact on the design of
production facilities (product cells) and their operation (just in
time, zero in-process inventory, lean manufacturing, and so on).
Finally, supplierconsumer information sharing has been looked
on as a means to reduce inventories needed to provide a desired
service level. Although these efforts, together with the wider use of
enterprise-wide transactions databases, have achieved remarkable
productivity gains, further improvements in global SC lead times
and QoS are critically required. The revolution in computational in-
telligence and communication capabilities, assisted more recently
by the emergence of sensor networks with dynamically reconfig-
urable topology, has brought these improvements within reach.
The lead time at each link of a SC contains information that is
critical for effective coordination. Lead times change across weeks
in the planning horizon. In fact, they vary nonlinearly with load,
production mix, lot sizes, detailed scheduling, and other opera-
tional practices adopted during each week of the planning hori-
zon. Nevertheless, widely used material requirements planning
0278-6125/$ see front matter 2008 The Society of Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.jmsy.2008.05.001
O.M. Anli et al. / Journal of Manufacturing Systems 26 (2007) 116134 117
(MRP) systems assume that lead times are constant across the
whole planning horizon to avoid the task of estimating and com-
municating variable lead time information. The use of limited
information in the current state-of-the-art industrial practice is
responsible for inefficient planning and often chaotic and unsta-
ble operations hampered by chronic backlogs and widely oscillat-
ing inventories. Two major barriers preventing more extensive use
of information are (i) the cost of collecting, processing, and com-
municating the requisite information and (ii) computational and
algorithmic challenges in using this information to plan and man-
age SCs optimally. A time-scale decomposition and information
communication architecture framework is proposed that is capa-
ble of exploiting sensor networks, and overcome the communica-
tion barrier. An iterative decentralized coordination algorithm is
also proposed that provides proof of the concept that the compu-
tational barrier can be overcome as well.
1.2. Current industry practice
Whereas capacity is ignored and dynamics are modeled by con-
stant lead times in the vanilla version of materials requirement
planning (MRP) models [1], advanced planning system (APS) ap-
proaches include adequate representationof material flowdynam-
ics and detailed representation of effective (or expected) capacity.
APS models rely on mathematical programming techniques and
hierarchical decomposition [2,3] to overcome combinatorial com-
plexity explosion barriers while capturing the details of capac-
ity restrictions. This task is particularly onerous in the face of
discrete part integrality and complex production rules and con-
straints, which together with uncertainty, render stochastic inte-
ger programming formulations computationally intractable.
Past approaches employed to bypass these hurdles include the
theory of constraints, scheduling algorithms, and fluid model ap-
proximations. The theory of constraints [46] approximates the
model of the production system by a small number of bottleneck
components that are modeled in great detail; production is sched-
uled around those components through constraint propagation
over time. Two main shortcomings of the theory of constraints ap-
proach are: first, the difficulty in identifying and modeling bottle-
neck components, and second, the fact that delays or lead time
dynamics along part routes are nonlinear and difficult to model.
The systematic modeling of individual facilities could be possibly
used to alleviate the first shortcoming. However, it is very diffi-
cult to overcome the second shortcoming. A variety of schedul-
ing algorithms ranging from mathematical programming and
Lagrangian relaxation to genetic algorithms have been used, of-
ten effectively [715]. Fluid model approximations have also been
used extensively and with considerable success [1621] but have
not adequately addressed dynamic lead time modeling. Systemdy-
namics simulation models have been proposed [22] that capture
nonlinearly increasing lead times as functions of the production fa-
cility utilization. It has been shown that deterministic fluid model
approximations of stochastic discrete production networks can be
employed to predict the qualitative nature of optimal scheduling
rules [8,9,16,18] and to determine the stability and robustness of
the approximated stochastic discrete networks [2326]. The pro-
posed algorithm exploits this line of research with particular em-
phasis on extending fluid network approximations to improve the
dynamic lead time modeling capabilities.
Past efforts to model lead time in production planning are note-
worthy [27,28] but are limited to static lead times estimated for
average or typical production conditions. Incorporating dynamic
lead times into production planning poses modeling analysis and
computational difficulties leading to deliberate choices on simpli-
fying approximations and relaxations. A mixed-integer production
planning model has been proposed [29] that employs piecewise-
linear functions to capture the effect of alternative routings and
subcontractors on load-dependent lead times. The quantitative re-
lationships between work in process and production are estimated
via Monte Carlo simulation and used as constraints in a nonlin-
ear production planning model [30] solved through linearization.
The effect of inventory on the quality of service has been also
studied in an uncapacitated single-product multi-class-QoS supply
chain through queuing approximations [31] and in a multi-product
single-facility fixed lead time setup [32]. A recent literature sur-
vey [33] provides an extensive overview of dynamic lead time
modeling in production planning and points out the use of the
aforementioned nonlinear relationship in supply chain production
planning. This paper explores further in that direction.
1.3. Overview of the proposed approach
The time-scale-driven decentralized information estimation
and communication architecture that are proposed in Section 2
enable coordination, planning, and operational decisions of
manufacturing cells, transportation activities, inventory, and
distribution facilities in a SC. It is shown that this can be achieved
through optimal and consistent production targets and safety-
stock levels scheduled for each part type produced by each SC
facility. Proposed is a framework of iterative information exchange
between three decision-making/performance-evaluation layers
that is indeed capable of achieving this coordination. The
framework consists of a centralized planning coordination layer,
a centralized QoS coordination layer, and finally a decentralized
performance evaluation and demand information layer. The
planning layer determines facility-specific production targets
using performance and sensitivity information it receives from
the decentralized performance evaluation and information layer.
The QoS layer combines interacting facility production capabilities
and requirements (that is, targets) to determine hedging inventory
requirements that achieve exogenously specified QoS levels. The
decentralized performance evaluation and demand information
layer analyze short-term (hourly) stochastic dynamics of each
facility to derive expected (weekly) work-in-process and safety-
stock inventory for each facility and their sensitivity w.r.t. planning
level targets.
The major objective of the proposed framework is to capture
second-order effects of the steady-state cell dynamics in order to
model dynamic lead time effectively at the coarse (varying weekly)
production planning dynamics layer. Weekly time averages are
a statistic with relatively low variance due to the law of
large numbers effects, and they can be effectively modeled as
deterministic quantities within the planning layer. Furthermore,
detailed information on machine-specific queue and setup states
is not globally available, hence, it is practical to share state
information that is (i) time averaged to the coarse time scale and
(ii) grouped by facility. To this end, capacity, work-in-process,
and production requirements are facility-specific aggregates.
Production planning dynamics are thus constrained to satisfy
minimum weekly average lead time requirements. Note that
although facility lead times and interfacility hedging inventory
requirements are averages over the fine (hourly) time-scale
dynamics modeled at the decentralized performance evaluation
layer, they are dynamic relative to the coarse (weekly) time
scale of the planning layer. Lead times and hedging inventory
requirements are modeled as functions of production planning
decisions (loading and mix). This constitutes the second-order
information that has been shown can be used [34] to significantly
decrease inventory and backlog costs. The planning coordination
layer employs an iterative interaction of a single production
planning master problem on the one hand with the hedging
118 O.M. Anli et al. / Journal of Manufacturing Systems 26 (2007) 116134
policy QoS layer and the performance evaluation layers multiple
decentralized facility-specific subproblems on the other.
The effectiveness of our planning layer model depends crucially
on the quality with which the operational dynamics of the
production facilities are modeled in the performance evaluation
and information layer. To this end, the framework relies on the
following two building blocks:
1. Dynamic lead time modeling: Performance analysis results
for stochastic queuing networks are used to accurately estimate
average weekly lead times as functions of capacity utilization,
productionmix, productionpolicies, anddistributions of stochastic
disturbances such as failure and repair times. This provides
delivery requirements to upstream facilities and available supply
to downstreamfacilities, whichare necessary for efficient planning
of production over a multi-week horizon. These nonlinear lead
time functions, denoted by g (), are incorporated as weekly
constraints on decision variables in production scheduling.
2. Provisioning of quality of service (QoS) guarantees: Constraints
are introduced that bound the probability of backlog at a SC
facility. It is believed that probabilistic constraints reflect customer
satisfaction considerations and followclosely the industry practice
of providing QoS guarantees. These guarantees are modeled as
nonlinear constraints in the production scheduling framework,
denoted by

