You are on page 1of 12

Transportation Research Part C 19 (2011) 821832

Contents lists available at ScienceDirect

Transportation Research Part C


journal homepage: www.elsevier.com/locate/trc

A bi-objective supply chain design problem with uncertainty


Y. Cardona-Valds, A. lvarez, D. Ozdemir *
Universidad Autnoma de Nuevo Len, San Nicols de los Garza, N.L., Mexico

a r t i c l e

i n f o

Article history:
Received 1 December 2009
Received in revised form 9 April 2010
Accepted 20 April 2010

Keywords:
Supply chain design
Multi-objective stochastic program
Transportation channel
e-Optimality
L-shaped method

a b s t r a c t
We consider the design of a two-echelon production distribution network with multiple
manufacturing plants, customers and a set of candidate distribution centers. The main contribution of the study is to extend the existing literature by incorporating the demand
uncertainty of customers within the distribution center location and transportation mode
allocation decisions, as well as providing a network design satisfying the both economical
and service quality objectives of the decision maker within two levels supply network setting. We formulate the problem as two stage integer recourse problem to nd a set of optimal network conguration and assignment of transportation modes and the respective
ows in order to minimize total cost and total service time, simultaneously. We develop
a stochastic optimization model under demand uncertainty, where the inherent risk is
modeled by scenarios. Finally, we propose solution methods for our stochastic optimization problem based on L-shaped algorithm within an -optimality framework and present
numerical results demonstrating the computational effectiveness.
2010 Elsevier Ltd. All rights reserved.

1. Introduction
In this paper, we address the design of a supply chain of a two-echelon distribution system. Ballou (1999) classies the
supply chain planning as decisions concerning inventory, transportation and facility location. This work is devoted to facility
location that denes the distribution network in the supply chain.
Network design decisions determine the supply chain conguration and have a signicant impact in logistics costs and
responsiveness (Chopra and Meindl, 2004). For instance, facility location has a long-term impact in the supply chain because
of the high cost to open a facility or to move it. While cost related to opening a new facility and inventory pooling costs induce to reduce the number of facilities, responsiveness causes a contrary effect. A high number of facilities may reduce the
lead time to deliver a product to the nal customer. In certain products fast lead time can be viewed as an added value so
that the rm that makes them available rst can obtain short- and long-term competitive advantages in the market.
In this study we consider a set of potential locations for new distribution centers. Each candidate site has a xed cost for
opening a facility with a limited capacity. In the supply network the number and location of plants and customers are known.
There are several transportation options available for each pair of facilities between echelons. The alternatives are generated by the supplier from different companies, the availability of different types of service at each company (e.g. express
and regular), and the use of different modes of transportation (e.g. truck, rail, airplane, ship or inter-modal). Commonly,
these alternatives involve an inverse correspondence between time and cost, i.e. a faster service will be more expensive.
In order to balance the economic concerns with prompt demand satisfaction, our approach is to minimize the total cost
and the maximum time needed for shipping the product across the whole supply chain. The presence of multiple objectives

* Corresponding author. Address: Apdo. Postal 111-F, Ciudad Universitaria, San Nicol s de los Garza, C.P. 66450, N.L., Mexico. Tel.: +52 081 1492 0368;
fax: +52 081 1052 3321.
E-mail address: deniz@yalma.me.uanl.mx (D. Ozdemir).
0968-090X/$ - see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.trc.2010.04.003

