European Journal of Operational Research 167 (2005) 96–115

www.elsevier.com/locate/dsw

Production, Manufacturing and Logistics

A stochastic programming approach for supply chain

network design under uncertainty
Tjendera Santoso, Shabbir Ahmed *, Marc Goetschalckx, Alexander Shapiro
School of Industrial & Systems Engineering, Georgia Institute of Technology, 765 Ferst Drive, Atlanta, GA 30332, USA
Received 25 June 2003; accepted 17 January 2004
Available online 18 May 2004

Abstract

This paper proposes a stochastic programming model and solution algorithm for solving supply chain network
design problems of a realistic scale. Existing approaches for these problems are either restricted to deterministic
environments or can only address a modest number of scenarios for the uncertain problem parameters. Our solution
methodology integrates a recently proposed sampling strategy, the sample average approximation (SAA) scheme, with
an accelerated Benders decomposition algorithm to quickly compute high quality solutions to large-scale stochastic
supply chain design problems with a huge (potentially infinite) number of scenarios. A computational study involving
two real supply chain networks are presented to highlight the significance of the stochastic model as well as the effi-
ciency of the proposed solution strategy.

2004 Elsevier B.V. All rights reserved.

Keywords: Facilities planning and design; Supply chain network design; Stochastic programming; Decomposition methods; Sampling

1. Introduction planning involves deciding the configuration of

the network, i.e., the number, location, capacity,
A crucial component of the planning activities and technology of the facilities. The tactical level
of a manufacturing firm is the efficient design planning of supply chain operations involves
and operation of its supply chain. A supply chain deciding the aggregate quantities and material
is a network of suppliers, manufacturing plants, flows for purchasing, processing, and distribution
warehouses, and distribution channels organized of products. The strategic configuration of the
to acquire raw materials, convert these raw mate- supply chain is a key factor influencing effi-
rials to finished products, and distribute these cient tactical operations, and therefore has a long
products to customers. Strategic level supply chain lasting impact on the firm. Furthermore, the fact
that the supply chain configuration involves the
commitment of substantial capital resources over
*
Corresponding author. Tel.: +1-404-894-2320; fax: +1-404-
long periods of time makes the supply chain
894-2301. network design problem an extremely important
E-mail address: sahmed@isye.gatech.edu (S. Ahmed). one.

0377-2217/$ - see front matter
2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2004.01.046
 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115 97

Beginning with the seminal work of Geoffrion ministic network design problems [8], to generate
and Graves [8] on multi-commodity distribution robust designs. Computational testing of the
system design, a large number of optimization- algorithm using randomly generated test problems
based approaches have been proposed for the de- with networks comprised of up to 33 nodes, 100
sign of supply chain networks (see for example arcs, 64 commodities, and 11 scenarios was carried
[1,9,32]). However, the majority of this research out. MirHassani et al. [22] considered a two-stage
assumes that the operational characteristics of, model for multi-period capacity planning of sup-
and hence the design parameters for, the supply ply chain networks. Here the first stage decisions,
chain are deterministic. Unfortunately, critical comprised of openings and closings of the plants
parameters such as customer demands, prices, and and distribution centers and setting their capacity
resource capacities are quite uncertain. Moreover, levels, are to be decided prior to the realization of
the arrival of regional economic alliances, for in- future demands. Then, based upon the particular
stance the Asian Pacific Economic Alliance and demand scenario realized, the production and
the European Union, have prompted many cor- distribution decisions are to be decided optimally.
porations to move more and more towards global The overall objective is to minimize the cost of the
supply chains, and therefore to become exposed to first-stage strategic decisions and the expected
risky factors such as exchange rates, reliability of production and distribution costs over the uncer-
transportation channels, and transfer prices [33]. tain demand scenarios. The authors used Benders
Unless the supply chain is designed to be robust decomposition to solve the resulting stochastic
with respect to the uncertain operating conditions, integer program, and presented computational
the impact of operational inefficiencies such as results on supply chain networks involving up to 8
delays and disruptions will be larger than neces- plant sites, 15 distribution centers, 30 customer
sary. A recent study [12] found that after a com- locations, and with 100 scenarios. Tsiakis et al. [29]
pany announces a supply chain disruption, such as also considered a two-stage stochastic program-
a production or shipment delay, its stock price can ming model for supply chain network design under
decrease significantly, with an average decrease of demand uncertainty. The authors developed a
8.6% on the day of the announcement, and is often large-scale mixed-integer linear programming
followed by further decreases, as much as 20% model for this problem, and presented a case study
over the next six months. using a European supply chain network involving
The significance of uncertainty has prompted a 14 products, 18 customer locations, 6 distribution
number of researchers to address stochastic center locations, and 3 demand scenarios. Alonso-
parameters in tactical level supply chain planning Ayuso et al. [2] proposed a branch-and-fix heu-
involving distribution of raw materials and prod- ristic for solving two-stage stochastic supply chain
ucts (see for example [6,17,34]). At the strategic design problems. Computational results on net-
level, there is a great deal of research in the facility works involving 6 plants, 12 products, 24 markets,
location component of supply chain network de- and 23 scenarios were presented.
sign under uncertainty (see for example [25]). As evident from the above discussion, the
However, research addressing comprehensive de- existing stochastic programming approaches for
sign of supply chain networks under uncertainty is supply chain design under uncertainty are suited
significantly smaller in number. Gutierrez et al. for a very small number of scenarios. Consider a
[10] proposed a robust optimization framework distribution network with just 50 facilities, each
for network design under uncertainty. This ap- with an uncertain operating level. Assume that the
proach seeks network configurations that are operating level for a facility can be one of only
good (nearly optimal) for a variety of scenarios three possibilities and are independent across
of the design parameters at the expense of being facilities. Then there are a total of 350
7 · 1023
sub-optimal for any one scenario. The authors scenarios for the joint realization of the uncer-
proposed a modification of the Benders decom- tainties! This is far more than any of the existing
position algorithm [3], commonly used for deter- stochastic programming approaches for supply
98 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115

chain design can handle. In this paper, we The supply chain configuration decisions con-
integrate a recently proposed sampling strategy, sist of deciding which of the processing centers to
the sample average approximation (SAA) scheme, build (major configuration decisions) and which
with an accelerated Benders decomposition algo- processing and finishing machines to procure
rithm to solve supply chain design problems with (minor configuration decisions). We associate a
continuous distributions for the uncertain param- binary variable yi to these decisions, yi ¼ 1, if a
eters, and hence an infinite number of scenarios. processing facility i is built or machine i is pro-
The proposed algorithmic technique is imme- cured, and 0 otherwise. The operational decisions
diately applicable to problems with finite but an consist of routing the flow of product k 2 K from
extremely large number of scenarios. Our compu- the supplier to the customers. We let xkij denote the
tational results involving two real supply chain flow of product k from a node i to a node j of the
networks indicate that the proposed methodology network where ðijÞ 2 A. We are now ready to
can serve as a viable strategic planning tool, state a deterministic mathematical model for our
allowing planners to investigate a multitude of supply chain design problem.
possibilities, and investigate a variety of solutions. X X X
min ci yi þ qkij xkij ð2:1Þ
The remainder of this paper is organized as
i2P k2K ðijÞ2A
follows: In the next section we describe a two-stage
model for supply chain design under uncertainty s:t: y 2 Y  f0; 1gjPj ; ð2:2Þ
and identify some of the challenges involved in X X
xkij  xkjl ¼ 0 8j 2 P; 8k 2 K;
developing a solution strategy. In Section 3, we i2N l2N
develop the proposed solution methodology. We
ð2:3Þ
explain the SAA technique, its integration with X
Benders decomposition, and develop techniques xkij P djk 8j 2 C; 8k 2 K; ð2:4Þ
for accelerating Benders decomposition. In Section i2N
X
4, we present a computational study involving two xkij 6 ski 8i 2 S; 8k 2 K; ð2:5Þ
real supply chain networks to illustrate the pro- j2N
posed methodology. Finally, we offer some con- !
X X
cluding remarks in Section 5. rjk xkij 6 mj yj 8j 2 P; ð2:6Þ
k2K i2N