h ().
The main purpose of this article is to demonstrate that dynamic
lead times and hedging inventory requirements can be modeled
and included in the tractable determination of faster SC production
plans while maintaining the desired quality of service guarantees.
The aim is to provide a proof of concept regarding the feasibility
of modeling dynamic lead times and quality of service guarantees
as part of the production planning process. Through a variety
of numerical examples, the potential cost savings achievable by
the proposed approach are studied relative to traditional constant
lead time based production planning approaches that represent
the bulk of todays industry practice. The comparison supports
the viability of implementing the approach in real life provided
that the cost of estimating and processing the required lead time
information is affordable. It is not claimed that the proposed model
is a perfect model of reality, particularly as far as the decentralized
queuing network subproblem model is concerned. More general
and accurate models have been developed in the queuing network
and simulation literature, and undoubtedly further extensions are
forthcoming from the formidable and research-active community
studying these topics. Moreover, the adoption of radio frequency
identification tag (RFID) and sensor network technologies will also
contribute to the affordability of dynamic lead time information.
This contribution is for showing how this information can be used
in a tractable, computationally efficient, and robust production
planning algorithm.
The demonstration of significant reduction in inventory costs
when the nonlinear relationship of facility lead times is modeled in
the SC production planning process is not the major contribution
of this paper. Most practitioners will argue from experience that
this is hardly surprising given the widely observed inadequacy of
MRP-based production schedules that rely on the constant lead
time assumption. The major contribution of this paper is in its
proposal and implementation of a practical, efficient, tractable, and
robust algorithm capable of actually achieving these cost savings.
The aim is to prove the concept that SC production planning on
constant lead times is not a necessary evil imposed by the incorrect
presumption of insurmountable computational complexity. In fact,
it is claimed that SC planning no longer has to live with the
undesirable consequences of the constant lead time assumption
impeding todays industry practice. This contribution supports the
notion that detailed production facility models and/or adoption
of RFID technologies can provide additional value added through
their ability to extract reliable, albeit affordable, dynamic lead
time information and make it available to the production planning
optimization process.
The next section introduces the proposed time-scale-driven
data communication architecture. The SC problem and the
performance evaluation, QoS, and planning layers are then
described, following by computational experience that shows
the value of dynamic lead time and probabilistic QoS constraint
information in the determination of a SCs coordinated production
schedule. A three-facility SC producing five different part types
is used to develop various representative examples of SCs.
Comparison to production schedules that are characteristic of
current industry practice indicates that substantial improvements
are possible.
2. Time-scale-driven decentralized data communication and
decision support architecture
The multitude of strategic, planning, and operational decisions
made routinely by SC participants are far too complex and the
requisite information is far too large to handle in a centralized
manner. Decentralized decision making has therefore been the
norm. However, since the consequences of various decisions
are interdependent, it follows that appropriate coordination can
foster desirable efficiencies. Consider a decentralized decision-
making agent as a decision node in a network of communicating
decision nodes. A key determinant of successful coordination is
the systematic conversion of data available at a certain decision-
making node i to a compact representation of information
relevant to the decision-making process at node j. Relevant is
construed here to mean incorporating all information about the
state, dynamics, and decision policies in node i that may contribute
to efficient decision making in node j. Compact representations
of relevant information may take, for example, the form of a
statistic: the time-averaged lead time in a production system, the
probability distribution and autocorrelation of a demand process,
or a performance target, such as the desired weekly output of a
manufacturing process. These compact representations provide
key enabling efficiencies in both the estimation of the relevant
information (which can be done in a decentralized distributed
manner) as well as in its communication (the transmission of a
statistic requires less bandwidth and energy than the time series
it describes). Although several issues are still to be resolved,
intelligent communicating mobile sensor networks have the
potential to both estimate and communicate relevant information
in ways that are superior to conventional alternatives in terms of
cost, flexibility, and reliability.
Proposed is a time-scale-driven assignment of SC decisions to
nodes that is suggestive of the relevant information exchange
architecture. The idea of time-scale-driven decomposition is not
new. In fact, it has been widely used to great advantage in control
theory [35]. The main idea here is the fact that decisions are
characterized by a characteristic frequency and its corresponding
time scale. For example, while machine operating decisions are
made every few minutes, major resource acquisition decisions
are made every few months or years. It is further noticed that
supply chain decisions characterized by functionality (for example,
resource allocation, planning, sequencing) and scope (for example,
enterprise, plant, cell, process) are associated with a decreasing
time scale as the scope narrows and the functionality changes
from resource allocation to sequencing. Table 1 provides such a
classification example where time scales decrease as the decision
of interest moves to the southeast.
The SC planning algorithm proposed here employs a decen-
tralized decision-making and information exchange architecture
Table 1
Example of time-scale-driven classification
that is an instantiation of the time-scale-driven approach of
Table 1. Fig. 1 presents the information exchange architecture that
supports factory-scope production planning decisions, cell-level
performance evaluation, and process-level operation control. Note
that:
The factory production planning node passes down weekly
production targets to each cell.
Each cell evaluates its performance during each week in the
planning horizon, determines variability distributions, and
aggregates its hourly dynamics to weekly time averages of
relevant performance measures such as work in process (WIP),
lead time (LT), and their sensitivity with respect to weekly
production targets passed down from the factory planning
node.
Contiguous cells coordinate horizontally to determine safety
inventory of semi-finished and finished goods that assures
desiredquality of supply levels at eachcell andquality of service
to customers.
Each cell communicates weekly averages and sensitivities
up to the factory planning node and variability distributions
horizontally to upstream and downstream cells.
Using WIP, LT, and sensitivity information, the factory planning
node adjusts production targets so as to achieve material flows
across cells that meet required WIP and safety-stock levels
while minimizing SC WIP and LT.
While the remainder of the paper proposes algorithms that
can make practical use of the information flow described
above and reach a stable and optimal production plan, it
must be emphasized that the proposed information architecture,
in addition to distributing computational effort (performance
evaluation and handling of short-time-scale stochastic dynamics
modeling are done in a decentralized manner at each cell) also
reduces communication requirements to the relevant information.
For example, the factory planning node does not need to know
the cell production details: labor and other resources available,
machine capacities, and manufacturing process specifics. It needs
to know, however and it does know the weekly lead times
at each cell and the hedging inventory between cells that are
consistent with the production targets that the planning node
sends to each cell.
The general philosophy of the time-scale-driven communica-
tion and decision support architecture described in this section
provides useful guidelines but not a mindless recipe. In the rest
of this paper, these guidelines are used to propose a SC production
planning algorithm that is computationally tractable and outper-
forms two proxies of the state of the art in industry practice that it
is compared to.
Fig. 1. Example of information architecture.
3. Supply chain management problem
This section develops a supply chain (SC) planning algorithm
that utilizes the principles of the time-scale-driven architecture
discussed above. Following anoverview, the three layers employed
are described in detail, as well as their interaction in providing the
optimal SC production plan.
3.1. Supply chain problem overview
To describe our SC model and establish the notation, the system
depicted in Fig. 2 is considered, and the associated information
exchange and decision layers are shown in Fig. 3. Although a tree
network of SC links or facilities can be modeled, C production
facilities connected in series are considered here, for ease of
exposition but without loss of generality. Production planning
decisions and the resulting WIP and QoS hedging inventory
requirements vary in the medium term (say, across weeks), and
the characteristic scale of their dynamics is called a period and
denoted by t {1, 2, . . . , T}. On the other hand, performance
evaluation and demand dynamics vary many times within a period
(say, across hours) and their characteristic scale is called a time slot
denoted by the subscript k of k {1, 2, . . . , K}.
The QoS layer determines a hedging point or safety-stock
inventory policy at each facility c, which guarantees that the
probability of stockout or starvation of facility c 1 does not
exceed 1
c
(t), where
c
(t) is the quality of service that facility
c offers to facility c 1. In other words, the probability that the
material release requirements of facility c 1 are met on time
equals or exceeds
c
(t). The QoS layer models random behavior
of short-term facility production capacity and final demand, while
the planning layer models expected values or time averages during
Fig. 2. A multi-class supply chain with limited production capacity at each facility.
Fig. 3. Information exchange among the coordination layers and the decentralized layer.
a period (for example, a week) under the underlying assumption
that the period is long enough for the time-slot stochastic process
dynamics to reach steady state.
External demand is met from the available finished goods
inventory in facility 1, and it is backordered if finished goods
inventory (FGI) is not available. Every facility c {1, 2, . . . , C}
produces a set of products and has a limited production capacity.
Whereas facility C can draw from an infinite pool of inventory,
production of facilities (C 1, . . . , 2, 1) is constrained by the
production capacities of workstations and in addition by the work
in process (WIP) in each facility. WIP is in turn constrained by
(i) the FGI available upstream to release input into a facility that
replenishes WIP and (ii) the production that depletes WIP. Again,
for ease of exposition and without loss of generality, it is assumed
that all facilities process the same set of part types, {1, 2, . . . ,
P}, and one part of part type p is required from an upstream
facility c + 1 to produce one unit of the same part type at
facility c. The serial SC problem presented in Fig. 2 retains the
most salient features of the more general problem, particularly in
terms of the general demand and service distributions allowed. As
suggested by past experience in the literature [3638], results for
the simpler system can be routinely generalized to accommodate
assembly/disassembly features.
D
p
(t
k
) denotes the amount of external orders for product of
class p arriving during time slot k of period t.
p
c,m
(t
k
) denotes the
part type p production capacity in isolation of facility c workstation
m during time slot k of period t. X
p
c
(t
k
) denotes the number of
type p parts facility c produces during time slot k of period t.
R
p
c
(t
k
) is the amount of WIP released into facility c from facility
c + 1 FGI. Finally, Q
p
c
(t
k
) denotes the type p WIP at facility c, and
I
p
c
(t
k
) denotes the type p FGI at facility c available at time slot k
of period t. Only at facility 1, the FGI is allowed to take negative
values to denote backordering. Following standard conventions,
_
I
p
1
(t
k
)
_
+
and
_
I
p
1
(t
k
)
_
denote, respectively, max

_
0, I
p
1
(t
k
)
_
and
max
_
0, I
p
1
(t
k
)
_
.
Because the period containing K time slots is the relevant time
scale in the planning layers dynamics, and because it is assumed
that it is long enough for the stochastic processes active at the
time-slot scale to reach steady state, the following time-averaged
variables are defined:

R
p
c
(t) =
1
K
K
k=1
E
_
R
p
c
(t
k
)
_
, and similarly
X
p
c
(t),

Q
p
c
(t),

I
p
c
(t),
p
c,m
(t), and

D
p
(t). We use vector notation
X
c
(t) =
_
X
1
c
(t), . . . ,

X
P
c
(t)
_
.
The SC management problem is implemented in three layers,
exchanging information as shown in Fig. 3 and described below.
3.2. Performance evaluation and information layer
The performance evaluation and information layer shown in
Fig. 3 models the short-term stochastic dynamics of production
facilities at the operational level and develops the steady-state
or time-averaged performance measure estimates of interest at
the longer time scale of the planning layer. More specifically,
performance evaluation means:
1. The transformation of production targets in each period to
estimates of minimum average WIP required during that period
in each facility to meet the production targets set by the planning
layer. This estimate will generally depend on production targets,
X
c
(t), the probability distribution of all relevant random variables
P
c
(t), and other operational policies,
c
(t), during that period. The
mapping of these inputs to the average WIP in facility c,

Q
c
(t), is
implicitly represented by function g
p
c
_
X
c
(t), P
c
(t),
c
(t)
_
.
2. The estimation of sensitivities (or derivatives) of g
p
c
()
with respect to production targets. This is needed for tractable
representation of the highly nonlinear relationship embodied in
the g
p
c
() function.
3. The transformation of production targets, hedging inven-
tory levels, w
c
(t), and operational policies to the minimum
average FGI required to meet the QoS constraint. The min-
imum average FGI requirements are represented by function
h
p
c
_
X
c
(t),