822

Y. Cardona-Valds et al. / Transportation Research Part C 19 (2011) 821832

in a problem, in principle gives rise to a set of optimal solutions (largely known as Pareto-optimal solutions), instead of a
single optimal solution so we focus on getting the Pareto front.
The contributions of the paper are twofold. First, we extend a previously studied problem in order to include more realistic and practical assumptions. We model it as a bi-objective stochastic mixed integer linear problem. In addition, we propose a solution algorithm based on the fusion of the -constraint and the L-shaped method.
The remainder of this paper is organized as follows: in the next section we present related studies found in academic literature. Section 3 describes the problem and provide the mathematical model. Section 4 proposes our solution methodology,
based on the fusion of the -constraint and the L-shaped method. Section 5 reports numerical results concerning our solution
methods performance as well as a study about the impact of some design system parameters on the problem complexity.
Section 6 concludes with our main ndings and future research avenues.
2. Literature review
This bi-objective problem was rst introduced by Olivares et al. (2007) as the Capacitated Fixed Cost Facility Location
Problem with Transportation Choices. In their contribution, the authors consider that all the design parameters are
deterministic.
However, in practice, supply chains are characterized by numerous sources of technical and commercial uncertainties.
Critical parameters as customer demands, prices and future facility capacities are quite uncertain. The fact that meeting customer demand is what mainly drives most supply chain initiatives motivated us to study the problem considering that the
demand is a random variable whose value is not known at the time of designing the network.
Literature reveals that there are several studies which deal with uncertainty in supply chain management (SCM) at different levels. In tactical level supply chain planning we can mention, for example, recent papers related to the distribution of
raw materials and products (Van Landeghem and Vanmaele, 2002; Yu and Li, 2000). At the strategic level, there is a great
deal of research in the facility location component of supply chain network design under uncertainty. A good review can
be found in the study of Snyder (2006).
The optimization focus in traditional SCM problems are maximizing prot or minimizing costs as a single objective (Tsiakis et al., 2001; Santoso et al., 2005; Elhedhli and Gzara, 2008). Nevertheless, other criteria to meet customer demand on time
such as customer response time, or ll rate should be taken into account because they are related to the most basic functions
of the SCM: to meet customer requirements.
In the last years customer response time related considerations have been revisited in the distribution network design
(Erol and Ferrell, 2004; De Toni and Tonchia, 2001). Controlling lead time is becoming a competitive advantage for many
rms because of the transformation of the manufacturing-distribution chains throughout the world. This parameter has effects on costs and also can be affected by the supply chain conguration.
Papers involving an integrated design of supply chain networks under uncertainty and considering several objectives is
signicantly smaller in number (Sabri and Beamon, 2000; Chen et al., 2008; Guilln et al., 2005).
The previous related works consider either single aspect of the supply chain design (only assignment or transportation or
other), or consider only single level. In most of the studies single objective is considered and uncertainty on parameters are
ignored. To the best of our knowledge, the problem presented in this study has not been considered in literature in multiechelon systems.
In the next section we describe the problem we are considering and its mathematical formulation.
3. Problem description
In this study we assume that the response time is inuenced by the selection of the transportation channel between facilities. The existence of third party logistic companies allows that different transportation services are available in the market,
so in this paper we consider several alternatives to transport the product between facilities, where each option represents a
type of service with associated cost and time parameters. The implicit assumption is that a faster transportation mode is also
the more expensive one, creating a tradeoff between cost and time that affects the distribution network conguration.
The selection of a transportation channel has commonly been limited to the transportation mode. In this study, the term
transportation channel is more generic and includes not only choices for the transportation mode but also for different
types of services from one or several transportation companies. Although the principles may be the same, this distinction
is important to describe a more general case.
Therefore, on one side, the objective of this problem is to select the appropriate sites to open distribution centers (DC) and
determine the ow between facilities to minimize the total cost (involving both infrastructure and transportation costs), taking into account the uncertainty in customers demands. On the other side, it is desired to minimize the transportation time
from the plant to the customers. This part of the problem determines which services will be selected in order to reduce the
transportation time in each echelon of the supply chain. Hence, the tradeoff between cost and time creates a bi-objective
problem. One criterion tries to minimize the combined cost of transportation and facility location. The other criterion looks
for the reduction of the time to transport the product along the supply chain. Even though there are studies that identify the
cost-time tradeoff as an important element in supply chain design, they do not relate this balance to the availability of