x 2 RjAjjKj
þ : ð2:7Þ
2. Problem description
In the above model ci denotes the investment cost
Let us first describe a deterministic mathemat- for building facility i or procuring machine i, and
ical formulation for the supply chain design qkij denotes the per-unit cost of processing product
problem considered in this paper. Consider a k at facility i and/or transporting product k on arc
supply chain network G ¼ ðN; AÞ, where N is ðijÞ. All cost components are assumed to be on an
the set of nodes and A is the set of arcs. The set N annualized basis. The objective function (2.1)
consists of the set of suppliers S, the set of pro- consists of minimizing total investment and oper-
cessing facilities P and the set of customers C, i.e., ational costs. Constraint (2.2) includes the set Y of
N ¼ S [ P [ C. The processing facilities include logical dependencies and restrictions, such as
manufacturing centers M, finishing facilities F yi 6 yj for all i 2 Nj and j 2 P or F, i.e., machine
and warehouses W, i.e., P ¼ M [ F [ W . Fur- i 2 Nj should only be procured if facility j is built.
ther, a manufacturing center i 2 M or a finishing This constraint also enforces the binary nature of
facility i 2 F consists of a set of manufacturing or the configuration decisions for the processing
finishing machines Ni . Thus the set P includes the facilities. Constraint (2.3) enforces the flow con-
processing centers as well as the machines in these servation of product k across each processing node
centers. Let K be the set of products flowing j. Constraint (2.4) requires that the total flow of
through the supply chain. product k to a customer node j, should exceed the
 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115 99

demand djk at that node. Constraint (2.5) requires certain realizations of the uncertain parameters,
that the total flow of product k from a supplier we include an additional cost term to penalize
node j, should be less than the supply skj at that shortfall. The resulting formulation is as follows:
node. Constraint (2.6) enforces capacity con-
min f ðyÞ :¼ cT y þ E½Qðy; nÞ ð2:15Þ
straints of the processing nodes. Here rjk denotes y
per-unit processing requirement for product k at s:t: y 2 Y  f0; 1g ;
jPj
ð2:16Þ
node j. The capacity constraint then requires that
the total processing requirement of all products where Qðy; nÞ is the optimal value of the following
flowing into a processing node j should be smaller problem:
than the capacity mj of facility j if it is built min qT x þ hT z ð2:17Þ
(yj ¼ 1). If facility j is not built (yj ¼ 0) the con- x;z

straint will force all flow variables xkij ¼ 0 for all s:t: Nx ¼ 0; ð2:18Þ
i 2 N. Finally, constraint (2.7) enforces the non-
Dx þ z P d; ð2:19Þ
negativity of the flow variables corresponding to
an arc ðijÞ 2 A and product k 2 K. Sx 6 s; ð2:20Þ
It will be convenient to work with the following Rx 6 My; ð2:21Þ
compact notation for model (2.1)–(2.7):
x2 RjAjjKj
þ : ð2:22Þ
min c T y þ qT x ð2:8Þ
Note again that n in (2.15) is a random vector
jPj
s:t: y 2 Y  f0; 1g ; ð2:9Þ corresponding to the uncertain processing and/or
Nx ¼ 0; ð2:10Þ transportation costs, demands, supplies, and
capacities, and the optimal value Qðy; nÞ of the
Dx P d; ð2:11Þ
second stage problem (2.17)–(2.22) is a function of
Sx 6 s; ð2:12Þ the first stage decision variable y and a realization
Rx 6 My; ð2:13Þ (or a scenario) n ¼ ðq; d; s; MÞ of the uncertain
parameters. The expectation in (2.15) is taken with
x 2 RjAjjKj
þ : ð2:14Þ
respect to the probability distribution of n which is
Above vectors c, q, d, and s correspond to invest- supposed to be known. The variable z in constraint
ment costs, processing/transportation costs, de- (2.19) and the cost component hT z in (2.17) cor-
mands, and supplies, respectively. The matrices N , responds to the penalty incurred for failing to meet
D and S are appropriate matrices corresponding to demand. This penalty can be interpreted as the
the summations on the left-hand-side of the cost of outsourcing unmet demand. Although, in
expressions (2.3)–(2.5), respectively. The notation this paper, we assume that the outsourcing costs
R corresponds to a matrix of rjk , and the notation M are deterministic, uncertainty in these costs can be
corresponds to a matrix with mj along the diagonal. easily accommodated without any added concep-
To extend the above model to a stochastic set- tual difficulties.
ting, we assume that processing/transportation Model (2.15)–(2.22) is a two-stage stochastic
costs, demands, supplies, and capacities are sto- program. The first-stage consists of the deciding
chastic parameters with known joint distribution. the configuration decisions y, and the second-stage
We use bold face to denote random variables consists of processing and transporting products
(random vectors) in order to distinguish them from suppliers to customers in an optimal fashion
from their particular realizations. In particular, based upon the configuration and the realized
n ¼ ðq; d; s; MÞ represents the random data vector uncertain scenario. The objective is to minimize
while n ¼ ðq; d; s; MÞ stands for its particular current investment costs cT y and expected future
realization. The design objective is to minimize the operating costs E½Qðy; nÞ. The demand-shortage
sum of current investment costs and expected fu- penalty hT z guarantees that Qðy; nÞ < 1 for all y
ture processing and transportation costs. Given and n. Furthermore, we assume that possible
that it may be impossible to meet demand for realizations of the processing and transportation
100 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115

