Journal of the Operational Research Society

ISSN: 0160-5682 (Print) 1476-9360 (Online) Journal homepage: https://www.tandfonline.com/loi/tjor20

Capacity planning and warehouse location in

supply chains with uncertain demands

E Aghezzaf

To cite this article: E Aghezzaf (2005) Capacity planning and warehouse location in supply
chains with uncertain demands, Journal of the Operational Research Society, 56:4, 453-462,
DOI: 10.1057/palgrave.jors.2601834

To link to this article: https://doi.org/10.1057/palgrave.jors.2601834

Published online: 21 Dec 2017.

Submit your article to this journal

Article views: 999

View related articles

Citing articles: 2 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tjor20
Journal of the Operational Research Society (2005) 56, 453–462 r 2005 Operational Research Society Ltd. All rights reserved. 0160-5682/05 $30.00

www.palgrave-journals.com/jors

Capacity planning and warehouse location

in supply chains with uncertain demands
E Aghezzaf*
University of Ghent, Technologiepark, Belgium
We discuss the strategic capacity planning and warehouse location problem in supply chains operating under
uncertainty. In particular, we consider situations in which demand variability is the only source of uncertainty. We first
propose a deterministic model for the problem when all relevant parameters are known with certainty, and discuss
related tractability and computational issues. We then present a robust optimization model for the problem when the
demand is uncertain, and demonstrate how robust solutions may be determined with an efficient decomposition
algorithm using a special Lagrangian relaxation method in which the multipliers are constructed from dual variables of
a linear program.
Journal of the Operational Research Society (2005) 56, 453–462. doi:10.1057/palgrave.jors.2601834
Published online 13 October 2004

Keywords: capacity planning; warehouse location; distribution; robust optimization

Introduction total distribution and transfer costs, while satisfying given

market demands. We then discuss properties of the
Production capacity planning and warehouse location
deterministic model and possible solution methods.
decisions are among the most significant capital investment
The uncertainty inherent in predicting unstable market
decisions that supply chain managers must periodically
demands makes it not viable and unrealistic for managers to
make. These decisions are crucial for supply chains
make important and vital capacity expansion investments
operating in a volatile business environment. In this paper,
and warehouse location decisions only on the strength of
we discuss the multi-period strategic capacity planning and
deterministic solutions. In fact, a solution that is optimal for
warehouse location problem for supply chains serving one particular forecast of future market demand may turn
markets with uncertain and unpredictable demands. In out to be far from optimal and very costly if this demand
particular, we are concerned with supply chains having a does not materialize. We therefore propose to extend the
given set of plants producing for potential interconnected deterministic model so that the resulting model can be used
warehouses, that in turn serve a set of markets with changing to generate capacity expansion programs and warehouse
demands. The two important decisions to be made are (1) location plans that are ‘robust’ to uncertain realizations of
the amounts of capacity to be installed in the production future market demand. In other words, we seek capacity
plants in each period; and (2) where to locate and operate expansion programs and warehouse location plans which
the warehouses. We assume that there is no lead-time to may not be optimal for a particular realization of market
acquire any amount of capacity decided in any period. We demand, but are probably the least expensive to adapt for
also allow some warehouses to transfer some or their entire any highly probable realization of market demand. To obtain
inventory to other warehouses at the end of a period for an these robust programs and plans we use the concept of robust
effective utilization of the available production capacity and optimization combined with Lagrangean relaxation methods.
an improved management of the warehousing capacity. A recent paper by Mulvey et al1 discussed the notion of
We first develop a deterministic capacity planning and robust optimization for planning large-scale systems that are
warehouse location model for the supply chain when market subject to noisy, uncertain, or incomplete input data. In
demands are known with certainty. The objective of the particular, they developed a general modelling and optimi-
model is to find an optimal capacity expansion program and zation framework that controls the sensitivity of an optimal
warehouse location plan that minimizes total capacity solution to noise (or variations) in the input data. Malcolm
investment and warehouse operating costs augmented by and Anandalingam2 used robust optimization for power
systems capacity expansion, and proposed a methodology
*Correspondence: E Aghezzaf, Department of Industrial Management, that integrates the capacity expansion process directly with
Campus Ardoyen, University of Ghent, Technologiepark 903, 9052
Gent-Zwijnaarde, Belgium. decision-analysis criteria for determining robustness. In this
E-mail: ElHoussaine.Aghezzaf@Ugent.be way, they can find the most robust possible plan relative to
454 Journal of the Operational Research Society Vol. 56, No. 4