Y. Cardona-Valds et al. / Transportation Research Part C 19 (2011) 821832

823

transportation choices between facilities (Zhou et al., 2003; Truong and Azadivar, 2005). In the paper of Eskigun et al. (2005),
authors use an aggregated function for time and cost. Although different transportation modes are included in their model
(rail and truck), the problem is to select between a direct and an inter-modal shipping strategy. They do not have transportation choices between each pair of locations.
The cost-time tradeoff, in conjunction with the uncertainty in demands, means that we are handling a novel multi-objective stochastic optimization problem.
The Stochastic Capacitated Fixed Cost Facility Location Problem with Transportation Choices (SCFCLP-TC) is based on a
two-echelon system for the distribution of one product in a single time period. In the rst echelon the manufacturing plants
send product to distribution centers. The second echelon corresponds to the ow of product from the distribution centers to
the customers. The number of customers as well as the number, locations and manufacturing capacities of the plants are
known. There is a set of potential locations to open distribution centers. The number of open DC is not dened a priori. Each
candidate site has a xed cost for installing a DC, where the DC will have a limited operational capacity. There are several
transportation options available for each pair of facilities between echelons. Each option represents a type of service with
associated cost and time parameters. Each customer has an associated product demand, which must be supplied from a
single DC. The exact demand realization is not known in advance. Thus the demand will be considered as a random variable
modeled through scenarios. We are aiming at a single network design, that is, the network will be the same in any scenario
realization. However, the ows may be different, depending on the demand scenario realizations. The decisions to be taken
are the location of the distribution centers to be open, the selection of the transportation channels and the ow between
facilities to meet the customers demands in any scenario realization. The decision related to the transportation options
has an impact on the transportation time from the plant to the customer. The tradeoff between cost and time must be considered in the formulation of a mathematical model that minimizes both criteria. Hence, the problem should be addressed
through a bi-objective optimization model. Following this approach, one criterion minimizes the combined cost of transportation and facility location. The other criterion looks for the minimum time to transport the product along any path from the
plants to the customers. In order to formulate de mathematical model let us introduce the following notation.
Sets:
I
J
K
LPij
LWjk
S

set
set
set
set
set
set

of
of
of
of
of
of

plants i
potential locations j for distribution center
customers k
arcs between nodes i and j
arcs between nodes j and k
scenarios for the demand

Parameters:
CPij
cost of transporting one unit of product from plant i to distribution center j using arc
cost of sending one unit of product from distribution center j to customer k using arc
CWjk
time for transporting any quantity of product from plant i to distribution center j using arc
TPij
time for transporting any quantity of product from distribution center j to customer k by arc
TWjk
capacity of plant i
MPi
capacity of distribution center j
MWj
xed cost for opening distribution center j
Fj
demand at customer k in scenario s
Dsk
probability of realization of scenario s
ps
Decision
X sij
Y sjk
Zj
Aij
Bjk

variables
quantity transported from plant i to DC j in scenario s using arc
quantity transported from DC j to customer k in scenario s using arc .
binary variable equal to 1 if DC j is open and equal to 0 otherwise
binary variable equal to 1 if arc connecting plant i and DC j is used and equal to 0 otherwise
binary variable equal to 1 if arc connecting DC j and customer k is used and equal to 0 otherwise

The mathematical formulation of the addressed problem is described in the following lines. As mentioned before, our
model has two objectives to capture: the economical and service quality considerations of the decision maker. The objective
that minimizes the cost for opening DC and the expected value of the transportation costs is represented by the following
function:

min f1

X
j2J

F j Zj

X
s2S

0
ps @

XX X
i2I

j2J 2LP ij

CP ij X sij

XX X
j2J

1
CW jk Y sjk A

k2K 2LW jk

The objective that minimizes the sum of the maximum lead time from the plants to the customers through each distribution center is expressed as follows:

824

Y. Cardona-Valds et al. / Transportation Research Part C 19 (2011) 821832

min f2 maxmaxTPij Aij maxTW jk Bjk


j

i;

k;

To guarantee that the demand of each customer is satised in any possible scenario, it is necessary to establish the following equality:

X X

Y sjk Dsk

8s 2 S; k 2 K

j2J 2LW jk

For not exceeding capacity limits of the plants, constraints (4) are imposed.