costs q, and the penalty costs h are sufficiently high In the SAA scheme, a random sample n1 ; . . . ; nN
such that Qðy; nÞ > 1 for all y and n, and hence of N realizations (scenarios) of the random vector
Qðy; nÞ is finite valued for all y 2 Y and possible n is generated, and the expectation E½Qðy; nÞ
realizations of the random data. We assume fur- is approximated
PN by the sample average func-
ther that the expected value E½Qðy; nÞ is well de- tion N 1 n¼1 Qðy; nn Þ. Consequently, the ‘‘true’’
fined and finite valued for the considered problem (2.15)–(2.16) is approximated by the
distribution of n. Consequently problem (2.15)– problem
(2.16) has a well-defined objective function f ðyÞ ( )
1 XN
and, since the set Y is non-empty and finite, pos- min fbN ðyÞ :¼ c y þ
T n
Qðy; n Þ : ð3:1Þ
sesses an optimal solution. y2Y N n¼1
There are two potential sources of difficulty in
Let vN and b y N be the optimal value and an optimal
solving problem (2.15)–(2.16): First, an evaluation
solution vector, respectively, of the SAA problem
of the objective function f ðyÞ for a given config-
(3.1). Note that vN and b y N are random in the sense
uration y, involves computing the expected value
that they are functions of the corresponding ran-
of the linear programming value function Qðy; nÞ.
dom sample. However, for a particular realization
For continuous distributions, exact computation
n1 ; . . . ; nN of the random sample, the problem (3.1)
of this expectation involves taking multiple inte-
is deterministic and can be solved by appropriate
grals and is quite impossible. For discrete distri-
optimization techniques.
butions, computing the expectation might involve
It is possible to show that under mild regularity
solving a prohibitively huge number of linear
conditions, as the sample size N increases, vN and
programs (2.17)–(2.22), one for each scenario of
b
y N converge with probability one to their true
the uncertain problem parameter realizations.
counterparts, and moreover b y N converges to an
Second, even if the expectation in (2.15) can be
optimal solution of the true problem with proba-
computed exactly, optimization of this function
bility approaching one exponentially fast [16]. This
presents significant difficulties. It is well known
convergence analysis suggests that a fairly good
that E½Qðy; nÞ is a convex non-linear (often non-
approximate solution to the true problem (2.15)–
smooth) function of y (see, e.g., [4]). However this
(2.16) can be obtained by solving an SAA problem
function is not available in a closed analytical form
(3.1) with a modest sample size. In particular,
and is only implicitly defined. Thus problem
suppose that the SAA problem is solved to an
(2.15)–(2.16) involves minimizing an implicitly
absolute optimality gap of d P 0 (i.e., b y N is a d-
defined non-linear function with respect to binary
optimal solution to (3.1)) and let e > d and
variables, and can be quite difficult.
a 2 ð0; 1Þ. Then a sample size of


3r2max jY j
3. Algorithmic strategy NP log ð3:2Þ
ðe  dÞ2 a

In this section, we detail an algorithmic strategy guarantees that the SAA solution b y N is an e-opti-
for solving the stochastic supply chain network mal solution (i.e., a solution with an absolute
design problem (2.15)–(2.16). Our method inte- optimality gap of e) to the true problem with a
grates a sampling strategy with an accelerated probability of at least 1  a. Here r2max is a maxi-
Benders decomposition scheme. mal variance of certain function differences (see
[16] for details of the estimate (3.2)). Note that
3.1. Sample average approximation jY j 6 2jPj , and hence log jY j 6 ðlog 2ÞjPj. Even
though the above estimate (3.2) is too conservative
As mentioned earlier, a key difficulty in solving for practical applications, it suggests that the re-
the stochastic program (2.15)–(2.16) is in evaluat- quired sample size is at most linear in the number
ing the expectation in the objective. We deal with of processing facilities jPj. In practice, the SAA
this problem using the SAA scheme (cf. [16,21,28]). scheme involves repeated solutions of the SAA
 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115 101

problem (3.1) with independent samples. Statisti- Note that f~N 0 ð
y Þ is an unbiased estimator
cal confidence intervals are then derived on the of f ð
y Þ. Since
y is a feasible solution to the
quality of the approximate solutions. This proce- true problem, we have f ð
y Þ P v . Thus
dure is described next. f~N 0 ð
y Þ is an estimate of an upper bound on
0
v . If the sample n1 ; . . . ; nN is independent
The SAA algorithm
identically distributed (iid), then the vari-
ance of this estimate can be estimated as
Step 1. Generate M independent samples each of
size N , i.e., ðn1j ; . . . ; nNj Þ for j ¼ 1; . . . ; M. r2N 0 ð
y Þ
N0 
X 2
For each sample solve the corresponding 1
:¼ 0 cT
y þ Qð
y ; nn Þ  f~N 0 ð
y Þ :
SAA problem 0
ðN  1ÞN n¼1
( ) ð3:5Þ
T 1 X N
n
min c y þ Qðy; nj Þ :
y2Y N n¼1 Step 4. Compute an estimate of the optimality gap
of the solution
y using the lower bound
Let vjN and b
y Nj , j ¼ 1; . . . ; M, be the cor- estimate and the objective function value
responding optimal objective value and an estimate from Steps 2 and 3, respectively,
optimal solution, respectively. as follows:
Step 2. Compute
gapN ;M;N 0 ð
y Þ :¼ f~N 0 ð
y Þ 
vN ;M : ð3:6Þ
1 XM

vN ;M :¼ vj and The estimated variance of the above gap
M j¼1 N
estimator is then given by
1 XM
r
2vN;M :¼ ðvj 
vN ;M Þ2 : ð3:3Þ r2gap ¼ r2N 0 ð
y Þ þ r
2vN;M : ð3:7Þ
ðM  1ÞM j¼1 N
It is natural to choose
y as one of the calculated
It is well known that the expected value of
estimates b y Nj with the smallest estimated objective
vN is less than or equal to the optimal
value f~N 0 ð
y Þ. Let us emphasize again that in order
value v of the true problem (see, e.g.,
to estimate the corresponding true objective func-
[21,23]). Since
vN ;M is an unbiased estima-
tion value f ð
y Þ one needs to generate a sample
tor of E½vN , we obtain that E½
vN ;M  6 v .
independent of the samples used in calculation of
Thus
vN ;M provides a lower statistical
y Nj . The above SAA procedure involving statistical
b
bound for the optimal value v of the true
evaluation of candidate solutions was suggested in
problem, and r
2vN ;M is an estimate of the
Norkin et al. [23,24] and further developed in Mak
variance of this estimator.
et al. [21]. Theoretical analysis of the SAA algo-
Step 3. Choose a feasible solution
y 2 Y of the true
rithm for solving stochastic programs with discrete
problem, for example, use one of the com-
first-stage variables was carried out in Kleywegt et
puted solutions b y Nj . Estimate the true objec-
al. [16]. Computational studies using the SAA
tive function value f ð
y Þ as follows:
method for solving stochastic linear programs is
N0 presented in Linderoth et al. [18], and for solving
1 X
f~N 0 ð
y Þ :¼ cT
y þ 0 Qð
y ; nn Þ: ð3:4Þ stochastic routing problems in Verweij et al. [31].
N n¼1
0
Here n1 ; . . . ; nN is a sample of size N 0 3.2. Accelerated benders decomposition
generated independently of the sample
used to obtain
y . Typically one can take N 0 Step 1 of the SAA method outlined in the pre-
much bigger than the sample size N used vious section calls for repeated solutions of the
in solving the SAA problems. This step sample average approximating problem (3.1). This
involves solution of N 0 independent sec- problem is itself a two-stage stochastic program,
ond-stage sub-problems (2.17)–(2.22). albeit with a much fewer number of scenarios than
102 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115