X
c1
(t), P
c
(t), P
c1
(t), w
c
(t),
c
(t)
_
. For purposes of
demonstrating the concept of dynamic lead times associated with
dynamic QoS guarantee provisioning, a limited pairwise coupling
of upstream and downstream facilities presented in Section 3.3 is
considered here.
4. The estimation of sensitivities (or derivatives) of the function
h
p
c
() with respect to production targets. Again, to serve the proof
of concept objective, relatively simple analytic approaches are used
for the determination of

h
p
c
() and its sensitivity requirement (see
Section 3.3).
5. A representation of an aggregate probabilistic model of
facility c short-term capacity availability for use by the QoS layer.
Whereas this can be in general a Markov-modulated process
(MMP) model, a simple, weighted bottleneck machine capacity
exponential model (see Section 3.3) is used here, again to capture
the correct factory physics and demonstrate the proof of concept
in the planning layer algorithm. In practice, MMP models for
production cells as well as for final demand can be estimated
by analyzing possibly autocorrelated production and shipment
transaction databases [39].
The estimates produced by the performance evaluation algo-
rithm are merely intended to demonstrate qualitatively appro-
priate behavior, because the objective is to concentrate on an
iterative planning layer. In real applications, the performance eval-
uation and information layer can be implemented using more ac-
curate approaches in a distributed/decentralized manner where
efficiency and robustness are important but not crucial.
Production target decisions are realizable at the desired QoS
level only if the requisite WIP and FGI are available at various
facilities in a manner consistent with material conservation
dynamics. Functions g
p
c
() and

h
p
c
() establish the minimum WIP
and FGI constraints employed at the planning layer discussed in
Section 3.4.
The evaluation of g
p
c
() and

h
p
c
() functions is a formidable
task. Analytic models have been used (mean value analysis [40]
in this paper and the queuing network analyzer elsewhere [34])
to quantify them and introduce their nonlinear characteristics
explicitly in an iterative decomposition planning algorithm for
the purpose of investigating convergence properties of the multi-
layer interactions depicted in Fig. 3. However, it is recognized that
Monte Carlo simulation or complex analytical models involving
Markovian or even more general stochastic decision processes may
be relevant and more accurate in practice and may be indeed used
in place of the more convenient models that were selected for
the purpose of proof of concept, and without loss of generality
since the qualitative behavior of the models is similar to that of
the more accurate models that may be selected in practice. The
SchweitzerBard Approximation [41,42] is used with our mean
value analysis algorithm, which enables calculation of g () values
for real valued

X
c
(t) and

Q
c
(t) vectors. Similar fluid approximation
enhancements as those used in the deterministic algorithms of
the planning layer are also relevant in the context of stochastic
models used at the decentralized layer [43,44]. These extensions
are not trivial. For example, key events that are responsible
for the efficiency of event-driven simulation algorithms (e.g., a
buffer fills or a buffer empties) proliferate (a buffer fills or
empties partially with multiple partial full/empty states) requiring
more sophisticated models [45]. The important advantage of
fluid production stochastic models (whether simulation based or
analytic) is their ability to provide sensitivity estimates more
tractably than finite differencing of stochastic discrete production
models.
Finally, the convexity of the feasible regions defined by the g ()
and

h () functions is crucial to the convergence of the planning
layer. Fig. 4 shows a realistic example of the feasible region
boundaries for a two-part type stochastic production network.
More specifically, the maximum value of

X
1
c
(t) subject to

Q
1
c
(t)
g
1
c
_
X
c
(t) , P
c
(t),
c
(t)
_
is plotted versus

Q
1
c
(t) and

X
2
c
(t).
Although the above constraints exhibit generally convex feasi-
ble regions, nonconvex feasible regions have been observed that
arise when either operational policies are flagrantly suboptimal
or facility designs are far from homogeneous (e.g., product classes
impose diverse production time requirements on facility work-
stations) [46]. Consider Fig. 5 depicting mildly nonconvex and
severely nonconvex feasible regions in contrast to the convex ex-
ample in Fig. 4. Robust iterative master problem subproblem al-
gorithms have been constructed that converge even under rather
severe nonconvexity conditions [47].
Fig. 4. Feasible production

X
p
c
(t) as a function of WIP

Q
p
c
(t).
3.3. Quality of service coordination layer
The QoS Layer interacts with other layers, as shown in Fig. 3.
Its objective is to estimate a production policy that achieves
the desired probabilistic QoS guarantees. Accurate modeling of
QoS coordination policies is an important research problem in
itself being monitored [4850]. A number of methodologies have
been used, including multi-class queuing network analysis (QNA),
Monte Carlo simulation, stochastic system approximations, and
large deviations asymptotics.
Given the stated objective to demonstrate the ability to
construct a robust and efficient planning layer, a simple but
certainly near-optimal [9] hedging policy is elected, which works
as follows: Facility c produces at full capacity as long as the amount
of work in its output buffers,
p
c
(t)I
p
c
(t), where 1/
p
c
(t) is the
bottleneck capacity of facility c for part type p, is belowthe hedging
inventory level expressed in units of work, w
c
(t), set for week t
by the QoS layer. The hedging inventory level is selected by the
QoS layer so that the probability of a stockout of the downstream
facility c 1 does not exceed the desired level 1
c
(t). The idea
is implemented using the following model:
1. The desired stockout probabilities, 1
c
(t), at intermediate
FGI positions and the final demand (c = 0, 1, 2, . . . , C and
t = 1, 2, . . . , T) are determined exogenously by the SCplanner.
2. The hedging inventory level is estimated so as to achieve a
maximally allowedstockout probability specifiedfor facility c
1 as 1
c
(t) under item 1 above by using the large deviations
approach described in [49] and summarized in 5. This approach
provides an efficient and accurate method for determining the
parameters of a hedging point policy and the associated average
inventory of semi-finished goods inventory in the output buffer
of facility c as a function of the QoS required at facility c 1, and
the first twomoments of (i) the effective service time of facility c
and(ii) the effective demandfor material release into facility c
1. The associated average inventory in

I
p
c
(t) is then estimated
by a G/G/1 approximation of the interaction of facilities c
and c 1 where each multi-machine facility is approximated
by a fictitious single machine with a general service time
distribution. 5 describes the modeling of the hedging inventory
requirements, including the special case of c = 1.
3. The simulation quantified the functional relationship between
QoS and the coefficient of variation. The first two moments
of the effective interrelease times of parts processed by each
facility c and released into FGI c are estimated through
simulation for a set of representative production targets and
hedging points. To this end, multiple Monte Carlo simulation
runs are employed as follows:
Fig. 5. Examples of mildly nonconvex and severely nonconvex feasible regions.
Fig. 6. Monte Carlo simulation of two-facility, one-part type SC.
(i) Selected production target values for each part type
produced in facility c,

X
p
c
(t), are fixed as inputs. In each
simulation run, the production targets across facilities are
set equal to each other. The approximate mean value
analysis (MVA) algorithm that was used to calculate g ()
in Section 3.2 is employed here again to determine the
required constant work-in-process (ConWIP) vector K
c
(t)
that guarantees facility c can produce in isolation at an
average rate

X
c
(t) =
_
X
1
c
(t),

X
2
c
(t), . . . ,

X
P
c
(t)
_
.
(ii) The SC is simulated for a range of hedging point w
c
(t)
values that correspond roughly to QoS levels in the range
of 80%99%. Each facility is modeled as a fixed routing
proportion queuing network with the material release
protocol described below using as a key parameter the
MVA-calculated ConWIP vector K
c
(t).
(iii) The system of Fig. 6, where q
c
(t
k
) is defined as a bin
holding fully processed parts remaining inside the facility,
is thensimulatedwiththe following material release policy
described for simplicity for the special case of a one-part-
type SC:
if Q
c
(t
k
) +q
c
(t
k
) < K
c
(t) +w
c
(t)/
c
(t)I
c
(t
k
), where
x is the smallest integer greater than x, then facility
c absorbs material from I
c+1
(t
k
) until there is equality
in the expression above or until I
c+1
(t
k
) empties. (Note:
under this rule it is possible to temporarily accumulate
parts inside facility c that exceed K
c
(t))
when a parts processing is completed in facility c, the
facility
proceeds to increment I
c
(t
k
) if I
c
(t
k
) < w
c
(t)/
c
(t)
or it proceeds toincrement q
c
(t
k
) if I
c
(t
k
) = w
c
(t)/
c
(t)
with contents q
c
(t
k
) counted as part of WIP. Note that
q
c
(t
k
) = 0 when I
c
(t
k
) < w
c
(t)/
c
(t).
(iv) Interarrival times into I
c
(t
k
) are sampled conditional
on I
c
(t
k
) < w
c
(t)/
c
(t) and used to estimate the
conditional variance of effective service times of facility c
denoted by
c
(t). Conditional variance values are stored in
a table entry together with the value of exogenous input
quantities used in that simulation.
Several simulations are performed, each corresponding to
different values of exogenous inputs of facility production
rate targets and hedging points. Each entry of the resulting
table stores a value of the squared coefficient of variation
of interrelease times into I
c
(t
k
). Note that this is a function
scv
c
_
c+1
(t),

X
c
(t) ,
c
(t), P
c
(t),
c
(t)
_
with the argument
list showing the significant dependencies. For simplicity
is written scv
c
(t) = scv
c
(
c+1
(t) ,
c
(t)), where
c
(t)
represents the utilization level of facility c, defined more
precisely later. The table is in effect a representation of the
function scv
c
(
c+1
(t),
c
(t)), whose values canbe interpolated
from the table entries.
Simulation results verify the a priori expectation that
scv
c
(
c+1
(t),
c
(t)) is increasing in

X
c
(t) and decreasing in
c+1
(t). Associating the largest index facility with the raw
material vendor, a range of vendor hedging point values is
simulated for the same hedging point values at the remaining
facilities. Because each hedging point value at the vendor
results in a different QoS level for the production facility that
the vendor supplies, the impact of QoS on that facility can
be calibrated. In general, the tabulated results were able to
fit a smooth nonlinear function that represents the behavior
of scv
c
(t), which is then used to model the nonlinear QoS
constraints in the planning coordination layer described in
Section 3.4. Fig. 7 graphs the scv
c
(t) function whose coefficients
are estimated to fit the simulation table entries. Its algebraic
representation is:
scv
c
(t) = 12.339 (
c
(t))
3
25.522 (
c+1
(t))
3
+26.205 (
c
(t))
2
c+1
(t) 19.602
c
(t) (
c+1
(t))
2
+85.276 (
c+1
(t))
2
+ 1.722
c
(t)
c+1
(t)
82.355
c+1
(t) + 27.425.
For simplicity, and because the purpose of the paper is to show
that optimal production planning optimization can account
for nonlinear QoS constraints, the reasonable approximation is
employed that the scv
c
(t) depends significantly only on
c+1
(t)
and
c
(t), while dependence on the utilization or QoS of other
facilities is negligible. This assumption is indeed supported by
simulation experience.
Fig. 7. Smooth interpolation of simulation results of hedging inventory replenish-
ment process coefficient of variation.
4. The SC production planning algorithm described in Sec-
tion 3.4 employs interpolation to determine variance val-
ues. The large deviations results of [49] are then used to
determine w
c
(t) values as a function of scv
c
(
c+1
(t),
c
(t)),
scv
c1
(
c
(t),
c1
(t)), and
c
(t). The G/G/1 approximation is
finally employed to determine the required average inventory
levels