other plans, and they can also explicitly determine the cost of u0i initial capacity in plant i at the beginning of the
robustness. This paper extends their approach to the planning horizon
integrated capacity planning and warehouse location pro- mti the maximum capacity expansion in plant i during
blem and tries to develop robust plans for supply chains with period t
uncertain demands. ktj capacity of warehouse j in period t
The remainder of the paper is organized as follows: first dkt demand of market k in period t (assumed to be
we present a deterministic capacity planning and warehouse stochastic and discrete)
location model when the demand is known with certainty
Model variables
and discuss possible solution methods. Next, we review the
general modelling framework of robust optimization, and xti binary decision variable set to 1 if the capacity in
extend the deterministic capacity planning and warehouse plant i is to be expanded during period t, and 0
location model to include uncertainty inherent in predicting otherwise
volatile market demands. We then present a numerical uti capacity expansion level for plant i in period t
example and some computational results, before concluding ytj binary decision variable set to 1 if warehouse j is to
the paper with a discussion of the results and possible be operated during period t, and 0 otherwise
extensions. vtij proportion of the available capacity in plant i
assigned to warehouse j in period t
wtjk proportion of market k demand dtk that is satisfied
The deterministic capacity planning and warehouse from the warehouse j in period t
location problem (CPWLD) fjjt 0 amount transferred from warehouse j to warehouse
We now present the optimization model for the strategic j 0 at the end of period t
capacity planning and warehouse location problem. The The model
main distinguishing features of our model from the usual
capacity expansion planning models discussed in literature (CPWLD):
are: !
XX X
Minimize ZIP ¼ ati xti þ bti uti þ gtij vtij

The proposed model integrates the issue of capacity t2T i2I j2J
expansion with distribution through the location of !
XX X X
warehouses. þ ctj ytj þ jtjj0 fjjt0 þ dtjk wtjk

The model considers transfer of commodities between t2T j2J j 0 2J k2k
warehouses allowing for a better management of ware-
house capacities and holding cost optimization. subject to

X X
t
The following mixed-integer programming formulation vtij  uti pu0i 8i; t ð1Þ
provides a deterministic model for the problem when j2J t¼1
demands are known with certainty. The set I represents
the facilities for which a capacity expansion program is to be
uti  mti xti p0 8i; t ð2Þ
developed, J is the set of potential warehouse locations, and
K is the set of markets with known demands that should be X X X X
satisfied during the planning horizon T. vtij þ fjt1
0j  wtjk  fjjt 0 ¼ 0 8j; t ð3Þ
i2I j 0 2J k2K j 0 2J

Model parameters X X
wtjk þ fjjt 0  ktj ytj p0 8j; t ð4Þ
ati discounted fixed cost for capacity expansion in plant k2K j 0 2J
i in period t
bti discounted variable cost for capacity expansion in X
wtjk ¼ dkt 8k; t ð5Þ
plant i in period t
k2K
ctj fixed cost of operating warehouse j during period t
jtjj0 transfer cost from warehouse j to warehouse j 0 at the
end of period t when jaj 0 , and inventory holding uti ; vtij ; fjjt 0 ; wtjk X0; xti ; ytj 2 f0; 1g 8i; j; j 0 ; k; t
cost in warehouse j at the end of period t when j ¼ j 0
gtij shipment cost from plant i to warehouse j during Constraints (1) ensure that, for any plant iAI, in any
period t period tAT, the initial capacity available in plant i
dtjk shipment cost from warehouse j to market k during augmented by the cumulated capacity expansion amounts
period t till period t, is sufficient to cover the quantities distributed to
 E Aghezzaf—Capacity planning and warehouse location 455