the true problem (2.15)–(2.16). Recall that the

1 XN
function Qðy; nn Þ in (3.1) is the value function of a fbN ðy i Þ ¼ cT y i þ Qðy i ; nn Þ
linear program (2.17)–(2.22) and is piece-wise lin- N n¼1
ear and convex in y. Consequently the SAA corresponding to the current feasible
problem (3.1) involves minimizing the sum of a solution y i . If fbN ðy i Þ < ub, update the
fairly large number, N , of piece-wise linear convex upper bound ub ¼ fbN ðy i Þ and the incum-
functions (plus the first stage linear cost function bent solution b y ¼ yi.
cT y) over 0–1 variables. Moreover since the func- Step 3: If ub  lb < d, where d P 0 is pre-specified
tions Qðy; nn Þ are not available in closed form, the tolerance, stop and return b y as the optimal
problem is quite difficult. Cutting plane algorithms solution and ub as the optimal objective
such as Benders decomposition [3] (also known as value; otherwise proceed to Step 4.
the L-shaped decomposition method in the sto- Step 4: For each n ¼ 1; . . . ; N , let ðlni ; mni ; pni ; qni Þ be
chastic programming literature [30]) is quite suit- the optimal dual solutions for the sub-
able for this class of problems. We briefly state the problem corresponding to y i and nn solved
general Benders decomposition algorithm as it in Step 2. Compute the cut coefficients
applies to problem (3.1).
Benders decomposition algorithm 1 XN
T
aTiþ1 ¼ ðqn Þ M n
Initialization step. Set lower and upper bounds N n¼1 i
lb ¼ 1 and ub ¼ þ1, respectively. Set the
iteration counter i ¼ 0. Let b y denote the incum- and
bent solution. N  
1 X
biþ1 ¼ ðmni ÞT d n þ ðpni ÞT sn :
Step 1: Solve the master problem N n¼1

lb ¼ min cT y þ h Update i ¼ i þ 1 and go to Step 1.

y;h

s:t: y 2 Y; In Benders decomposition, the master problem

h P aTk y þ bk ; k ¼ 1; . . . ; i: solved in Step 1 involves a lower
PN approximation of
i
the non-linear function N 1 n¼1 Qðy; nn Þ through
Let y be an optimal solution of the master the h variables and the cuts h P aTk y þ bk . This
problem. problem is a mixed-integer linear program
Step 2: For n ¼ 1; . . . ; N , solve the sub-problems involving jPj binary variables and one continuous
(2.17)–(2.22) corresponding to y i and variable. The optimal objective value and the
nn ¼ ðqn ; d n ; sn ; M n Þ. We re-state this prob- optimal solution obtained from the master prob-
lem for convenience: lem correspond to a lower bound to the optimal
Qðy i ; nn Þ ¼ min ðqn ÞT x þ hT z objective value and a feasible solution for the SAA
s:t: Nx ¼ 0 ðlÞ; problem (3.1), respectively. The objective value
corresponding to this feasible solution is evaluated
Dx þ z P d n ðmÞ;
by solving the sub-problems in Step 2. These N
Sx 6 sn ðpÞ; linear programs can be solved independent of each
Rx 6 ðM n Þy i ðqÞ; other, allowing for a computationally convenient
x2 RjAjjKj ; decomposition. Since the current solution is fea-
þ
sible, this yields an upper bound information. If
where the Greek terms in parentheses next the lower and upper bounds are sufficiently close
to the constraints denote the correspond- together we conclude optimality. Otherwise, dual
ing dual variables. Using the sub-problem information from the sub-problems is used to
objective values, compute the objective compute an optimality cut which is added to the
function value master problem. The new optimality cut serves to
 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115 103

PN
improve the approximation of N 1 n¼1 Qðy; nn Þ in the feasible master problem solutions are all ex-
the master problem, which is then resolved and the treme points of the ½0; 1jPj hypercube, a trust re-
process iterates. gion with size greater than or equal to one would
By duality it follows that the optimality cut include all binary solutions, and a trust region with
added at the end of the ith iteration is ‘‘exact’’ at size less than one would include only the previous
y i , i.e., feasible solution. Consequently, such a scheme is
X
N not meaningful. In our implementation, we use a
aTiþ1 y i þ biþ1 ¼ N 1 Qðy i ; nn Þ; trust region that bounds the Hamming distance
n¼1 [11] of the next master problem solution from the
previous solution. Suppose y i is the master prob-
and is a lower approximate at all other solutions.
lem solution obtained in the ith iteration and let
Using this fact together with the finiteness of the
Y i :¼ fj : yji ¼ 1g. We impose the following trust
solution set it is easy to show that the algorithm
region constraint in the master problem of itera-
terminates in a finite number of iterations with an
tion i þ 1:
d-optimal solution to the SAA problem (3.1). X X
While the Benders decomposition algorithm is a ð1  yj Þ þ yj 6 Di ;
finite scheme, the number of iterations required j2Y i j62Y i
may be too large in practice. To improve the
where Diþ1 < jPj is the trust region size at iteration
convergence behavior of the generic Benders
i þ 1. The trust region size can be constant or
decomposition algorithm outlined above, we use a
dependent on the iteration. Unfortunately, unlike
number of acceleration techniques. These strate-
in case of continuous first-stage variables, con-
gies are described next.
vergence of the algorithm cannot be ensured if a
Trust region non-redundant trust region is maintained
An undesirable feature of cutting plane algo- throughout the algorithm. Since the initial itera-
rithms such as Benders decomposition is that in tions exhibit poor convergence behavior, we im-
the early iterations, the solutions tend to oscillate pose the trust region in the initial iterations, and
wildly from one region of the feasible set to an- drop it once the iterates have stabilized.
other, thereby slowing convergence [13]. For con-
Knapsack inequalities
tinuous problems, this drawback is effectively
Let h P aTi y þ bi be the optimality cut derived at
controlled either by adding a regularizing term to
the end of the ith iteration, and let ub be the current
the objective of the master problem that penalizes
best known upper bound. Since ub P cT y þ h, we
the ‘2 -distance of the master problem solution
have the following valid inequality for the master
from the previous solution [13,26]; or by adding
problem in iteration i þ 1:
constraints that bound the ‘1 distance of master
problem variables from the previous solution to bðc þ ai ÞT cy 6 bub  bi c;
within a trust region [19]. These extensions prevent
where bac is the component-wise floor of a. If a
the master problem solution from moving too far
good upper bound ub is available, then adding the
from the previous iterate. The cutting plane algo-
above knapsack inequality along with the opti-
rithm is modified to allow for increasing or
mality cut can have a significant impact in gener-
decreasing the regularizing penalty or trust region
ating a good quality solution from the master
size based upon the progress. By proper control of
problem in iteration i þ 1. A state-of-the-art solver
these, convergence of the modified algorithm can
such as CPLEX can derive a variety of valid
be ensured [19,26].
inequalities from the above knapsack inequality,
In our supply chain network design problem,
and expedite convergence.
the first-stage variables are binary. In this case,
adding a ‘2 regularizing term would lead to a Upper bounding heuristic
mixed-integer quadratic master problem. On the The upper bound and incumbent solution
other hand, if a ‘1 trust region is used, then since identified in Step 2 of the Benders decomposition
104 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115