I
p
c
(t).
5. Recall that the function

h
c
(
X
c
(t),

X
c1
(t), P
c
(t), P
c1
(t), w
c
(t),
c
(t)) represents the minimum average FGI requirements that
the planning layer must observe in order to model dynamic lead
time and QoS requirements. This average must hold both dur-
ing the fast time-slot scale that characterizes production ca-
pacity variations (for example, failures and repairs) as well as
during the coarse time scale of the production planning func-
tion. The approach in [48] is adopted and the fast stochastic
dynamics are modeled with a discrete Markov-modulated pro-
cess model whose characteristic time scale is a day. The results
in [48] are then used to obtain the hedging point required to
obtain the desired QoS level as a multiple of the average daily
production capacity. For a QoS level that is reasonably close to
one, a good approximation for

h
c
() is a quantity proportional
to w
c
()

L
c
(t), where

L
c
(t) denotes the average queue length
of a G/G/1 queuing system corresponding to a model of the re-
verse material flow relative to the actual system. In the reverse
flow G/G/1 model, starvation is equivalent to blocking that oc-
curs in the actual system when the inventory level reaches the
hedging point level. Thus, the zero queue event corresponds to
the event where the inventory equals the hedging point in the
actual system. Furthermore, a strictly positive queue event in
the G/G/1 model corresponds to an inventory level strictly be-
low the hedging point in the actual system. The Krmer and
Langenbach-Belz two-moment approximation is used to calcu-
late

L
c
(t) as in [48]. This gives:
L
c
(t) =
(
c
(t))
2
(scv
c
(t) + scv
c1
(t))
2 (1
c
(t))
exp
_
2 (1
c
(t)) (1 scv
c1
(t))
3
c
(t) (scv
c
(t) + scv
c1
(t))
_
+
c
(t)
if scv
c1
(t) 1, otherwise
L
c
(t) =
(
c
(t))
2
(scv
c
(t) + scv
c1
(t))
2 (1
c
(t))
exp
_
(1
c
(t)) (scv
c1
(t) 1)
4scv
c
(t) + 4scv
c1
(t)
_
+
c
(t)
where
c
(t) is the weighted sum of utilization levels of the
workstations in facility c caused by production targets

X
c
(t).
The weight used for each workstation equals the workstations
utilization normalized so that the weights sum to unity. The
squared coefficient of variation of the interarrival times into
I
c
(t
k
) is denoted by scv
c
(t).
Using the large deviations results of [48], the hedging point is
calculated as:
w
c
(t) =
1
c
(t)
log
_
1
c
(t)
c
(t)
L
c
(t)
_
where
c
(t) is the solution of Eq. (11) in [48], assuming that the
demand and service processes in the G/G/1 model are sampled
at the beginning of each time slot from a normal distribution
where the mean and variance values of the effective processing
times are
_
c
(t); scv
c
(t) (
c
(t))
2
_
and (1; scv
c1
(t)), respectively.
In general, autocorrelations may be modeled and a numerical
solution can be easily obtained as described in [48]. This paper uses
the simple case of a single-state Markov-modulated model where
the solution can be written explicitly as:
c
(t) =
2 (1
c
(t))
scv
c
(t) + scv
c1
(t) (
c
(t))
2
.
Note that the variance values reflect the stochastic dynamics
at facilities c and c 1 that are compatible with the planned
production targets and the specified QoS levels. Average utilization
rates correspondto productiontargets inthe optimizedproduction
schedule and are expressed in terms of work assigned to facility c
and the production capacity of facility c.
In case of multiple part types, the minimum average FGI
requirement calculated above is divided among part types
proportional to the workload each imposes on facility c.
The hedging inventory policy described above achieves the
specified QoS levels between facilities and decouples the imple-
mentation of operational decisions across facilities. A single ficti-
tious part type aggregation is employed, and virtual workstations
that behave similarly in each SC facility are utilized. The aggrega-
tion of part types to a single fictitious part is done so as to preserve
the probabilistic behavior of work needed to produce all individual
part types. Therefore, QoS, and hence stockout probability, is de-
fined in terms of the probability that the work incorporated in the
intermediate FGI of facility c falls short of the work incorporated in
the part-type quantities required during the relevant time period
by downstreamfacility c1. In that sense, the hedging point quan-
tity is a scalar that is not part-type specific and is defined in terms
of work corresponding to the quantity of the fictitious part that
is sufficient to achieve the desired stockout probability threshold
for the fictitious part. It should be noted that the proposed supply
chain production planning algorithm does not provide part-type-
specific hedging FGI quantities needed for production floor im-
plementation. Part-type-specific quantities are approximated for
holding-cost accounting purposes by disaggregating the fictitious
part work-hedging quantity to part-type-specific work in propor-
tion to the work rate incorporated in each parts production tar-
get rate relative to the total, and then converting the work parts.
However, it is not proposed that these part-type-specific quanti-
ties be used on the production floor. Among others, the reason is
that the material flow protocol is not fully defined because a dy-
namic policy allocating production capacity among specific part
types in real time would be required under these circumstances.
It is clear that not any such policy, as for example a priority pol-
icy, will achieve the desired QoS for each part type. Such a policy
is not specified, as it is considered to be out of the scope of this
paper. Instead, what is used is the reasonable assumption that an
appropriate policy that achieves part-type-specific QoS guarantees
does exist. Furthermore, it is conjectured that, given the aggregate
fictitious part-hedging quantity that provides the requisite excess
capacity, an adequate real-time policy can be easily devised on the
production floor.
Fig. 8. Horizontal coordination constraint function scaled by
c
(t),
c
(t)
h
c
(), for
a single-part-type system.
The large deviation estimates of the required hedging point
and average queue given above characterize the FGI required to
meet the desired quality of service, assuming that the hedging
point material flow policy is implemented in real time and is
based on the single-part variability of production and demand.
However, in practice, the hedging point material flow policy is
not implemented in real time but over a longer period that
lies between the hourly time scale of production and demand
events and the weekly planning period time scale. This reduces
the coefficient of variation of total production and demand over
the longer material flow implementation period. Hence, the
minimum FGI required to meet the desired quality of service
over the longer material flow period is smaller. To model this
issue, the real-time material flow policy FGI is divided by a
factor of d that must be larger than or equal to one. For the
purposes of the numerical examples, and to avoid the tedious
work of estimating explicitly FGI under a reduced-variability
assumption, d = 6 is used, which is approximately the number
of working days per week. Thus,

h
c
() =
1
d
_
w
c
(t)

L
c
(t)
_
.
Finally, distributing this aggregate quantity among part types using
weights
p
c
(t)

X
p
c
(t)/
P
r=1

r
c
(t)
X
r
c
(t) that sum up to one and
converting from work units back to part units by multiplying with
1/
p
c
(t), the minimum average FGI requirements function for part
type p is obtained:
h
p
c
_
X
c
(t),

X
c1
(t), P
c
(t), P
c1
(t), w
c
(t),
c
(t)
_
=
1
p
c
(t)
p
c
(t)

X
p
c
(t)
P
r=1
r
c
(t)

X
r
c
(t)
1
d
_
w
c
(t)

L
c
(t)
_
.
As mentioned, the coefficients
p
c
(t) convert each unit of FGI,
I
p
c
(t), to units of work defined as the processing time needed
at the upstream facility to produce the required amount of
inventory.
p
c
(t) is used to equal the reciprocal of the minimum
production capacity for part type p of all machines in facility
c,
p
c
(t) =
_
min
mM
c

p
c,m
(t)
_
1
. Fig. 8 shows an instance
of the FGI constraint surface corresponding to upstream and
downstream facilities, each consisting of three identical machines
with production capacity of 50 parts per time period and
c+1
(t) =
c
(t) = 96%.
The hedging inventory requirements of facility c = 1 is cal-
culated by treating scv
0
(t), the squared coefficient of variation of
interarrival times of the demand for final products at the first fa-
cility, as an exogenous parameter. The QoS provided to the cus-
tomers,
1
(t), is also determined by the user. The required hedging
inventory at c = 1,

h
p
1
_
X
1
(t),

D
p
(t) , P
1
(t), P
0
(t) , w
1
(t),
1
(t)
_
,
Fig. 9. FGI hedging constraint function scaled by
p
c
(t),
1
(t)
h
1
(), for a single-
part-type system.
is therefore a function of only one endogenous variable, the pro-
duction target vector of the first facility. An instance of the FGI con-
straint surface corresponding to facility c = 1 comprising three
identical machines with production capacity of 50 parts per time
period,
2
(t) = 96%,
1
(t) = 95%, and scv
0
(t) = 0.6, is plotted
in Fig. 9.
3.4. Planning coordination layer
The planning layer shown in Fig. 3, and its role in the
collaborative framework where the master and subproblemlayers
interact in an iterative algorithm that produces the optimal
production plan, are described next.
3.4.1. Master problem optimization algorithm
Given the longer (weekly) time scale of the planning layer, a
linear programming (LP) based fluid model approximation of the
discrete part production planning problem is used. Moreover, the
fluid model is extended to represent WIP and FGI-driven dynamic
lead times through the nonlinear constraint surfaces defined in
Sections 3.2 and 3.3. These constraints are key components of the
planning layer model (see inequalities (6)(8) in Exhibit 1).
Exhibit 1 (Planning Layer Optimization Problem).
min
X
p
c
(t),
R
p
c
(t)
p,t
_
C
c=1

p
c

Q
p
c
(t) +
C
c=2
p
c
I
p
c
(t) +
p
1
_
I
p
1
(t)
_
+
+
p
1
_
I
p
1
(t)
_
+
p
1
_
J
p
1
(t)
_
_
(1)
subject to:
Q
p
c
(t + 1) = Q
p
c
(t) +

R
p
c
(t)

X
p
c
(t) p, c, t (2)
I
p
c
(t + 1) = I
p
c
(t) +

X
p
c
(t)

R
p
c1
(t) p, c 2, t (3)
I
p
1
(t + 1) = I
p
1
(t) +

X
p
1
(t)

D
p
(t) p, c = 1, t (4)
P
p=1
X
p
c
(t)

p
c,m
(t)

c
(t) m, c, t (5)
Q
p
c
(t) g
p
c
_
X
c
(t), P
c
(t),
c
(t)
_
p, c, t (6)
I
p
c
(t)

h
p
c
_
X
c
(t),

X
c1
(t) , P
c
(t), P
c1
(t) , w
c
(t),
c
(t),
c1
(t)) p, c 2, t (7)
I
p
1
(t) J
p
1
(t)

h
p
1
_
X
1
(t),

D
p
(t) , P
1
(t), P
0
(t) , w
1
(t),
1
(t)) p, c = 1, t. (8)
The planning layers objective is to determine targets of weekly
production and release that minimize weighted WIP, FGI, and
backlog costs over the horizon T, subject not only to the capacity
and material conservation constraints, but also the nonlinear
constraints on weekly average WIP and hedging inventory that
capture the essence of faster time-scale stochastic dynamics driven
by QoS provisioning and other operational policies that may be in
place at some facilities.
The planning layer solves the optimization problem given in
Exhibit 1 where minimization takes place over weekly production
target,