the warehouses. Constraints (2) reflect the maximum procurement-planning problem in knockdown production
capacity expansion levels planned for the plant i in period systems.15 The problem is formulated as a mixed-integer
t. Constraints (3) are the in- and outflows of materials linear program and solved by a two-stage heuristic
through each warehouse. Constraints (4) reflect the ware- procedure using branch-and-cut.
house capacities and constraints (5) are the flow distribution
restrictions from the warehouses to the markets.
Clearly the problem (CPWLD) belongs to the class of NP- A brief review of the robust optimization paradigm
hard problems, since it can be viewed as a multiple-source There has been, in recent years, a considerable interest in
capacitated economic lot-sizing problem. A special case of robust solutions for many decision problems. This increasing
the latter problem is the single-source capacitated lot-sizing interest is motivated by the fact that important parameters
problem known to be NP-hard.3 To solve this difficult of these decision problems depend heavily on the changing
problem many branch-and-bound, branch-and-cut, and realization of the future. Recently, Mulvey et al1 elaborated
branch-and-price exact methods as well as many heuristic on the notion of robust optimization and presented a
methods were proposed. Wolsey4 reviewed the capacitated framework to address the problem of finding robust
lot-sizing problem and some of its extensions, and provided solutions to some stochastic optimization problems. Using
efficient formulations for a wide variety of these problems. the standard terminology in stochastic programming, we let
The problem (CPWLD) contains also the facility location x denote the design variables whose optimal values are
problem as a sub-problem. This latter problem has also independent of any realization of the uncertain parameters.
attracted considerable attention. In particular, Cornuéjols We then let y denote the control variables that can be
et al5 presented a survey of some mathematical program- adjusted once the uncertain parameters are observed. The
ming models for the uncapacitated version of this facility optimal values of control variables depend both on the
location problem. Many specific branch-and-bound algo- realization of the uncertain parameters, and on the values of
rithms, including dual and primal-dual approaches, were the design variables. Now, consider the following optimiza-
developed for this version of the problem.6 Akins and tion problem LP:
Khumawala7 presented an efficient branch-and-bound algo-
rithm for the capacitated version of the problem. The pure (LP):
capacity expansion problem is also a sub-problem of Minimize cx þ dy
(CPWLD), also extensively studied, Cormier and Gunn8 ð6Þ
Subject to Ax ¼ b
developed a dynamic programming model to establish a
warehouse capacity expansion schedule and underlying Ex þ Fy ¼ g
multi-item inventory policy that are jointly optimal. Lucas ð7Þ
x; yX0
et al9 studied the capacity planning problem with uncertain
demands and used Lagrangian relaxation to solve it. where c, d, A, b, E, F and g are parameters that define the
Rajagopalan et al10 studied the problem of capacity inputs of the model. The first constraints denote the design
expansion and replacement in markets with uncertain constraints that are not impacted by uncertainty. The other
technological breakthroughs. They formulated and solved constraints denote the control constraints in which the
the problem using a dynamic programming algorithm that coefficients may be subject to uncertainty.
takes advantage of some structural properties of the To define a robust optimization problem associated with
formulation. the mathematical program (LP), let PS ¼ {1, 2,y, S} be a set
Various authors have also tackled the strategic capacity of future possible scenarios, each with a probability of
P
planning integrated with facility location and distribution. occurrence ps such that s2PS ps ¼ 1. Also, for each
The integrated capacity planning and facility location scenario sAPS, let {Es, Fs, ds, gs} be the set of realizations
problem consists of determining where to locate the facilities, for the coefficients of the control constraints and the
how much capacity to install in each location and how to objective function of the mathematical program (LP).
assign customers to these facilities over a multi-period An optimal solution of the mathematical program (LP) is
planning horizon. Among others Fong and Srinivasan11,12 considered to be ‘solution-robust’, if it remains ‘close’ to
considered a multi-period location and capacity planning optimal for any realization of the scenario sAPS. If this
model where fractional assignment of customer demands is solution remains ‘almost’ feasible for any realization of
allowed, and proposed a heuristic solution for it. Klincewicz sAPS, it is considered to be ‘model-robust’. Of course, it is
et al13 proposed an iterative heuristic procedure for solving unlikely that a solution for problem (LP) remains both
the dynamic location and capacity planning problem where feasible and optimal for any realization of sAPS. It is thus
fractional assignment of customer demands is not allowed. necessary for a robust model to permit the measurement of a
Lim and Kim14 proposed a Lagrangean relaxation-based trade-off between solution and model robustness. The
branch-and-bound strategy for the exact solution of this following model formalizes a way to measure this tradeoff.
problem. They have also studied the plant location and It also controls both solution and model robustness.
456 Journal of the Operational Research Society Vol. 56, No. 4

Let {y1, y2,y, yS} be the control variables for each assume that the capacity planning and warehouse location
scenario sAPS. Given the realization of different scenarios, decision variables (for the entire planning horizon) are the
there is no guarantee that the control constraints will always design variables valid for all scenarios, and we let the
be satisfied. Therefore, ‘error variables’ {e1, e2,y, eS} are distribution decisions form the control variables, which can
introduced to measure the infeasibility in the control be adjusted once the demand is observed. Thus, the capacity
constraints under scenario s. Now, consider the following expansion and warehouse location plan is decided, and then
formulation of the robust optimization model (RP): optimal decisions regarding distribution patterns are made
for each scenario realization.
Minimize sðx; y1 ; y2 ; :::; ys Þ þ orðe1 ; e2 ; :::; es Þ Consequently, we will have distribution variables for each
ð6Þ
Subject to Ax ¼ b
scenario, which means vtij, fjjt, and wtjk get an extra index s and
become vts ts ts
ij , fjj 0 , and wjk. We also introduce the error variables
Es x þ Fs y þ es ¼ gs ts ts
ei and ej to indicate any capacity shortages in the plants and
ð8Þ
x; ys X0 warehouses. We assume that all parameters except the
demand are the same for different scenarios, thus only dtk
The first term of the objective function measures solution- gets an extra index s and becomes dts k . Finally, two extra
robustness, the second term measures model-robustness. The parameters are introduced, namely o1 and o2, the costs per
o is a weighting factor that measures the relative importance unit capacity shortage in the plants and warehouses.
of obtaining a model-robust solution versus a solution-
robust solution. Note that there is a variety of choice for the
cost function s(x, y1, y2,y, yS) and the feasibility penalty A robust optimization model
function r(e1, e2,y, eS) used to penalize violations of the
control constraints. It is, however, necessary that the chosen To present the robust optimization model for the capacity
functions lead to consistent preferences between alternative planning and warehouse location problem, let us first define
decisions. for each scenario the objective function xs
An example of a solution-robustness measure is the !
XX X
t t t t t ts
maximum regret function sðx; yÞ ¼ maxs2PS ðxs  xs Þ where xs ¼ ai xi þ bi ui þ gij vij
x is the objective function to be minimized. The value xs is t2T i2I j2J
!
the objective function value given that the design plan is XX X X
made (eg, a particular capacity expansion plan) and then þ ctj ytj þ jtjj 0 fjjts0 þ dtjk wts
jk
t2T j2J j 0 2J
scenario s becomes true. The value x*s is the objective k2k