algorithm corresponds to the solution y i of the these alternative cuts are valid, and exact at the
master problem. This solution may be quite far current solution y i , one cut may be dominated by
from the optimal solution, and may cause the another in the vicinity of the optimal solution. For
algorithm to explore inferior parts of the feasible example, let ða; bÞ and ða; bÞ be two alternative
region. If a good solution is available through optimality cut coefficients at y i corresponding to
some heuristic, we can replace y i by this solution in alternative sets of dual optimal solutions, then
Step 2 and proceed from there. Our implementa-
aT y i þ b ¼ aT y i þ b:
tion makes use of a particular heuristic strategy for
this purpose which is described next. Suppose y  is the optimal solution, then cut ða; bÞ
In our implementation of the generic Benders dominates cut ða; bÞ if
decomposition algorithm, we observed that after
aT y  þ b > aT y  þ b:
the optimality gap becomes quite small, the upper
bound exhibits a tailing off behavior as the itera- Clearly cut ða; bÞ is preferable in iteration i since it
tions progresses. This is due to the fact that, in will typically lead to better lower bounds and
these iterations, the incumbent solutions all have expedite convergence. Unfortunately, since the
identical or very similar major configuration optimal solution y  is not known a priori it is
decisions (corresponding to the processing centers) difficult to identify dominating cuts. Magnanti and
solutions, and the only changes are due to the Wong [20] considered a core point, i.e., a point in
minor configuration decisions (corresponding to the relative interior of the convex hull of feasible
the machines). Since the minor configuration region, as a proxy for the optimal solution y  .
decisions have relatively small implication on the They proved that if a cut is selected such that it
objective, the upper bound changes very little. To attains the maximum value at a core point
avoid this behavior, we call the following heuristic amongst the set of all alternative cuts, then such a
strategy in Step 2 if the upper bound has not im- cut is not dominated by other cuts at any feasible
proved in several successive iterations: solution. Such a cut is termed Pareto-optimal. We
follow this strategy to identify good dual solutions
Step 1: Fix all the major configuration decisions and hence Pareto-optimal cuts. Typically, a core
to those in the incumbent solution. point is not easy to identify. In our implementa-
Step 2: Consider a subset of the sampled scenar- tion, we choose a fractional optimal solution from
ios, and solve the corresponding determin- the LP relaxation of the master problem y 0 as a
istic equivalent problem to solve for the candidate core point. Although, such a solution is
minor configuration decisions. not guaranteed to be a core point, it is often in the
Step 3: Evaluate the objective function corre- neighborhood of the integer optima. Furthermore,
sponding to the solution (major and min- after some cuts are added to the master problem,
or configuration) obtained above by the LP relaxation solution will typically be in the
solving all the sub-problems as in Step 2. interior of the convex hull of Y , and hence satisfy
If the solution is better than y i (the current the requirement of being a core point. After solv-
master problem solution), replace y i by ing the sub-problem (2.17)–(2.22) to compute
this solution. Update upper bound and Qðy i ; nn Þ in Step 2, we solve the following linear
incumbent solution if necessary and pro- program to identify good dual solutions:
ceed as before.
max ðqT M n Þy 0 þ mT d n þ pT sn
l;m;p;q
Cut strengthening s:t: N T l þ DT m þ S T p þ RT q 6 qn ;
Note that the sub-problem (2.17)–(2.22) has a
network sub-structure. Typically such problems m 6 h;
have multiple dual optimal solutions. Conse- l unrestricted; m P 0; p 6 0; q 6 0;
quently, there may be a number of alternatives for
the optimality cut computed in Step 3. While all of ðqT M n Þy i þ mT d n þ pT sn ¼ Qðy i ; nn Þ:
 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115 105

The first three constraints of the above linear !

X X
program enforces the dual feasible region of the rik xn;k
il 6 mni yi :
sub-problem (2.17)–(2.22) corresponding to y i and k2K l2N
nn , and the fourth constraint restricts the feasible
dual solutions to the set of alterative dual optima. Since rik > 0, and j 2 N, the above implies, for a
mni
The objective function corresponds to maximiz- given k 2 K, xn;k ij 6 ri i
k y : Summing over all
ing the cut value at y 0 . Note that, the above i 2 PðjÞ, we have
scheme for identifying alternative dual solutions X n;k X mn
i
does not prevent the Benders decomposition xij 6 k i
y: ð3:9Þ
i2PðjÞ
r
i2PðjÞ i
algorithm to converge even when y 0 is not a core
point. Combining inequalities (3.8) and (3.9), we have the
following valid inequality:
Logistics constraints
In the early iterations, the master problem X mn
i
k i
y þ zn;k n;k
j P dj 8j 2 C; k 2 K;
solved in Step 1 of the Benders decomposition i2PðjÞ
r i
algorithm includes only the logical constraints Y
and a few cuts. Consequently, the configurations n ¼ 1; . . . ; N :
produced consist of very few open facilities and Taking the probability weighted sum of the above
therefore correspond to a small objective function inequalities, and substituting a new variables
z as
value (lower bound). On the other hand, for such the average of all the zn variables, we have
sparse configurations, little demand can be met X m
i
in the scenario sub-problems, thereby resulting y þ
zkj P d
jk 8j 2 C; k 2 K;
k i
ð3:10Þ
r
i2PðjÞ i
high shortage penalties. Therefore the resulting
upper bound is very high. To avoid such bad where d
and m
i are the average demand of product
master problem solutions, it is important to add k for customer j, and capacity of facility i,
some sub-problem information into the master. respectively. We can add the above deterministic
However adding additional constraints and vari- constraint to the master problem, along with the
ables into the master may impede solution effi- new variable
zkj which captures the shortfall from
ciency. Therefore, a trade-off between the master meeting the average demand. The penalty of
problem size and the quality of the solution gen- shortage also needs to be added to the master
erated needs to reached. Next, we derive a sim- problem, P i.e., the
ple set of constraints that we found to be quite P objective
k k
of the master should
include k2K zj . However, since the
j2C hj

useful in improving the master problem solu- shortage penalties are quite high, and we would
tion. want a configuration that is capable of satisfying
Consider the sub-problem corresponding to nn , at least the average demands, we drop the penalty
and a customer node j 2 C. Expanding the de- variables
zkj , and modify the original master
mand constraint (2.19) for this customer node, we problem, by adding the inequalities
have X m
i
X n;k y P d
jk 8j 2 C; k 2 K:
k i
ð3:11Þ
xij þ zn;k n;k
j P dj ; ð3:8Þ i2PðjÞ
r i
i2PðjÞ
Cut disaggregation
where PðjÞ is the set of processing centers directly Recall that in the proposed Benders decompo-
supplying to customer node j. Note that we have sition algorithm, one optimality cut is added in
indexed the x and z variables also by n 2 each iteration. The cut approximates the sample
f1; . . . ; N g since these variables are local to the average of the second-stage value function at the
current sub-problem only. Now consider a pro- current solution. Instead one could add N cuts to
cessing node i 2 PðjÞ. Expanding the capacity approximate the individual second-stage value
constraint (2.21), we have functions corresponding to each of the N sampled
106 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115