X
p
c
(t), and release,

R
p
c
(t), decision variables. For all facilities,

p
c
and

p
c
denote, respectively, WIP and FGI holding cost rates,
while
p
1
,
p
1
, and
p
1
denote respectively, holding, shortage, and
hedging constraint penalty cost rates
1
at c = 1 applied to
the end-of-period values of FGI. Average WIP and FGI values at
facilities c > 1 are related to the end-of-period values via
the approximations

Q
p
c
(t) = 0.5Q
p
c
(t 1) + 0.5Q
p
c
(t) and
I
p
c
(t) = 0.5Q
p
c
(t 1) + 0.5Q
p
c
(t).
c
(t) is the maximum allowed
utilization level for workstations in facility c during time period
t. The optimization problem of Exhibit 1 is subject to the usual
positivity (with the exception of I
p
1
(t) and J
p
1
(t) being unrestricted
p, t), initial condition, capacity (5), material conservation (2)(4),
and nonlinear constraints (6)(8) that capture lead time and QoS
dynamics. P
c
(t) and
c
(t) are as defined in Section 3.2. Vector
[w
c
(t) c = 1, 2, . . . , C] is determined at the QoS layer. It denotes
the hedging inventory level implemented by facility c to control
its production schedule so as to prevent the downstream stockout
frequency from violating the desired probabilistic QoS constraint.
Notice that the hedging inventory level is strictly enforced on
I
p
c
(t) in constraint (7), whereas constraint (8) is a soft constraint.
Violation of the hedging inventory requirement on I
p
1
(t) is defined
as J
p
1
(t) = I
p
1
(t)
h
p
1
(), and its negative part,
_
J
p
1
(t)
_
, is penalized
in the objective function at a rate of
p
1
. This allows partitioning
the FGI cost into three parts: (i) I
p
1
(t)

h
p
1
(), with a cost rate of
p
1
, (ii) 0 I
p
1
(t)

h
p
1
(), with a cost rate of
p
1
for each unit of
negative deviation from

h
p
1
(), and (iii) I
p
1
(t) 0, with a cost rate
of
p
1
+
p
1
(because both
_
I
p
1
(t)
_
and
_
J
p
1
(t)
_
are positive). This

constraint is defined on the FGI end-of-period values, I
p
1
(t), rather
than the average time period values,

I
p
1
(t). The justification of this
modeling choice is that it constitutes a reasonable simplification
(indeed, a conservative cost estimate) given the nonlinear nature
of the inventory/backlog cost trajectory).
3.4.2. Iterative algorithm for layer collaboration and coordination
An iterative single master problem (centralized planning coor-
dination layer) multiple subproblem (decentralized performance
and information layer) algorithm has been developed to model
the nonlinear constraints and derive optimal production plans that
explicitly account for variable lead time and QoS constraints (see
Fig. 3). The efficient representation of constraints (6) and (7) in
Exhibit 1 requires point estimates and sensitivity estimates so as
to approximate the nonlinear constraint boundaries. In fact, the it-
erative fine-tuning of a finite number of appropriately selected lo-
cal approximations leads to convergence under mild convexity or
quasi-convexity conditions [46]. This iterative algorithm is sum-
marized below and also shown in Fig. 3:
1. The planning layer, at iteration n, calculates tentative
production and material release schedules
n
X
p
c
(t),
n
R
p
c
(t)p, c, t,
1
Note that penalizing for back orders at the planning layer does not contradict
the probabilistic QoS guarantees provided by the hedging point policy implemented
through production policy thresholds w
c
(t). The planning layers constraints (7)
and (8) assure that weekly production targets are consistent with probabilistic QoS
requirements.
and conveys them to facility subproblems along with the WIP and
FGI levels
n

Q
p
c
(t),
n
I
p
c
(t), that result from the production schedule
through the material flow constraints (2)(4) in Exhibit 1.
2. For the tentative production target assigned to it by the
planning layer and the QoS it receives from the upstream portion
of the SC, each facility conveys to the quality of service layer a
Markov-modulated model that represents the stochastic behavior
of its production capacity. For the same tentative production
targets, the quality of service layer proceeds to calculate tentative
hedging points w
c
(t) that each facility needs to employ to provide
the required QoS to its downstream facility.
3. Each facility evaluates its performance and calculates its
average WIP as well as the WIPs sensitivity with respect to the
facilitys tentative production target. For each pair of upstreamand
downstream facilities, the pairs performance is also evaluated to
calculate the average FGI needed to achieve the QoS requirements
as well as the FGIs sensitivity with respect to the tentative
production targets of both facilities. The planning layer receives
the following feedback: (i) the requiredWIP andFGI
_
g
p
c
() ,

h
p
c
()
_
,
and (ii) their sensitivities, namely g
p
c
() and
h
p
c
() w.r.t. the
production targets.
4. The planning layers linear constraint set is augmented
using the hyperplanes tangent to functions g
p
c
() and

h
p
c
() at
the most recent iterations production target values
n
X
c
(t). More
specifically, the additional constraints are:
Q
p
c
(t) g
r
c
_
n
X
c
(t)
_
+ g
r
c
_
n
X
c
(t)
_
X
c
(t)
n
X
c
(t)
_
p, c, t (9)
I
p
c
(t)

h
r
c
_
n
X
c
(t),
n
X
c1
(t)
_
+
h
r
c
_
n
X
c
(t),
n
X
c1
(t)
_
_

X
c
(t)
n
X
c
(t)
X
c1
(t)
n
X
c1
(t)
_
p, c 2, t (10)
I
r
1
(t) J
r
1
(t)

h
r
1
_
n
X
1
(t)
_
+
h
r
1
_
n
X
1
(t)
_
X
1
(t)
n
X
1
(t)
_
p, c = 1, t. (11)
Note that this set of hyperplanes (linear constraints) (9)(11),
are generated at iteration n to approximate the nonlinear WIP
constraints (6) and FGI constraints (7) and (8), respectively, around
the point
n
X
c
(t). Note also that the probability distributions P
c
(t),
operational
c
(t), and the hedging points w
c
(t) are omitted for
notational simplicity.
5. Using the tangent hyperplanes (9)(11) accumulated over
past iterations 1 through n and the master problem constraints
(2)(5), the planning layer solves again the master problem as a
linear program to produce a new set of tentative targets,
n+1
X
p
c
(t),
n+1
R
p
c
(t)p, c, t. Iterations continue to convergence defined by
a tolerance with respect to which the nonlinear constraints are
satisfied.
Fig. 10 depicts this iterative process for iterations n and n +
1 on a single-part-type WIP constraint. Tentative planning layer
solution
__
n

Q
p
c
(t),
n
X
p
c
(t)
__
is associated with the nominal point
_
g
p
c
_
n
X
c
(t)
_
,
n
X
p
c
(t)
_
on the surface of the nonlinear constraint
where the tangent hyperplane constraint is generated. The next it-
erations tentative solution,
_
n+1
Q
p
c
(t),
n+1
X
p
c
(t)
_
, satisfies all con-
straints added so far. A new tangent hyperplane constraint is gen-
erated after iteration n +1 at the point
_
g
p
c
_
n+1
X
c
(t)
_
,
n+1
X
p
c
(t)
_
.
Note that the approximation accuracy of the nonlinear con-
straint surfaces increases monotonically with the addition of
tangent hyperplanes. As the algorithm approaches convergence,
tentative solutions are close to each other, resulting in the addition
of an increasing number of tangent hyperplane constraints in the
vicinity of the optimal solution, rendering the linear approximation
error of the nonlinear constraints arbitrarily small. This process of
outer linearization of the nonlinear constraints is a smart lineariza-
tion because it is not a uniform linearization that is, in general,
Fig. 10. Tentative target
_
n

Q
p
c
(t) ,
n
X
p
c
(t)
_
and the associated tangent hyperplane
depicted on a single-part-type WIP constraint.
Fig. 11. Feasible production as a function of WIP with various deterministic fluid
and stochastic model approximations.
computationally intractable, but is progressively dense and hence
accurate where it matters, that is, in the vicinity of the optimal so-
lution.
An example to introduce the basic idea of lead time dynamics
at the coarse time scale of the planning layer and its significance
in yielding a superior production plan is employed next. Consider
a simple manufacturing facility consisting of a 10-workstation
transfer line. Each workstation has stochastic processing time and
average production capacity of 10 parts per week. This is also the
capacity of the facility.
Fig. 11 shows the average weekly facility production rate,

X,
that is achievable as a function of the average available WIP,
denoted by

Q, under three successively more accurate fluid
model approximations (FMAs). The naive FMA does not impose
any average WIP constraints, and the production rate is simply
constrained by the average production capacity. The certainty
equivalent FMA considers deterministic workstation capacities
equal to the average capacity availability. As a result, it disregards
queuing delays. It underestimates the average WIP needed for
sustaining a certain average production rate, setting it equal to
the sum of the 10 machine utilizations. This is a linear constraint
involving production and WIP. Finally, the optimal open loop
feedback controller (OOLFC) employs a stochastic model that
captures the fact that effective workstation processing times are
random variables. This results in the correct estimate of average
WIP, and hence of average lead time, which reflects the average
queue levels in front of workstations in addition to the average
number of parts being processed.
Fig. 13. OOLFC, CE, and LIX.
In the context of the SC planning coordination layer, the
OOLFC model approximation is implemented via the inclusion
of nonlinear constraints on production