function value for the optimal plan had we known that Then, this is the robust model
scenario s was going to be true. This measure ensures that (CPWLR):
the expected greatest deviation of the robust solution from !
the scenario optimal solution is minimized. Alternatively, a X XX XX
positive power of the difference between the objective value Minimize ZR ¼ ps xs þ o1 ets
i þ o2 ets
j
s2PS t2T i2I t2T j2J
of implementing the design plan and the scenario optimal
P subject to
objective could be used: sðx; yÞ ¼ s2PS ps ðxs  xs Þq . This
function reduces to minimizing expected regret when q X X
t

i pui 8i; t; s
uti  ets 0
vts
ij  ð9Þ
equals 1, and is the same as the stochastic linear program- j2J t¼1
ming objective. Values of q that are higher than 1 lead to
risk-averse measures of solution-robustness. uti  mti xti p0 8i; t ð2Þ
For model-robustness measures, functions that force the
error (slack or surplus) in the constraints to be close to zero X X X X
vts
ij þ fjt1;s
0j  wts
jk  fjjts0 ¼ 0 8j; t; s ð10Þ
for all scenarios are typically sought. For instance, to i2I j 0 2J k2K j 0 2J
measure model-robustness one could use the function rðeÞ ¼ X X
j p0 8j; t; s
maxs2PS ðkes kÞ to make sure the maximum violation is wts
jk þ fjjts0  ktj ytj  ets ð11Þ
minimized. Alternatively, the expected violation rðeÞ ¼ k2K j 0 2J
P
s2SP ps ðkes kÞ can also be minimized. X
wts ts
jk ¼ dk 8j; t; s ð12Þ
k2K
The robust capacity planning and warehouse location 0
problem uti ; vts ts ts ts ts t t
ij ; fjj 0 ; wjk ; ei ; ej X0; xi ; yj 2 f0; 1g 8i; j; j ; k; t; s

To incorporate demand uncertainty in the development of To tackle the stochastic capacity planning problem, some
optimal capacity expansion and warehouse location plans, authors adopted a two-stage stochastic programming
we adopt the concept of robust optimization. Thus, we approach.16 In this approach, it is assumed that capacity
 E Aghezzaf—Capacity planning and warehouse location 457

planning and warehouse location decisions have to be made uts t

i  ui ¼ 0 8i; t; s ð16Þ
here and now. Once capacities are decided, time unfolds and
a certain scenario of demand realizes, and then the optimal yts t
j  yj ¼ 0 8j; t; s ð17Þ
decision regarding distribution plans can be made. The
approach used in this paper for the solution of model 0
uti ; uts ts ts ts t ts t ts
i ; vij ; fjj 0 ; wjk X0; xi ; xi ; yj ; yj 2 f0; 1g 8i; j; j ; k; t; s
(CPWLR) consists of reformulating the problem and
applying the Lagrangean relaxation method to make it
decomposable into sub-problems. The following paragraph where the objective function xs’ is given as follows:
!
presents the proposed decomposition-based algorithm. 0
XX X
t ts t ts t ts
xs ¼ ai xi þ bi ui þ gij vij
t2T i2I j2J
!
A decomposition-based algorithm for the CPWLR XX X X
problem þ ctj yts
j þ jtjj 0 fjjts0 þ dtjk wts
jk
t2T j2J j 0 2J k2k
A careful analysis of the robust optimization model
(CPWLR) presented above, shows that the variables x, u, The reformulated model has a large number of additional
and y, and constraints (2) are the only components linking variables and restrictions. Constraints (15)–(17) guarantee
the solutions of the capacity planning and warehouse that the ‘optimal’ capacity expansion and warehouse
location problem for each scenario. Constraints (2) state location plans are the same for all scenarios. Note also
that the capacity expansion plan must be valid for all that some of the constraints are redundant but were kept
scenarios. If we relax these restrictions, the resulting problem in the model. These constraints will be used to provide
reduces to a collection of ‘deterministic’ problems, one for us with an alternative way to generate the Lagrangean
each scenario. We expect then that the Lagrangean multi- multipliers.
pliers acting in each scenario problem will lead to an Now, observe that if constraints (15)–(17) are relaxed and
‘optimal’ robust solution. To exploit this observation, we dualized with their corresponding Lagrangean multipliers
R
reformulate the model (CPWLR) as explained below. lts ts ts
i , mi , and Zj , the problem (CPWLR) decomposes into
We introduce new variables xts ts ts
i , ui , and xj defining, (|PS| þ 1) sub-problems. The first |PS| sub-problems are
respectively, for each scenario sAPS, the decision to expand exact replications of the model (CPWLD) for each scenario s,
the capacity of plant i in period t, the amounts of capacity except for error variables, and the last sub-problem deals
expansion in plant i in period t, and the decision to open with determining the common capacity expansion and
warehouse j during period t. Using these variables we can warehouse location plans that are valid for all scenarios.
rewrite the model (CPWLR) as follows: More precisely, for each scenario s 2 PS we get a sub-
problem defined by the following model:
(CPWLR
R):
! (CPWLLR
R (s)):
X 0
XX XX
Minimize ZR ¼ ps xs þ o1 ets
i þ o2 ets
j 00
XX XX
s2PS t2T i2I t2T j2J Minimize Z LR ðsÞ ¼ xs þ ps o1 ets
i þ ps o2 ets
j
t2T i2I t2T j2J
subject to
subject to
X X
t
X X
t
i  ei pui 8i; t; s
uts 0
vts
ij 
ts
ð9Þ
i  ei pui 8i; t
uts 0
vts
ij 
ts
ð9Þ
j2J t¼1
j2J t¼1