scenarios. In this case the master problem solved 4.1. Data and implementation
in Step 1 of the ith iteration takes the form
Our first test problem is that of the design of a
1 XN
domestic supply chain network for a US company
min cT y þ hn
y;h N n¼1 that supplies cardboard packages to breweries and
s:t: y 2 Y; soft-drink manufacturers. The problem is adapted
T from [7]. Henceforth this problem is referred to as
hn P ðank Þ y þ bnk ; k ¼ 1; . . . ; i; the ‘‘domestic’’ problem.
n ¼ 1; . . . ; N : Our second test problem is a global supply
The cut coefficients ani and bni are computed as chain network design problem described in [33].
The chain encompasses the US and seven Latin
T T
ðani Þ ¼ ðqni1 Þ M n American countries. This problem is referred to as
the ‘‘global’’ problem in the remainder of this
and
paper. The global problem is slightly different from
T T
bni ¼ ðmni1 Þ d n þ ðpni1 Þ sn ; the cost-minimization formulation (2.15)–(2.22).
This problem involves maximizing the expected net
where ðmni1 ; pni1 ; qni1 Þ are the optimal dual solu-
cash flow (NCF) after subtracting the first-stage
tions for the sub-problem corresponding to y i1
design costs from the expected after-tax profits
and nn solved in Step 2 of iteration ði  1Þ. The
(revenues minus the processing and transportation
disaggregation of the optimality cut can provide a
costs) of the second-stage operational decisions.
better approximation of the sample average of the
The second-stage problem involves maximizing
second-stage value functions, and thereby improve
after tax profits. The model formulation is
convergence, at the expense of a larger master
problem. The trade-off lies in between the fewer max f ðyÞ :¼ cT y þ E½Qðy; nÞ
y
number of outer iterations of Benders decompo-
sition and the more computational time required s:t: y 2 Y  f0; 1gjPj ;
to solve the master problem in each iteration. where Qðy; nÞ is the optimal value of the following
For problems where N is not too large, this ap- problem:
proach may be viable. This variant of Benders
decomposition algorithm is often referred to as the max qT x
x
multi-cut version [5]. A similar cut disaggregation
approach has been shown to be very effective in s:t: Nx ¼ 0;
Benders decomposition methods for solving Dx 6 d;
deterministic supply chain design problems in [7]. Sx 6 s;
Rx 6 My;
x 2 RjAjjKj
þ :
4. Computational results
Here the parameters q denote per unit after-tax
In this section we describe numerical experi- profits. These parameters are computed after tak-
ments using the proposed methodology for solving ing into account the import and export taxes and
two realistic supply chain design problems. We currency exchange rates across different countries
first describe the characteristics of the test prob- of the supply chain network. Note that it is not
lems and some implementation details, then dem- required that all demand be satisfied, and hence no
onstrate the computational efficiencies afforded by shortage penalties are imposed. It should be clear
our method, and finally comment on the quality of that conceptually the above model is no different
the stochastic programming solution in compari- from the cost minimization formulation (2.15)–(2.22).
son to those obtained using a deterministic ap- The supply chain topology with all possible
proach. center locations and transportation channels for
 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115 107

the domestic and global problem are presented in Recall from Section 3.1 that the SAA method
Figs. 1 and 2, respectively. The main characteris- calls for the solution of M instances of the
tics of these two networks are presented in Table 1. approximating stochastic program (3.1), each
We used the same data for the deterministic involving N sampled scenarios. Statistical valida-
parameter values as in [7] (for the domestic prob- tion of a candidate solution is then carried out by
lem) and [33] (for the global problem). To obtain evaluating the objective function using N 0 sampled
stochastic instances of the problems, we assumed scenarios. In our implementation, we used N ¼ 20,
that the product demands and facility capacities 30, 40 and 60; M ¼ 20; and N 0 ¼ 1000. As will
are uncertain. We used log-normal distributions be evident in our numerical results, these param-
for these uncertain parameters. The non-negativity eters were sufficient to obtain good quality solu-
of the parameter values are preserved by using tions with high confidence levels. To illustrate
log-normal distributions. Furthermore, recently the complexity of solving (3.1) within the SAA
Kamath and Pakkala [15] provided evidence that scheme, we present the sizes of the deterministic
log-normal distributions are well suited for mod- equivalents of the SAA problems corresponding to
elling economic stochastic variables such as de- the different values of N in Table 2.
mands. We used the values for the demand and The proposed algorithmic scheme was imple-
capacity parameters used in the deterministic mented in C++ with CPLEX 7.0 [14] callable li-
models of [7] and [33] as the mean values for the brary routines for solving linear and mixed-integer
random parameters. The standard deviations for the programming sub-problems. All computations
distributions were chosen as certain fractions of the were carried out on a Pentium II 400 MHz PC
mean values. running Windows NT.

Fig. 1. Supply chain network for the domestic problem.

108 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115

Fig. 2. Supply chain network for the global problem.

Table 1 tational efficiency of the solution procedure is of

Supply chain network characteristics great significance. Here we discuss the perfor-
Domestic Global mance of the accelerated Benders decomposition
Product types 13 29 algorithm for solving the SAA problems.
Total facilities 142 87 Figs. 3 and 4 chart the growth of the compu-
Internal suppliers 2 6
tational time with the growth of the sample size N .
Manufacturing plants 8 8
Machines 28 10 For each of the two problems, we compare the
Finishing facilities 9 10 CPU seconds required for solving an SAA prob-
Final process machines 93 36 lem instance as a monolithic deterministic equiv-
Warehouses 2 17 alent problem using the CPLEX MIP solver, the
Customers 238 17
standard Benders decomposition algorithm, and
Transportation channels 1559 239
Country 1 7 the accelerated Benders decomposition algorithm
proposed in Section 3.2. The efficacy of the pro-
posed acceleration strategies is clearly observed.
4.2. Performance of acceleration techniques Table 3 compares the performance of various
combinations of the acceleration schemes de-
Recall that, since the SAA method calls for scribed in Section 3.2 for solving SAA instances of
repeated solutions of the SAA problems, compu- 20 scenarios. The acceleration schemes are denoted
 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115 109

Table 2
Size of the deterministic equivalent of the SAA problem
N Domestic Global
Constraints Variables Constraints Variables
Equality Inequality Continuous Binary Equality Inequality Continuous Binary
1 3498 4324 20,912 140 1065 402 6824 70
20 69,960 86,480 418,240 140 21,300 8040 136,480 70
40 139,920 172,960 836,480 140 42,600 16,080 272,960 70
60 209,880 259,440 1,254,860 140 63,900 24,120 409,440 70

Fig. 3. CPU seconds versus sample size for the domestic problem.

as follows: logistics constraints (LC); trust region ‘‘All’’ corresponds to all five acceleration schemes
(TR); cut disaggregation (CD); Knapsack (LC, TR, KI, UH, and CS for the domestic
inequalities (KI); upper-bounding heuristic (UH); problem, and LC, CD, KI, UH and CS for the
and cut strengthening (CS). Each row of the table global problem) used together. For the domestic
presents computational performance measures for problem, we compare the relative optimality gap
various combinations of these acceleration after the first iteration (1st gap), the relative opti-
schemes. The trust region scheme was not effective mality gap after the 10th iteration (10th gap), the
for the global problem and the cut disaggregation total CPU seconds, and the total number of iter-
scheme was not tested for the domestic problem. ations. Similar performance measures are used for
Hence, the notation ‘‘TR/CD’’ in the rows of Ta- the global problem, except that we consider the
ble 3 refers to the trust region scheme for the relative optimality gap after the 50th iteration
domestic problem and the cut disaggregation (50th gap) instead of the 10th iteration (10th gap).
scheme for the global problem. The row marked From Table 3 it can be observed that, although
‘‘Standard’’ corresponds to the standard Benders the acceleration schemes improves the convergence
decomposition algorithm, and the row marked behavior over standard Benders decomposition,
110 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115

Fig. 4. CPU seconds versus sample size for the global problem.