X
p
c
(t) represented by
the WIP and FGI functions, g
p
c
() and

h
p
c
(), introduced in
Sections 3.2 and 3.3. Several approaches have been successfully
implemented to handle these constraints, so far based on iterative
tangent or piecewise-linear approximations [34] that allow LP
optimization. When used in a manner of an open-loop control,
errors introduced via shortcomings of modeling choices in
estimating these functions, g
p
c
() and

h
p
c
(), will propagate over the
planning horizon. The magnitude of errors depend on: (i) accuracy
of the model used in estimation, and (ii) the time-scale separation
required for steady-state convergence [35]. The propagation and
impact can be evaluated through simulation. Use of the rolling
horizon approach requires implementation of only the first-period
results, thus mitigating the impact of error propagation. Although
the preliminary evaluations show that the impact is relatively
small, additional work in this area would be useful [33].
The nonlinear constraints render the production plans gener-
ated not only feasible but also superior to plans obtained by al-
ternative deterministic approaches based on certainty equivalent
formulations [20,51,46]. The introduction of the nonlinear con-
straints is equivalent to formulating and solving the open-loop
optimal controller known to dominate the certainty equivalent
controller [52]. The certainty equivalent controller, possibly ad-
justed to model a larger, worst-case, constant lead time delay in
each activity, represents todays industry practice, and, as such,
it is compared to the proposed open-loop optimal controller in
Section 4.
4. Algorithmic implementation and numerical evaluation of
benefits
This section discusses the computational experience of the
proposed dynamic lead time and QoS hedging inventory modeling
algorithm, referred to as the optimal open loop feedback controller
(OOLFC). OOLFC performance is compared to two algorithms
that represent industry practice to explore the value that one
may attribute to the dynamic lead time and QoS hedging
inventory information employed by the OOLFC approach. Whereas
production planning algorithms have been previously reported
that model explicitly dynamic lead times [34,53,51,46,54,55],
results reflecting probabilistic QoS policies are presented here for
the first time.
Fig. 12. An example supply chain with three facilities producing five part types.
Table 2
Average workstation production capacities
Facility 3 Facility 2 Facility 1
m
3,1
m
3,2
m
3,3
m
3,4
m
3,5
m
2,1
m
2,2
m
2,3
m
2,4
m
1,1
m
1,2
m
1,3
m
1,4
m
1,5
m
1,6
p.t. 1 70 70 55 55 55 55 55 55 55 40 35 40 40 35 35
p.t. 2 75 65 60 70 55 55 55 65 55 40 50 40 35 50 50
p.t. 3 75 65 75 75 80 55 60 55 65 60 50 55 50 60 55
p.t. 4 85 70 90 80 65 65 55 70 70 55 65 65 55 55 65
p.t. 5 70 85 80 75 90 70 70 65 65 65 65 70 70 65 70
Table 3
WIP, FGI holding, and FGI backlog cost coefficients
Facility 3 Facility 2 Facility 1
WIP
p
3
FGI
+

p
3
WIP
p
2
FGI
+

p
2
WIP
p
1
FGI
+

p
1
FGI

p
1
FGH

p
1
p.t. 1 20 27 30 55 60 150 1200 1800
p.t. 2 15 20 25 40 50 110 880 1320
p.t. 3 10 12 15 40 50 120 960 1440
p.t. 4 10 23 30 42 45 100 800 1200
p.t. 5 5 8 10 26 30 70 560 840
Fig. 14. Facility 1 optimal solution details of OOLFC under demand scenario 4.
Presented below is numerical experience from six demand
scenarios applied to a five-part type 3 facility SC depicted
in Fig. 12. Facilities 3, 2, and 1 having five, four, and six
workstations, respectively, with average production capacities,
are shown in Table 2, with part types corresponding to columns
and workstations to rows. Facilities are allowed a maximum
workstation utilization of
c
(t) = 0.9 c, t. QoS levels are set
at
(t) =
_
95% 97% 97% 97%
_
for facility indices 1 through
4.
4
(t) is the QoS provided to facility 3 by the raw material
vendor. The demand-squared coefficient of variation is set to 0.6
throughout the planning horizon. Initial levels of WIP and FGI are
set to zero in all facilities. Cost coefficients are shown in Table 3.
All demand scenarios consist of a 23-week planning horizon.
The first four weeks in scenarios 1 and 2, the first two weeks in
scenarios 3, 4, 5, and 6, and the last four weeks in all scenarios
were on purpose assigned zero demand to allow the algorithm
to fill and empty the system optimally without disadvantaging
the suboptimal industry practice alternative planning algorithms
Table 4
Weekly demand for scenarios 3, 4, 5, and 6; weeks 3 through 19
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Demand scenario 3 p.t. 1 3 9 4 1 1 1 1 2 2 2 2 7 11 10 8 10 10
p.t. 2 21 15 10 8 5 6 4 8 14 12 29 23 17 13 21 11 19
p.t. 3 4 3 1 1 1 3 3 3 1 1 2 4 5 3 1 1 3
p.t. 4 11 3 1 4 9 14 15 15 6 2 4 7 16 14 7 2 7
p.t. 5 5 9 5 4 4 4 4 6 6 6 5 4 5 7 6 6 11
p.t. 2 0 0 18 11 13 6 8 4 19 15 42 19 16 28 6 49 31
p.t. 3 0 0 4 2 1 1 4 5 1 1 3 6 4 1 1 4 7
p.t. 4 0 0 10 1 3 11 18 28 9 3 6 11 17 7 0 16 7
p.t. 5 0 0 4 9 4 4 5 7 7 7 6 5 9 7 9 23 10
p.t. 2 0 23 11 11 8 6 5 7 16 13 34 22 16 19 17 14 39
p.t. 3 0 4 2 1 1 2 3 4 1 1 2 5 5 3 1 1 4
p.t. 4 0 12 1 2 7 13 18 21 7 3 5 9 16 11 3 3 13
p.t. 5 0 5 8 5 4 4 4 6 6 6 5 5 8 8 8 10 18
p.t. 2 0 0 9 5 3 15 28 17 7 0 5 14 10 5 0 2 1
p.t. 3 0 0 8 7 6 10 4 2 11 6 5 6 4 8 6 4 1
p.t. 4 0 0 7 3 18 18 4 1 4 1 5 2 2 7 8 11 5
p.t. 5 0 0 4 1 33 13 1 4 3 18 7 7 1 4 3 4 9
considered. The first two scenarios have a constant demand for
each of the five part types during weeks 519 equal to vectors of
(7, 14, 8, 6, 10) and (11, 9, 8, 6, 3). Demand scenarios 3, 4, 5, and 6
during weeks 319 are given in Table 4. Part types correspond to
rows.
Scenarios 3, 4, 5, and 6 represent increasingly more costly
situations for the SC where, for a range of intermediate time
periods in the planning horizon, namely periods (1316), (1217),
(1419), and (719), respectively, the cumulative demand load
exceeds the cumulative production capacity.
To compare the proposed SC planning algorithmto the state-of-
the-art practice in industry, three planning approaches are evalu-
ated, the one proposed in this paper and two additional ones that
represent advanced versions of planning algorithms employing the
standard feature of the fixed lead time approximation. Given that
the fixed lead time approximation is a basic foundation of industry
Table 5
OOLFC, LIX, and CE relative cost comparison
Approach OOLFC LIX CE OOLFC LIX CE OOLFC LIX CE
Scenario 1 Scenario 2 Scenario 3
T.C. 124K 130 2698 77K 158 4102 120K 204 2397
WIP 18 26 15 25 45 24 18 28 15
FGI 14 15 12 17 18 17 15 15 13
FGI
+
62 83 9 57 92 0 62 83 8
FGI
0 0 1614 0 0 2472 0 3 1457

FGH
6 7 1048 1 3 1588 5 75 903

Scenario 4 Scenario 5 Scenario 6
T.C. 242K 136 1604 169K 126 1780 1141K 152 518
WIP 10 15 8 14 19 12 2 4 2
FGI 10 9 8 12 11 11 2 2 1
FGI
+
74 66 21 69 84 18 4 3 1
FGI
0 10 989 0 0 1072 33 54 326

FGH
6 37 579 4 11 667 59 89 188

practice algorithms, these two algorithms are treated as proxies
of industry practice. To improve the performance of the indus-
try proxy algorithms so as to provide a fair comparison to the
proposed algorithm, after the material release schedule to the up-
stream most facility is determined, the fixed lead time approxi-
mation is relaxed in modeling the actual production performance
of the supply chain. Thus, it is asserted that these two industry
practice proxies are actually representative of the state of the art
in industry practice. By contrast, the vanilla industry practice
in widespread MRP scheduling uses a fixed lead time to model
delays at all production facilities, giving rise to further work-in-
process inventory mismatches along the supply chain. The three
methodologies used for comparison purposes in the numerical ex-
perience section are described next in greater detail:
(i) Optimal open loop feedback controller approach. Recall
that the optimal open loop feedback controller approach (OOLFC)
is the short name used for the algorithm proposed in this
article and described in Section 3.4. It models the dynamics of
required lead times and QoS hedging inventory by employing
a smart, progressively increasing in accuracy outer linearization
approximation of the nonlinear lead time and hedging inventory
constraints. For a one-dimensional case, this can be visualized
Fig. 17. Facility 3 production targets and FGI levels comparison among OOLFC, LIX, and CE approaches under demand scenario 1.
as using a tentative polyhedral feasible decision variable region
area corresponding to an approximation of the union of areas A
2
and A
3
in Fig. 13. In Fig. 13, the top horizontal line represents
production capacity and the one below represents the maximum
allowed utilization, namely the linear constraint (5) in Exhibit 1.
Littles law is then employed to convert the WIP, denoted by