i  mi xi p0 8i; t; s
uts t ts
ð13Þ
i  mi xi p0 8i; t
uts t ts
ð13Þ
X X X X
vts
ij þ fjt1;s
0j  wts
jk  fjjts0 ¼ 0 8j; t; s ð10Þ X X X X
i2I j 0 2J k2K j 0 2J vts
ij þ fjt1;s
0j  wts
jk  fjjts0 ¼ 0 8j; t ð10Þ
X X i2I j 0 2J k2K j 0 2J
wts j  ej p0 8j; t; s
fjjts0  ktj yts ts
jk þ ð14Þ
X X
j 0 2J
j  ej p0 8j; t ð14Þ
k2K
wts
jk þ fjjts0  ktj yts ts

X k2K j 0 2J
wts ts
jk ¼ dk 8j; t; s ð12Þ
k2K X
wts ts
jk ¼ dk 8j; t ð12Þ
uti  mti xti p0 8i; t ð2Þ k2K

0
xts t
i  xi ¼ 0 8i; t; s ð15Þ uts ts ts ts ts ts
i ; vij ; fjj 0 ; wjk X0; xi ; yj 2 f0; 1g 8i; j; j ; k; t
458 Journal of the Operational Research Society Vol. 56, No. 4

where the objective function xs00 is given as follows: follows:

(CPWLLR R (s)LP):
XX
00
xs ¼ ðps ati  lts ts t ts ts XX
i Þxi þ ðps bi  mi Þui
0
LR
t2T i2I
Minimize ZLP ðsÞ ¼xs þ ps o1 ets
i
! t2T i2I
X XX XX
þ ps gtij vts þ ðps cti  Zts ts þ ps o2 ets
ij j Þyj j
j2J t2T j2J t2T j2J
!
X X
þ ps jtjj 0 fjjts0 þ ps dtjk wts subject to
jk
j 0 2J k2k
X X
t

i  ei pui 8i; t
uts 0
vts
ij 
ts
ð9Þ
The last sub-problem resulting from the decomposition is
j2J t¼1
given by:
(CPWLLRR (0)):
i  mi xi p0 8i; t
uts t ts
! ð13Þ
PP P
Minimize Z LR ð0Þ ¼ lts
i xti X X X X
t2T i2I
! ! s2PS
vts
ij þ fjt1;s  wts
jk  fjjts0
P i2I j 0 2J
0j

k2K j 0 2J
þ mts
i uti
s2PS
! ¼ 0 8j; t ð10Þ
PP P
þ Zts
j ytj X X
wts j  ej p0 8j; t ð14Þ
fjjts0  ktj yts ts
jk þ
t2T j2J s2PS
subject to k2K j 0 2J

uti  mti xti p0 8i; t ð2Þ

X
wts ts
jk ¼ dk 8j; t ð12Þ
uti X0; xti ; ytj 2 f0; 1g 8i; j; t k2K

With these last two models, we are now equipped to

generate robust solutions for the capacity expansion and xts ^ti ¼ 0 8i; t ð15Þ
i x
warehouse location problem. The following section discusses
the steps of the proposed solution algorithm.
uts ^ti ¼ 0 8i; t ð16Þ
i u

The algorithm yts ^tj ¼ 0 8j; t ð17Þ

j y
The problem (CPWLLR R (s)) is an NP-complete problem and
is hard to solve. Moreover, since this problem is to be solved
i ; vij ; fjj 0 ; wjk X0; 0pxi p1;
uts ts ts ts ts
for each scenario s, the computational time will increase 0
dramatically as the number of scenarios gets larger. To j p1 8i; j; j ; k; t
0pyts
resolve this difficulty, we propose an alternative approach to
Observe here that the above optimization problem
obtain good approximations to optimal robust solutions of
(CPWLLR R (s)LP) is a linear program, we can then
the original problem (CPWLR) using the models resulting
obtain the dual multipliers for the constraints (15),
from the above Lagrangian decomposition.
(16) and (17). The new Lagrangian multipliers are
updated by augmenting the current Lagrangian
Step 0: First, we start with all Lagrangian multipliers multipliers lts ts ts
i , mi , and Zj , with the dual multi-
taken to be lts ts ts
i ¼ mi ¼ Zj ¼ 0, for all t, i, j and s. pliers of the constraints (15), (16) and (17)
We also start with the most expensive capacity multiplied by an appropriate factor. With these
expansion and warehouse location plans obtained new Lagrangian multipliers we proceed to step 2.
by letting the decision variables xti ¼ x^ti ¼ 1,
uti ¼ u^ti ¼ mti , and ytj ¼ y^tj ¼ 1. We then proceed Step 2: To determine the common new capacity expansion
to Step 1. and warehouse location plans xti , uti , and ytj , we
solve the integer program (CPWLLR R (0)) with the
Step 1: With these values of lts ts ts
^ti , u^ti , and y^tj , we
i , mi , Zj , x new Lagrangian multipliers lts i , mts ts
i , and Zj as
solve the linear programming relaxation of the objective costs. We then let the new values of
0
model (CPWLLR R (s)), for each scenario s, with xs x^ti ¼ xti , u^ti ¼ uti , and y^tj ¼ ytj . These values are in
00
instead of xs and some additional restrictions as turn passed on to Step 1.
 E Aghezzaf—Capacity planning and warehouse location 459