Table 3
Comparison of acceleration schemes (N ¼ 20)
Acceleration scheme Domestic Global
1st Gap 10th Gap CPUs Iteration 1st Gap 50th Gap CPUs Iteration
(%) (%) (%) (%)
Standard >100 60 >4000 >30 >100 41 >13,000 >60
LC 31 8 >4000 >30 >100 27 >13,000 >60
TR/CD >100 40 >4000 >30 >100 18 >13,000 >60
KI 60 5 >4000 >30 >100 21 >13,000 >60
UH >100 60 >4000 >30 >100 41 >13,000 >60
CS >100 9 >4000 >30 >100 29 >13,000 >60
LC + TR/CD 31 0.7 >4000 >30 >100 22 11,900 56
LC + KI 31 0.1 3860 26 >100 12 >13,000 >60
LC + UH 31 0.08 2180 12 >100 27 >13,000 >60
LC + CS 31 0.5 >4000 >30 >100 19 12,300 58
TR/CD + KI 60 3 >4000 >30 >100 5 >13,000 >60
LC + TR/CD + KI 31 0.2 3600 23 >100 3 10,300 51
LC + KI + UH 31 0.01 1500 8 >100 12 >13,000 >60
LC + TR/CD + KI + UH 31 0.01 1380 7 >100 4 10,300 51
LC + TR/CD + KI + CS 31 0.06 3050 19 >100 <1 9800 45
All 31 0.01 1890 7 >100 <1 9800 45

no one acceleration technique is a clear winner. performance could be achieved by combining LC,
Rather, the schemes are most effective in concert. TR, KI, and UH. While, for the global problem,
In particular, for the domestic problem, the best the best combination is LC, CD, KI, and CS. Al-
 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115 111

though, the results in Table 3 correspond to SAA a sample size of just N ¼ 20 the SAA problem
instances of sample size 20, similar behavior has solutions are very close to being optimal for the
been observed for problems with different sample true stochastic supply chain design problem.
sizes (see [27] for detailed computational results). Table 5 compares the mean-value problem
solution to the SAA solution for the global prob-
4.3. Quality of stochastic solutions lem. Once again a sample size of N ¼ 20 was used.
In this case, the SAA method identified a single
In this section we compare the solutions of the ‘‘best’’ candidate solution (denoted by y 1 ). The
stochastic programming model to that of a deter- estimated optimality gaps clearly indicate the
ministic optimization problem involving the mean superiority of the SAA solution.
values of the uncertain problem parameters Tables 4 and 5 reveal that the solutions to the
(known as the mean-value problem). Recall that stochastic programming model are not only supe-
the SAA method produces a number of candidate rior to the mean-value problem solution in terms
solutions (at most M unique solutions). In Table 4, of the optimality gap corresponding to the ex-
we compare statistics of the (uncertain) total cost pected cost/NCF objective, but these solutions also
for the domestic problem corresponding to the lead to comparatively smaller variability of the
mean-value problem solution (denoted by y MVP ) to cost/NCF objective.
that of three candidate solutions (denoted by y 1 , In Tables 6 and 7, we present the optimality gap
y 2 , and y 3 , respectively) identified by solving estimates of the ‘‘best’’ solution identified by the
Mð¼ 20Þ SAA problem instances with N ¼ 20. SAA method for different sample sizes N . It is
The candidate solutions were chosen as the three clear that the SAA method with only a modest
‘‘best’’ solutions based upon their objective func- number of sampled scenarios can provide very
tion value and optimality gap estimates as pro- high quality solutions to the true stochastic supply
vided by the SAA method. As before, the total cost
statistics for each solution are computed using a Table 5
sample size of N 0 ¼ 1000. It is clear that the esti- NCF statistics for candidate solutions to the global problem
mated average total costs for all three candidate NCF (million $) y MVP y1
stochastic programming solutions are smaller than Average 51.03 54.002
that of the mean-value problem solution. Fur- SD 0.126 0.124
thermore, from rows 2–4 of Table 4, we observe Min 35.243 43.149
that the cost variability corresponding to the sto- Max 64.290 69.403
chastic programming solutions are also smaller gapN;M;N 0 3.18 0.21
rgap 0.54 0.05
than that of the mean-value problem solution. The
last two rows of Table 4 displays the estimated
optimality gap and the standard deviation of the Table 6
Optimality gap estimates for the domestic problem
gap estimate (computed using (3.6) and (3.7),
respectively) for the mean-value problem solution N gapN ;M;N 0 rgap
and each of the three best SAA solutions. Even for 20 0.16 0.04
30 0.11 0.03
40 0.08 0.02
Table 4 60 0.07 0.02
Costs statistics for candidate solutions to the domestic problem
Cost (million $) y MVP y1 y2 y3
Table 7
Average 116.77 111.03 111.03 111.05 Optimality gap estimates for the global problem
SD 0.34 0.11 0.11 0.11
Min 99.02 100.38 100.14 100.10 N gapN ;M;N 0 rgap
Max 173.30 122.57 122.08 122.11 5 1.24 0.22
gapN ;M;N 0 5.91 0.16 0.17 0.18 20 0.21 0.05
rgap 0.79 0.04 0.04 0.04 60 0.09 0.02
112 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115

Fig. 5. Effect of variability on the cost range for the domestic problem.

Fig. 6. Effect of variability on the NCF range for the global problem.

chain design problem involving potentially infinite Finally, to observe the effect of the variability of
number of scenarios. the uncertain problem parameters, we considered
 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115 113

Fig. 7. Effect of variability on the worst-case cost for the domestic problem.