Q,
versus production, denoted by

X, space to the equivalent lead time
versus production space. The OOLFC approach is compared to the
following two approaches.
(ii) Certainty equivalent approach. The certainty equivalent
(CE) approach provides a conservative upper cost bound to the
state of the art in industry practice. It generates a raw material
release schedule based on an optimistic, that is, relaxed, feasible
region that corresponds to the ideal situation in which lead times
equal the fixedsumof requiredprocessing time withnoaccounting
for queuing delays. The resulting raw material release schedule
generates shortfalls and almost surely provides a higher planning
cost (especially material shortage costs) than that of the second
proxy presented below. The relaxed feasible decision space used
to generate the raw material release schedule can be visualized
as the union of areas A
1
, A
2
, and A
3
in Fig. 13. Under the CE
approach, the production planning problem is first solved under
this optimistic fixed lead time constraint, ignoring QoS constraints.
The resulting production schedule violates the actual lead time and
QoS constraints. Under the CE approach, the raw material release
schedule associated with the infeasible production plan is then
fixed, and facility-specific production is reoptimized subject to the
fixed rawmaterial release schedule but with a full implementation
of the nonlinear WIP and FGI constraints. The feasible production
plan thus obtained is reported as the production planning cost of
the CE approach. A second, more favorable proxy of the state of the
art in industry practice that is also based on the fixed lead time
approximation is considered below.
(iii) Limited information exchange approach. The limited in-
formation exchange (LIX) approach is finally used to generate pro-
duction schedules that are the best representatives of the state of
the art inindustry practice. It models limitedinformationexchange
among planners and operators by generating a production and raw
material release schedule based on constant lead times that are
representative of expected facility loading conditions. More specif-
ically, the level of the constant lead times is set to the worst-case
level, namely the one required to sustain the maximum allowable
utilization. The resulting feasible decision space can be visualized
as area A
3
in Fig. 13. The LIX approach is a reasonable abstraction of
the limited information about actual lead times that production fa-
cility managers may provide on a yearly or quarterly basis to legacy
MRP-based planning systems in place. Again, to render the LIX ap-
proach a state of the art of industry practice representative, further
action is taken to ensure compatibility of WIP requirements and
actual production levels and to enforce interfacility QoS require-
ments. To this end, the rawmaterial release rates are again fixed to
the schedule determined by the linear fixed lead time constraints,
and the production schedule is reoptimized subject to the union of
the linear fixed lead time and the nonlinear lead time and QoS con-
straints. Note that the LIX approach provides feasible and virtually
optimal production schedules when the production facilities in the
SC are level loaded, whereas it is still disadvantaged when demand
and hence loading varies over time. It may therefore be concluded
that the performance of state of the art in industry practice today is
well represented by the LIX and CE approaches. Cost comparisons
reported below may therefore be interpreted accordingly.
Figs. 1416 present the optimal production schedules for
facilities 1, 2, and3 of the proposedOOLFCapproachunder demand
scenario 4. Part-type-specific schedules are shown from top to
bottom. The release rate of parts into each facility, average WIP in
each facility, facility-specific production rates, and FGI (facility 1)
or hedging inventory (facilities 2 and 3) schedules are also shown.
Demand trajectories are superimposed for easy reference on FGI
trajectories as a bar graph on the same scale. It is observed that for
part types 1, 2, and5 productionplans at facility 1 are characterized
by significant preproduction during the first half of the planning
horizon for part type 2 and in the second half for the other two. FGI
trajectories of part types 3 and 4 seemto balance the requirements
of the FGI hedging constraint and the implied cost of sharp changes
in FGI level.
InFigs. 15 and 16, the last column depicts the hedging inventory
located downstream to the associated facility. Past computational
experience indicates that the trajectory of the hedging inventory
resembles closely the upstream production schedule, which is
the major determinant of the prescribed hedging level. For the
reasonable cost coefficients and demand loads employed, only
under extreme cases does the optimization algorithm choose to
stock material in FGI for future use.
In Figs. 1719, the production targets of facility 1 obtained by
each of the three approaches are compared. Solid, dashed, and
dotted lines are used to distinguish OOLFC, LIX, and CE approach
weekly production schedules for three different demand scenarios.
Part type 1 through 5 production is shown from top to bottom,
respectively.
Demand scenario 1 schedule shown in Fig. 17 corresponds to a
constant workload of about 79% on facility 1 and a constant part
type demand mix. Because this load is close to the 90% for which
the constant lead time and hedging inventory constraint used in
the LIXapproachhas beencalculated, one wouldexpect a relatively
small cost difference in the performance of the LIX approach
relative to the OOLFC approach. The differences in performance
between LIX and OOLFC observed here (30% higher cost than
OOLFC as reported in Table 5) are primarily caused by (i) higher
FGI holding costs at the first facility due to LIX approachs need
to build up inventory earlier than necessary and then not being
able to reduce the inventory until the end of the planning horizon
because demand load remains high till the end of the horizon,
and (ii) higher WIP holding costs due to the lower production
target levels that are needed during time periods at the beginning
and the end of the planning horizon when the WIP and hedging
inventory levels are ramped up from zero levels or ramped
down to zero levels. In a production system processing a single-
part-type, or equivalently a family of perfectly similar part types,
level loading that is close to the maximum allowed utilization,
c
(t), yields minimal performance differences [56]. On the other
hand, scenario 2 has a constant workload of about 66%, which
differs significantly fromthe 90% utilization to which LIXfixed lead
times are calibrated. As a result, LIX performance is worse58%
higher as opposed to 30%. The level-loading, constant-production
mix observed in Fig. 17 verifies the proximity of the LIX and
OOLFC-generated production plans during the middle portion of
the planning horizon. The remaining differences are due to the
nonlinear lead time and hedging inventory requirements during
the transition period at the beginning/end of the planning horizon
from/to zero demand activity.
Fig. 18 shows the production target and FGI differences among
the three approaches under scenario 4. Although the overall
horizon workload is feasible, demand scenario 4 contains periods
of very high demand. Production schedules for each of the three
approaches attempt to balance preproduction and shortages in
different ways. There are multiple ramp-up and ramp-down
portions of the production schedule during which the LIXapproach
results followthe OOLFCapproachresults withdelay. This behavior
of production schedules is even more conspicuous in Fig. 19,
which presents the results of each of the three approaches on
scenario 6. Due to the fluctuating demand pattern of scenario 6,
none of the approaches can avoid temporary shortages of FGI at
facility 1. Furthermore, the delay of the LIX approach solution in
responding to scenario 6 demand has an amplification effect on
backlog and results in very high backlog costs that are in absolute
terms much higher than WIP costs. This is the main reason why the
LIX approach reports 42% higher total cost than OOLFC in scenario
6, in contrast to 26% in scenario 4.
As mentioned already, while OOLFC corresponds to the pro-
posed modeling of dynamic lead times through the incorporation
of nonlinear constraints (6)(8) of Exhibit 1, the LIX approach is
used as a representative of the current practice of planning with a
fixed lead time and safety stock assumptions, usually correspond-
ing to level-loading conditions at the maximumallowed efficiency
(worst-case analysis). Finally, the CE approach is representative of
an over-optimistic, that is, too low, lead time allowance that as-
sumes no queuing delays. Obviously, CE schedules are hampered
by shortfalls and backlogs, rendering the CE approach one order of
magnitude worse than the more advanced state of the art in indus-
try practice represented by LIX.
Table 5 summarizes the potential benefits of the proposed
OOLFCframework. Note that the Total Cost (T.C.) columnshows the
value of the optimal objective function for OOLFC and the relative
percentage of the cost produced by the CE and LIX approaches. The
remaining columns represent a decompositionof the total cost into
its WIP, positive FGI, and backlog components. The superiority of
the OOLFC dynamic lead time approach is clear.
In terms of computational effort, all three approaches under all
six scenarios converged in no more than 16 master problem sub-
problem iterations. In 15 runs out of the total of 18, convergence
was achieved in at most 12 iterations. The software implementa-
tion involved compiled C++ objects, Matlabs interpreted scripts,
and Ilogs CPLEX optimizer. On a Pentium 4 PC running a Linux
operating system, the calculations for each iteration takes about
five seconds (there is ample room for improvement if the whole
implementation is done in compiled objects), and the LP solution
takes less than half a second. Extensive empirical experience with
varying sizes of problem instances, up to eight part types, seven
facilities, and 40 time periods indicates that: (i) the number of it-
erations to convergence remains flat over all cases run, and (ii) the
total computational effort grows linearly in the number of time pe-
riods and number of facilities but polynomially in the number of
part types.
5. Conclusion
Computationally tractable and robust algorithms for SC coor-
dination and planning are presented that are capable of incorpo-
rating nonlinear lead time performance and probabilistic quality
of service requirements to reduce SC inventory and increase its
speed. Nevertheless, demonstration that significant reduction in
inventory costs is possible when the nonlinear relationship of fa-
cility lead times is modeled in the SC production planning process
is not the major contribution of this paper. The major contribution
of this paper is the proposal and implementation of a practical, ef-
ficient, tractable, and robust algorithm that is capable of achieving
these cost savings. In doing so, the concept is proved that plan-
ning on constant lead times is not a necessary evil imposed by the
presumption of insurmountable computational complexity. In fact,
this presumption is shown to be incorrect and that industry does
not have to live any longer with the undesirable consequences of
the constant lead time assumption impeding todays production
planning practice.
Furthermore, the results leverage the value of research in
stochastic systems and, in particular, queuing network models of
production systems, including analytical probabilistic approaches
as well as Monte Carlo simulation approaches that have attracted
and continue to attract a substantial portion of the research
community. Ongoing research is focusing on improving QoS layer
algorithms and enhancing fluid Monte Carlo simulation models
applicable to multi-echelon production systems.
Finally, it is emphasized that the inclusion of production
system stochastic dynamics in SC coordination and production
planning is particularly desirable in view of the emergence of
sensor network and related technologies such as RFIDs. The value
of these technologies ability to supply reliable, affordable, and
timely information about production conditions, transportation,
receiving, warehousing, and retail activities will be undoubtedly
enhanced if this information is translated to significant efficiency
gains in SC management.
Acknowledgment
NSF Grant DMI-0300359 is acknowledged for partial support of
the research reported here.
References
[1] Meal HC, Wachter MH, Whybark DC. Material requirements planning in
hierarchical production planning systems. International Journal of Production
Research 1987;25(7):94756.
[2] Hax AC, Meal HC. Hierarchical integration of production planning and
scheduling. In: Geisler MA, editor. Logistics, TIMS studies in the management
sciences. North Holland; 1975. p. 5369.
[3] Bitran GR, Tirupati D. Hierarchical production planning. In: Graves SC, Rinnooy
Kan AHG, Zipkin PH, editors. Logistics of production and inventory, vol. 4.