This process is repeated until no significant change in the
Shipment costs from plants to warehouses are constant
P
total cost s2PS ps Z
LR
ðsÞ þ ZLR ð0Þ is achieved. The and are set to gtij ¼ 1.
procedure is then stopped and the resulting plans are taken
Fixed cost of operating warehouses 1–3 during periods
as the best robust capacity expansion and warehouse 1–5, and the planned warehouse capacities during this
location plans. planning horizon (Table 2).
At this point, we need to make two important remarks.
Transfer cost from a warehouse to a warehouse at the end
The first, is that the solution of the problem (CPWLLR R (0)) of each period (Table 3).
requires integer solutions but the number of variables is
Shipment cost from warehouses to markets during the
small and the problem is not difficult to solve compared to planning periods (Table 4).
the original problem (CPWLR). On the other hand, the
problem (CPWLLR R (s)LP), that must be solved for each Table 2 Warehouse fixed costs and capacities
scenario, becomes in our decomposition procedure a linear Planning horizon
program that can be solved efficiently. Therefore, we can
reasonably expect that the amount of time and space 1 2 3 4 5
required to solve the original problem (CPWLR), directly
(a) Fixed cost warehouse
with a commercial optimizer, be much larger than the 1 100 25 25 100 25
amount of time and space required by our special 2 50 200 200 50 50
decomposition procedure. It is also important to report that 3 100 25 25 100 25
to insure the convergence of the above procedure we needed
to solve occasionally the integer program (CPWLLR (b) Capacity warehouse
R (s)) and
1 10 25 30 10 10
combine the Lagrangian multipliers obtained with the 2 50 50 50 50 50
classical sub-gradient method with our method described 3 10 25 30 10 10
in Step 1.
The following section presents an example to illustrate the
various steps involved in our decomposition-based algo- Table 3 Warehouse to warehouse transfer costs
rithm. The section thereafter presents the results of a Planning horizon
computational study conducted to measure the computa-
tional time and solution quality of our decomposition-based From warehouse To warehouse 1 2 3 4 5
algorithm against the average solution approach, and the 1 1 1 1 1 1 1
direct solution of the original problem (CPWLR). These tests 2 2 2 2 2 2
are carried out on a set of randomly generated problems 3 2 2 2 2 2
with specific settings for some critical parameters. 2 1 2 2 2 2 2
2 1 1 1 1 1
3 2 2 2 2 2
3 1 2 2 2 2 2
A computational example 2 2 2 2 2 2
3 1 1 1 1 1
We consider two facilities producing the same product, and
three potential warehouses serving five markets. We also
consider a planning horizon of five periods. The necessary Table 4 Warehouses to markets shipment costs
data are given below:
Planning horizon

Discounted capacity investment fixed costs are constant
From warehouse To market 1 2 3 4 5
and are set to ati ¼ 50.

Discounted variable investment costs are constant and are 1 1 2 2 2 2 2
set to bti ¼ 1. 2 1 2 1 1 2

Initial capacity in each plant and the maximum allowable 3 2 1 2 2 1
4 1 2 1 1 2
expansion level per period (Table 1). 5 2 2 2 2 2
2 1 2 1 2 2 2
2 2 2 2 2 1
3 1 2 1 2 2
Table 1 Initial capacity and expansion levels 4 2 1 2 1 2
Planning horizon 5 2 2 2 2 2
3 1 2 2 2 2 1
Plant Initial 1 2 3 4 5 2 1 1 1 1 2
3 2 2 2 2 2
1 10 5 5 10 5 5 4 2 1 1 1 2
2 10 5 5 5 5 5 5 1 2 2 2 1
460 Journal of the Operational Research Society Vol. 56, No. 4

Table 5 Market demand scenarios Table 6 Results of the three solution approaches
Demands Planning horizon Adapted
Solution of average The
Scenario Market 1 2 3 4 5 the problem solution decomposition
(CPWLR) (CPWLD) procedure
1 1 5 5 10 5 5
2 2 2 10 2 2 Solution value 1120 1361.2 1330
3 10 10 15 10 10 Computation 0.09 0.07 2.04
4 3 3 10 3 3 time (s)
5 10 10 15 10 10 x Period Values Values Values
2 1 6 7 9 8 3 Plant 1 1 1 1 1
2 1 1 5 1 3 2 1 0 1
3 13 12 18 10 12 3 1 1 1
4 4 6 8 4 4 4 0 0 0
5 8 12 15 7 9 5 0 0 0
3 1 5 5 4 4 3 Plant 2 1 1 1 1
2 1 2 3 1 2 2 1 1 1
3 9 11 12 12 12 3 0 0 1
4 6 4 2 4 6 4 0 0 0
5 13 7 12 13 10 5 0 0 0
4 1 3 6 7 5 4
2 1 3 1 2 2 u
3 11 8 10 7 12 Plant 1 1 5 5 5
4 4 6 2 4 2 2 5 0 5
5 12 10 13 7 12 3 10 10 10
5 1 7 5 8 7 6 4 0 0 0
2 1 3 1 2 3 5 0 0 0
3 11 12 13 13 11 Plant 2 1 5 5 5
4 6 5 2 5 5 2 5 3 5
5 7 10 10 9 7 3 0 0 5
4 0 0 0
5 0 0 0