Fig. 8. Effect of variability on the worst-case NCF for the global problem.

three different levels of the variability of the from solving SAA problems of a given sample size
uncertain problem parameters. For each level, we N . In Figs. 5 and 6, the ranges for the cost and
considered the best candidate solution obtained NCF corresponding to these candidate solutions
114 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115

are compared against those of the corresponding Institute of Paper Science and Technology (IPST),
mean-value problem solution. As before these and by the National Science Foundation under
statistics are computed using a sample size of grants DMI-0099726 and DMI-0133943.
N 0 ¼ 1000. As mentioned earlier, the stochastic
solutions have significantly smaller ranges than the
mean-value problem solution. Furthermore, with References
the increase in the variability of the uncertain
parameters, the cost/NCF variability (range) for [1] C.H. Aikens, Facility location models for distribution
planning, European Journal of Operational Research 22
the stochastic solutions increases at a slower rate
(3) (1985) 263–279.
than that of the mean-value problem solution. In [2] A. Alonso-Ayuso, L.F. Escudero, A. Garin, M.T. Ortuno,
Figs. 7 and 8, the worst-case cost and NCF cor- G. Perez, An approach for strategic supply chain planning
responding to the candidate stochastic solutions under uncertainty based on stochastic 0–1 programming,
are compared against those of the corresponding Journal of Global Optimization 26 (2003) 97–124.
mean-value problem solution. Once again, it is [3] J.F. Benders, Partitioning procedures for solving mixed
variables programming problems, Numersiche Mathema-
clear that the stochastic solution is more resilient tik 4 (1962) 238–252.
to the variability of the problem parameters. For [4] J.R. Birge, F. Louveaux, Introduction to Stochastic
the global problem, the resiliency is improved by Programming, Springer, New York, 1997.
including more scenarios in the SAA problem. [5] J.R. Birge, F.V. Louveaux, A multicut algorithm for two-
stage stochastic linear programs, European Journal of
Operational Research 34 (3) (1988) 384–392.
[6] R.K.-M. Cheung, W.B. Powell, Models and algorithms for
5. Concluding remarks distribution problems with uncertain demands, Transpor-
tation Science 30 (1996) 43–59.
In this paper, we have developed a practical [7] K. Dogan, M. Goetschalckx, A primal decomposition
method for the integrated design of multi-period produc-
methodology for large-scale supply chain network
tion–distribution systems, IIE Transactions 31 (1999)
design problems under uncertainty. The method 1027–1036.
integrates an accelerated decomposition scheme [8] A.M. Geoffrion, G.W. Graves, Multi-commodity distribu-
along with the recently developed SAA method. tion system design by Benders decomposition, Manage-
The proposed methodology provides an efficient ment Science 20 (1974) 822–844.
framework for identifying and statistically testing [9] A.M. Geoffrion, R.F. Powers, Twenty years of strategic
distribution system design: An evolutionary perspective,
a variety of candidate design solutions. We have Interfaces 25 (1995) 105–128.
provided empirical results for the design two [10] G.J. Gutierrez, P. Kouvelis, A.A. Kurawala, A robustness
realistic supply chain networks. Our results reveal approach to uncapacitated network design problems,
the computational efficacy of the proposed method. European Journal of Operations Research 94 (1996) 362–
Furthermore, we have demonstrated that the 376.
[11] R.W. Hamming, Error detecting and error correcting
candidate solutions identified by the proposed codes, Bell System Tech Journal 9 (1950) 147–160.
method are not only superior to traditional mean- [12] M. Hicks, When the chain snaps, eWeek––Enterprise News
value problem solutions in an expectation sense, & Reviews. Available from <http://www.eweek.com>,
but also result in significantly smaller cost/cash 2002.
flow variability, and this reduction is more pro- [13] J.-B. Hiriart-Urruty, C. Lemarechal, Convex Analysis and
Minimization Algorithms II, Springer-Verlag, Berlin, 1996.
nounced in case of higher variability in the [14] ILOG Inc, CPLEX 7.0 User’s Manual, ILOG CPLEX
uncertain environment. Division, Incline Village, NV, 2000.
[15] K.R. Kamath, T.P.M. Pakkala, A Bayesian approach to a
dynamic inventory model under an unknown demand
Acknowledgements distribution, Computers & Operations Research 29 (2002)
403–422.
[16] A.J. Kleywegt, A. Shapiro, T. Homem-De-Mello, The
This work has been funded in part by a grant sample average approximation method for stochastic
from the Sloan Foundation to the Center for Pa- discrete optimization, SIAM Journal of Optimization 12
per Business and Industry Studies (CPBIS) at the (2001) 479–502.
 T. Santoso et al. / European Journal of Operational Research 167 (2005) 96–115 115

[17] H. Van Landeghem, H. Vanmaele, Robust planning: A [27] T. Santoso, A comprehensive model and efficient solution
new paradigm for demand chain planning, Journal of algorithm for the design of global supply chains under
Operations Management 20 (2002) 769–783. uncertainty, Ph.D. Thesis, School of Industrial & Sys-
[18] J.T. Linderoth, A. Shapiro, S.J. Wright, The empirical tems Engineering, Georgia Institute of Technology,
behavior of sampling methods for stochastic programming, 2002.
Annals of Operations Research (forthcoming). [28] A. Shapiro, T. Homem de Mello, A simulation-based
[19] J.T. Linderoth, S.J. Wright, Implementing a decomposi- approach to stochastic programming with recourse, Math-
tion algorithm for stochastic programming on a computa- ematical Programming 81 (1998) 301–325.
tional grid, Computational Optimization and Applications [29] P. Tsiakis, N. Shah, C.C. Pantelides, Design of multi-
24 (2003) 207–250. echelon supply chain networks under demand uncertainty,
[20] T.L. Magnanti, R.T. Wong, Accelerating Benders decom- Industrial & Engineering Chemistry Research 40 (2001)
position: Algorithmic enhancement and model selection 3585–3604.
criteria, Operations Research 29 (1981) 464–484. [30] R. Van Slyke, R.J.-B. Wets, L-shaped linear programs with
[21] W.K. Mak, D.P. Morton, R.K. Wood, Monte Carlo applications to optimal control and stochastic program-
bounding techniques for determining solution quality in ming, SIAM Journal on Applied Mathematics 17 (1969)
stochastic programs, Operations Research Letters 24 638–663.
(1999) 47–56. [31] B. Verweij, S. Ahmed, A.J. Kleywegt, G. Nemhauser, A.
[22] S.A. MirHassani, C. Lucas, G. Mitra, E. Messina, C.A. Shapiro, The sample average approximation method
Poojari, Computational solution of capacity planning models applied to stochastic routing problems: A computational
under uncertainty, Parallel Computing 26 (2000) 511–538. study, Computational Optimization and Applications 24
[23] V.I. Norkin, G.Ch. Pflug, A. Ruszczy nski, A branch and (2003) 289–333.
bound method for stochastic global optimization, Math- [32] C. Vidal, M. Goetschalckx, Strategic production–distribu-
ematical Programming 83 (1998) 425–450. tion models: A critical review with emphasis on global
[24] V.I. Norkin, Y.M. Ermoliev, A. Ruszczy nski, On optimal supply chain models, European Journal of Operational
allocation of indivisibles under uncertainty, Operations Research 98 (1997) 1–18.
Research 46 (1998) 381–395. [33] C. Vidal, M. Goetschalckx, A global supply chain model
[25] S.H. Owen, M.S. Daskin, Strategic facility location: A with transfer pricing and transportation cost allocation,
review, European Journal of Operational Research 111 European Journal Of Operational Research 129 (2001)
(1998) 423–447. 134–158.
[26] A. Ruszczy nski, A regularized decomposition method for [34] C.-S. Yu, H.-L. Li, A robust optimization model for
minimizing a sum of polyhedral functions, Mathematical stochastic logistic problems, International Journal of
Programming 35 (1986) 309–333. Production Economics 64 (2000) 385–397.