Amsterdam: N. Holland; 1993 [Chapter 10].
[4] Chen H, MandelbaumA. Discrete flownetworks: Bottleneck analysis and fluid
approximations. Operations Research 1991;16(2):40847.
[5] Lambrecht M, Decaluwe L. JIT and constraint theory: The issue of bottleneck
management. Production and Inventory Management Journal 1988;29(3):
616.
[6] Goldratt E, Cox J. The goal: Excellence in manufacturing. North River Press;
1984. 262 pp.
[7] Sharifnia A, Caramanis M, Gershwin SB. Dynamic set-up scheduling and flow
control in manufacturing systems. Discrete Event Dynamic Systems: Theory
and Applications 1991;1(2):14975.
[8] Caramanis M, Liberopoulos G. Perturbation analysis for the design of flexible
manufacturing system flow controllers. Operations Research 1992;40(6):
110725.
[9] Liberopoulos G, Caramanis M. Dynamics and design of a class of parameterized
manufacturing flowcontrollers. IEEE Transactions on Automatic Control 1995;
40(6):101828.
[10] Kaskavelis C, Caramanis M. Efficient Lagrangian relaxation algorithms for real-
life-size job-shop scheduling problems. IIE Transactions on Scheduling and
Logistics 1998;30(11):108597.
[11] Khmelnitsky E, Caramanis M. One-machine n-part-type optimal set-up
scheduling: Analytical characterization of switching surfaces. IEEE Transac-
tions on Automatic Control 1998;43(11):15848.
[12] Deleersnyder J, Hodgson T, King R, OGrady P, Savva A. Integrating Kanban type
pull systems and MRP type push systems: Insights from a Markovian model.
IIE Transactions 1992;24(3):4356.
[13] Jain S, Johnson M, Safai F. Implementing setup optimization on the shop floor.
Operations Research 1996;43(6):84351.
[14] Goncalves JF, LeachmanRC, GasconA, Xiong Z. Aheuristic scheduling policy for
multi-item, multi-machine production systems with time-varying, stochastic
demands. Management Science 1994;40(11):145568.
[15] Brandimarte P, Alfieri A, Levi R. LP-based heuristics for the capacitated lot
sizing problem. CIRP Annals - Manufacturing Technology 1998;47(1):4236.
[16] Veatch M, Caramanis M. Optimal average cost manufacturing flowcontrollers:
Convexity and differentiability. IEEE Transactions on Automatic Control 1999;
44(4):77983.
[17] Chen H, Yao DD. Dynamic scheduling of a multiclass fluid network. Operations
Research 1993;41(6):110415.
[18] Bertsimas D, Paschalidis ICh, Tsitsiklis JN. Optimization of multiclass
queuing networks: Polyhedral and nonlinear characterizations of achievable
performance. The Annals of Applied Probability 1994;4(1):4375.
[19] Chen H, Yao DD. A fluid model for systems with random disruptions.
Operations Research 1992;40(2):23947.
[20] Connors D, Feigin G, Yao DD. Scheduling semiconductor lines using a fluid
network model. IEEE Transactions on Robotics and Automation 1994;10(2):
8898.
[21] Feng Y, Leachman R. A production planning methodology for semiconductor
manufacturing based on iterative simulation and linear programming
calculations. IEEE Transactions on Semiconductor Manufacturing 1996;9(3):
25769.
[22] Orcun S, Uzsoy R, Kempf K. Using system dynamics simulations to compare
capacity models for production planning. In: Proceedings of the 38th winter
simulation conference. 2006. p. 185562.
[23] Sharifnia A. Stability and performance of distributed production control
methods based on continuous-flow models. IEEE Transactions on Automatic
Control 1994;39(4):72537.
[24] Kumar PR, Meyn SP. Stability of queuing networks and scheduling policies.
IEEE Transactions on Automatic Control 1995;40(2):25160.
[25] Kumar PR, Seidman TI. Dynamic instabilities and stabilization methods in
distributed realtime scheduling of manufacturing systems. IEEE Transactions
on Automatic Control 1990;35(3):28998.
[26] Dai JG. On the positive Harris recurrence for multiclass queuing networks: A
unified approach via fluid models. The Annals of Applied Probability 1995;5:
4977.
[27] Graves SC. Atactical planning model for a job shop. Operations Research 1986;
34(4):52233.
[28] Lambrecht MR, Ivens PL, Vandaele NJ. ACLIPS: A capacity and lead time
integrated procedure for scheduling. Management Science 1998;44(11):
154861.
[29] Woodruff DL, Voss S. A model for multistage production planning with
load dependent lead times. In: Proceedings of the 37th Hawaii international
conference on system sciences. 2004. p. 19.
[30] Asmundsson J, Rardin RL, Uzsoy R. Tractable nonlinear production planning
models for semiconductor wafer fabrication facilities. IEEE Transactions on
Semiconductor Manufacturing 2006;19(1):95111.
[31] Ettl M, Feigin GE, Lin GY, Yao DD. A supply network model with base-stock
control and service requirements. Operations Research 2000;48(2):21632.
[32] Graves SC. Safety stocks in manufacturing systems. Journal of Manufacturing
and Operations Management 1988;1(1):67101.
[33] Pahl J, Voss S, Woodruff D. Production planning with load dependent lead
times: An update of research. Annals of Operations Research 2007;153(1):
297345.
[34] Caramanis M, Pan H, Anli OM. A closed-loop approach to efficient and stable
supply-chain coordination in complex stochastic manufacturing systems. In:
Proceedings of the American control conference. 2001. p. 13818.
[35] Saksena VR, OReilly J, Kokotovic PV. Singular perturbations and time-scale
methods in control theory: Survey 19761983. Automatica 1984;20(3):
27393.
[36] Clark AJ, Scarf H. Optimal policies for a multi-echelon inventory problem.
Management Science 1960;6:47590.
[37] Lambrecht MR, Muckstadt JA, Luyten R. Protective stocks in multi-stage
production systems. International Journal of Production Research 1984;22(6):
100125.
[38] Glasserman P, Tayur SR. Sensitivity analysis for base-stock levels in
multiechelon production inventory systems. Management Science 1995;
45(2):26381.
[39] Paschalidis ICh, Vassilaras S. On the estimation of buffer overflowprobabilities
from measurements. IEEE Transactions on Information Theory 2001;47(1):
17891.
[40] Suri R, Hildenbrant RR. Modeling flexible manufacturing systems using mean
value analysis. Journal of Manufacturing Systems 1984;3(1):2738.
[41] Schweitzer P. Approximate analysis of multiclass closed network of queues.
In: International conference on stochastic control and optimization. 1979.
[42] Bard Y. Some extensions to multiclass queuing network analysis. In: Arato AM,
Butrimenko, Gelenbe E, editors. Performance of computer systems. North
Holland; 1979.
[43] Caramanis M. Production system design: A discrete event dynamic system
and generalized Benders decomposition approach. International Journal of
Production Research 1987;8(25):122334.
[44] Kouikoglou VS, Phillis YA. An exact discrete-event model and control policies
for production lines with buffers. IEEE Transactions on Automatic Control
1991;36(5):51527.
[45] Caramanis M, Wang J, Paschalidis ICh. Enhanced fluid approximation models
discrete part dynamics while enabling Monte Carlo simulation-based I.P.A.
Boston University Center for Information and Systems Engineering working
paper; March 2003.
[46] Caramanis M, Anli OM. Modeling work in process versus production con-
straints for efficient supply chain planning: Convexity issues. In: Proceedings
of the IEEE CDC. 1999. p. 9006.
[47] Anli OM. Iterative approximation selection algorithms for supply chain
production planning. Boston University Center for Information and Systems
Engineering working paper; 2002.
[48] Bertsimas D, Paschalidis ICh. Probabilistic service level guarantees in make-to-
stock manufacturing systems. Operations Research 2001;49(1):11933.
[49] Paschalidis ICh, Liu Y. Large deviations-based asymptotics for inventory
control in supply chains. Operations Research 2003;51(3):43760.
[50] Paschalidis ICh, Liu Y, Cassandras CG, Panayiotou C. Inventory control for
supply chains with service level constraints: A synergy between large
deviations and perturbation analysis. The Annals of Operations Research
(Special Volume onStochastic Models of Production-Inventory Systems) 2004;
126:23158.
[51] Caramanis M, Anli OM. Dynamic lead time modeling for JIT production
planning. In: Proceedings of the IEEE robotics and automation conference, vol.
2. 1999. p. 14505.
[52] Bertsekas DP. Dynamic programming and optimal control, vol. I and II.
Belmont (MA): Athena Scientific; 1995.
[53] Caramanis M, Pan H, Anli OM. Is there a trade off between lean and agile
manufacturing? A supply chain investigation. In: Proceedings of the third
aegean international conference on design of manufacturing systems. 2001.
[54] Caramanis M, Paschalidis ICh, Anli OM. A framework for decentralized control
of manufacturing enterprises. In: Proceedings of the DARPA-JFACCsymposium
on advances in enterprise control. 1999. p. 99109.
[55] Caramanis M, Anli OM. Manufacturing supply chain coordination through
synergistic decentralized decision making. In: Proceedings of Rensselaers
international CAICIM. 1998.
[56] Anli OM. Supply chain production planning: Modeling dynamic lead times and
efficient intercell inventory policies. Boston University Center for Information
and Systems Engineering Technical Report and Ph.D. dissertation; September
2003.
Osman Murat Anli holds a B.Sc. degree (1994) in industrial engineering from
Istanbul Technical University, a M.Sc. degree (1997) in operations research and
statistics from Rensselaer Polytechnic Institute, and a Ph.D. degree (2003) in
manufacturing engineering from Boston University. He worked for the Mitsubishi
Electric Research Laboratories (MERL) (Cambridge, MA) in 2003 as a research intern
and as a research associate for the Manufacturing Engineering Dept. at Boston
University in 2004. He joined the Industrial Engineering Dept. at Isik University in
2005 as an instructor. Dr. Anlis current research interests include stochastic models
of supply chain manufacturing and online applications of e-learning in engineering
education focusing on virtual enterprise models.
Michael C. Caramanis (BS, Ch. Eng. Stanford University; MS and Ph.D., Engineering
and Decision and Control, Harvard University) is a professor in the College
of Engineering at Boston University, affiliated with the Dept. of Mechanical
Engineering, the Division of Systems Engineering, and the Center for Information
and Systems Engineering. He has served as editor of IIE Transactions on Design and
Manufacturing andis associate editor of several journals. Dr. Caramanis has beenthe
principal investigator on numerous National Science Foundation funded projects
and has collaborated widely with industry. His current research interests are in
the area of production planning and control of complex manufacturing systems
operating under uncertainty, with particular emphasis on the decomposition
of these systems to tractable subsystems and their coordination for enterprise
integration and supply chain management. Dr. Caramanis first career was in the
area of spot pricing of electricity and the liberalization of electricity markets, and
he is coauthor of the influential field initiating text Spot Pricing of Electricity, Kluwer
1988. From 2004 until recently, he served as the chair of the Greek Regulatory
Authority for Energy. Building on his energy systems research experience, he is
currently focusing on the creation of cyber infrastructure based real-time markets
capable of exploiting demand-side managements potential to provide the requisite
reserve services for significantly higher adoption of clean energy generation and the
electrification of the transportation sector.
Ioannis Ch. Paschalidis received the Diploma in ECE from the National Technical
University of Athens in 1991 and the SM and Ph.D. degrees in EECS from the
Massachusetts Institute of Technology in 1993 and 1996, respectively. He joined
Boston University in 1996, where he is an associate professor of electrical and
computer engineering and the codirector of the Center for Information and
Systems Engineering (CISE). He is also the academic director of the Sensor
Network Consortium (SNC)an industry consortium with companies active in
sensor networking. He has held visiting appointments with MIT and the Columbia
University Business School. Dr. Paschalidis current research interests lie in the
fields of systems and control, optimization, networking, operations research,
and computational biology. The main application areas he is targeting include
communication and sensor networks, supply chains, and protein docking. He has
received an NSF CAREER award (2000), the second prize in the 1997 George E.
Nicholson paper competition by INFORMS, and was an invited participant at the
2002 Frontiers of Engineering Symposium, organized by the National Academy of
Engineering. He is an associate editor of IEEE Transactions on Automatic Control and
of Operations Research Letters.

POMS AnliCaramanisPaschalidis

Uploaded by

Copyright:

Available Formats

POMS AnliCaramanisPaschalidis

Uploaded by

Document Information

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

POMS AnliCaramanisPaschalidis

Uploaded by

Copyright:

Available Formats

Journal of Manufacturing Systems 26 (2007) 116134

Contents lists available at ScienceDirect

denote, respectively, max

are positive). This

0 0 1614 0 0 2472 0 3 1457

6 7 1048 1 3 1588 5 75 903

0 10 989 0 0 1072 33 54 326

6 37 579 4 11 667 59 89 188

You might also like