Demands of the markets: The above table shows the
demands making the five scenarios considered for this y
Warehouse 1 1 0 0 0
illustrative example (Table 5). 2 1 1 1
3 1 1 1
We have determined the robust solution for this 4 0 0 0
illustrative example using the model (CPWLR) solved 5 0 0 0
directly with the commercial optimizer Cplex 8. We then Warehouse 2 1 1 1 1
determined the average solution obtained by solving the 2 0 0 0
3 0 0 1
deterministic problem (CPWLD) with average demands, and
4 1 1 1
we computed the average cost of this solution by averaging 5 1 1 1
costs obtained by adapting this average solution to each Warehouse 3 1 0 0 0
scenario s. Finally, we solved the problem with our 2 1 1 1
decomposition-based algorithm. The findings of these three 3 1 1 1
4 0 0 0
approaches are reported in Table 6. 5 0 0 0
As can be seen in this table our decomposition-based
algorithm provides a better solution value than the average
solution approach. The time required by our procedure is
higher than that required by the other two approaches. out a significant number of experiments on large instances to
However, this is due to the small number of scenarios analyse the behaviour of our algorithm, we have developed a
considered in this illustrative example. This computational design of experiment in which we distinguished three critical
time is generally much smaller than the time required to factors. The first factor is intended to reflect the effect of the
solve directly the model (CPWLR) when the number of problem size. Three levels were assumed for this factor: S for
scenarios gets larger. small-size problems with two plants, four warehouses and six
markets, M for medium-size problems with four plants, eight
warehouses and 10 markets, and Lr for large-size problems
Design of experiments and computational results
with five plants, 10 warehouses and 20 markets. The second
Recognizing the limitation of the above results from a factor is meant to reflect the effect of demand variability. It
performance analysis viewpoint, and the necessity to carry also has three levels: Lo for low demand variability, namely
 E Aghezzaf—Capacity planning and warehouse location 461

10% of the demand mean, M for medium demand algorithm to the optimal value obtained with the model
variability, 25% of demand mean and H for high demand (CPWLR).
variability, 40% of demand mean. The last factor is the Table 7 shows the computational times, and as expected
number of scenarios involved in the problems. We have also when the number of scenarios gets larger our decomposition-
chosen three levels for this factor: S for small number of based algorithm performs better in comparison to the direct
scenarios, 25 scenarios, M for medium number of scenarios, solution of the model (CPWLR). This is typically due to the
50 scenarios, and L for large number of scenarios, 75 fact that the large size sub-problems dealing with the
scenarios. The planning horizon is composed of six periods scenarios are solved as linear programs in our decomposi-
and the other fixed parameters are chosen randomly in the tion-based algorithm. We avoid in this manner the time-
same magnitude as in the example above. consuming branch-and-bound phase necessary to solve the
For each combination of these factors, five problems model (CPWLR). Of course for small problems the search
are generated and solved with three approaches. Namely, procedure, in our decomposition-based algorithm, for
the direct solution of the model (CPWLR) with a commercial optimal multipliers takes more time than the direct solution
optimizer (Cplex 8), the average solution approach consis- of the model (CPWLR).
ting in solving the deterministic problem (CPWLD*) Table 8 shows that our decomposition procedure per-
for average demands and then adapting the resulting forms better than the average solution, and that it generally
solution to each scenario to obtain the average cost, and produces better robust solutions. There are, however, some
finally with our proposed decomposition-based algorithm cases in which the average solution approach performs
(DEC-Alg). The following tables show, for each combina- better. These cases are usually few problems with low-to-
tion of the above factors, the average computational time medium demand variability. In these cases the extra costs to
and the average gap ratio of the solution values obtained by adapt the average solution to each scenario are not
the average solution approach and the decomposition-based important since the differences between scenario demands

Table 7 Computational times of the three solution approaches

Problem size

Small-size demand Medium-size demand Large-size demand

Computational time (0.1 s) variability variability variability

Scenario Approach Lo M H Lo M H Lo M H

Small CPWLR 38 57 46 49 103 84 156 173 164

CPWLD 23 27 31 44 86 108 177 178 177
DEC-Alg 159 184 204 205 205 266 276 315 297
Medium CPWLR 40 119 110 179 231 240 185 287 346
CPWLD* 41 80 74 68 109 96 134 159 173
DEC-Alg 214 230 300 133 266 163 189 324 241
Large CPWLR 162 221 198 155 262 209 277 420 421
CPWLD* 68 88 130 195 198 202 282 295 288
DEC-Alg 273 278 284 336 338 344 342 393 397

Table 8 Gap ratios of the solutions CPWLD and DEC-Alg to CPWLR

Problem size

Solution gap to direct Small-size demand Medium-size demand Large-size demand

solution of CPWLR variability variability variability

Scenario Approach Lo M H Lo M H Lo M H

Small CPWLD* 1.31 1.9 2.43 1.32 1.39 2.5 1.48 1.87 2.27
DEC-Alg 1.15 1.21 1.31 1.11 1.16 1.35 1.11 1.12 1.2
Medium CPWLD* 1.16 1.41 2.29 1.4 1.99 2.12 1.29 1.7 2.35
DEC-Alg 1.28 1.35 1.49 1.31 1.7 1.72 1.08 1.59 1.16
Large CPWLD* 1.71 1.61 1.58 1.43 1.53 1.89 1.62 2 2.02
DEC-Alg 1.8 1.7 1.8 1.49 1.63 1.42 1.23 1.25 1.26
462 Journal of the Operational Research Society Vol. 56, No. 4

are not significant. To conclude, the direct solution of the Acknowledgement—We thank Birger Raa for conducting the initial
computational tests on an earlier version of the model.
model (CPWLR) leads to the optimal solution, but it
requires a significant amount of computational time, the
average solution approach is fast in terms of computational
time, but it generally generates more expensive solutions. References
Our decomposition-based approach falls in-between; it
1 Mulvey JM, Vanderbei RJ and Zenios SA (1995). Robust
produces rather good solutions with a reasonable amount
optimization of large-scale systems. Opns Res 43: 264–281.
of computational time. 2 Malcolm S and Anandalingam G (2001). Robust optimiza-
tion for power sector planning in large countries. Opns Res
(forthcoming).
3 Bitran GB and Yanasse HH (1982). Computational complexity
Conclusion of the capacitated lot size problem. Mngt Sci 28: 1174–1185.
In this paper, we studied the capacity planning and 4 Wolsey LA (2002). Solving multi-item lot-sizing problems with
an MIP solver using classification and reformulation. Mngt Sci
warehouse location problem with uncertain demands. We 48: 1587–1602.
proposed a deterministic model for this problem when 5 Cornuéjols G, Nemhauser GH and Wolsey LA (1990). Discrete
demands are known with certainty and we discussed some of Location Theory. Lecture Note in Artificial Intelligence (LNAI
its computational properties. We then moved to the case of 1865). Wiley: New York.
6 Koerkel M (1989). On the exact solution of large-scale simple
uncertain demands and we developed a robust model based
plant location problems. Eur J Opns Res 39: 157–173.
on the concept of robust optimization. To solve this model, 7 Akins U and Khumawala BM (1996). An efficient branch and
we developed an algorithm based on the Lagrangean bound algorithm for the capacitated warehouse location
relaxation decomposition. This decomposition resulted in problem. Mngt Sci 23: 585–594.
two separate sub-problems, the first sub-problem contains a 8 Cormier G and Gunn EA (1999). Modelling and analysis for
capacity expansion planning in warehousing. J Opl Res Soc 50:
set of replications of the deterministic model for each
52–59.
scenario, and the second sub-problem determines the 9 Lucas C, MirHassani SA, Mitra G and Poojari CA (2001). An
common capacity expansion and warehouse location plans. application of Lagrangian relaxation to a capacity planning
Based on this decomposition, we developed a linear problem under uncertainty. J Opl Res Soc 52: 1256–1266.
program that provides us with upper bounds on the original 10 Rajagopalan S, Singh MR and Morton TE (1998). Capacity
expansion and replacement in growing markets with uncertain
problem and the Lagrangian multipliers to be used in the technological breakthroughs. Mngt Sci 44: 12–30.
Lagrangian relaxation procedure. The lower bounds are 11 Fong CO and Srinivasan V (1981). The multiregion dynamic
obtained, as usual, by solving the Lagrangian dual. This capacity expansion problem: Part I. Opns Res 29: 787–799.
approach is tested on a randomly generated set of problems, 12 Fong CO and Srinivasan V (1981). The multiregion dynamic
and is compared to the average solution approach and to the capacity expansion problem: Part II. Opns Res 29: 800–816.
13 Klincewicz JG, Luss H and Yu CS (1988). A large-scale
solution of the original robust model solved directly with a multilocation capacity planning model. Eur J Oper Res 34:
commercial solver. The results are reported and the major 178–190.
conclusion of these findings can be stated as follows: the 14 Lim SK and Kim YD (1999). An integrated approach to
direct solution of the original robust model produces dynamic plant location and capacity planning. J Opl Res Soc 50:
1205–1216.
optimal robust solutions, but it requires a significant amount
15 Lim SK and Kim YD (2001). Plant location and procurement
of computational time, the average solution approach is fast planning in knockdown production systems. J Opl Res Soc 52:
in terms of computational time, but it generally generates 271–282.
more expensive solutions. Our decomposition-based algo- 16 Birge JR and Louveaux F (1997). Introduction to Stochastic
rithm produces rather good solutions, which are generally Programming. Springer: New York.
better than those of the average solution approach, and are
obtained within a reasonable amount of computational time Received November 2003;
compared to the solution of the original robust model. accepted June 2004 after two revisions