X X

X sij 6 MP i

8s 2 S; i 2 I

j2J 2LP ij

The following constraints guarantee that the ow going out from a DC does not exceed the capacity of the facility. At the
same time ensure that the ow of product is done only through open DCs.

MW j Z j 

X X

Y sjk P 0 8s 2 S; j 2 J

k2K 2LW jk

In order to keep the ow balance at each DC we impose:

X X

X sij 

X X

i2I 2LP ij

Y sjk 0 8s 2 S; j 2 J

k2K 2LW jk

Constraint (7) establishes that each distribution center must be supplied by a single source.

X X

Bjk 1 8k 2 K

j2J 2LW jk

In order to guarantee single sourcing, at most one arc can be selected between nodes ij and nodes jk. Therefore, we
impose constraints (8) and (9) respectively.

Aij 6 1 8i 2 I; j 2 J

2LPij

Bjk 6 1 8j 2 J; k 2 K

2LW jk

The following set of constraints ensures that a distribution center is not open if it has no active incident arc on it, i.e. the
DC does not receive ow of product.

X X

Aij  Z j P 0 8j 2 J

10

i2I 2LP ij

As we are looking for a network design that should be identical for any scenario realization, then if one arc is used in some
scenario, it must be in the chosen network design as states constraints (11) and (12) below:

X sij  Aij P 0 8i 2 I; j 2 J; 2 LP ij

11

Y sjk  Bjk P 0 8j 2 J; k 2 K; 2 LW jk

12

s2S

X
s2S

The ow is allowed only through active arcs, that is, arcs chosen for the network design:

MPi Aij  X sij P 0 8s 2 S; i 2 I; j 2 J; 2 LPij


MW j Bjk 

Y sjk

P 0 8s 2 S; j 2 J; k 2 K; 2 LW jk

13
14

Constraint set (15) is for continuous nonnegative variables. Binary variables are required in constraint (16).

X sij1 ; Y sjk2 P 0 8s 2 S; i 2 I; j 2 J; k 2 K; 1 2 LP ij ; 2 2 LW jk

15

Z j ; Aij1 ; Bjk2 2 f0; 1g 8i 2 I; j 2 J; k 2 K; 1 2 LP ij ; 2 2 LW jk

16

If demands and capacities are integer it is not necessary to change constraint (15) to require integer values for X sij and Y sjk .
Since once the values for Xij, Yjk and Zj are xed, the remaining structure is a transportation problem. It is well known that
the unimodularity property produces integer results under such condition.
Note that the presented model is not linear, due to nonlinear time objective f2, which is represented in Eq. (2). In order to
get a linear model, the following auxiliary variables have to be dened.

Y. Cardona-Valds et al. / Transportation Research Part C 19 (2011) 821832

825

Auxiliary variables
E1j
auxiliary variable equal to the maximum time in the rst echelon of the supply chain for active DC j, i.e.
E1j maxTPij Aij
E2j
auxiliary variable equal to the maximum time in the second echelon of the supply chain for active DC j, i.e.
E2j maxTW jk Bjk
T
auxiliary variable equal to the maximum time for sending the product from any plant to any distribution center, i.e.
T maxE1j E2j
Moreover, constraints (17)(20) should be added to the model.

T  E1j  E2j P 0 8j 2 J

17

E1j

 TPij Aij P 0 8i 2 I; j 2 J; 2 LP ij

18

E2j

 TW jk Bjk P 0 8j 2 J; k 2 K; 2 LW jk

19

T; E1j ; E2j P 0 j 2 J

20

As a result, the stochastic bi-objective mixed integer linear model can be expressed as:

min f1

F j Zj

j2J

0
ps @

s2S

XX X
i2I

CP ij X sij

XX X

j2J 2LPij

j2J

1
CW jk Y sjk A

k2K 2LW jk

min f2 T
subject to constraints (3)(20).
3.1. Problem formulation as a two-stage stochastic model
Given the nature of the problem, the choice of locations has to be decided before knowing the actual realization of demand on the future. The product ow on the network is decided once the uncertainty of demand has been resolved. Therefore, it makes sense to develop a two-stage stochastic linear model with recourse, where a long-term anticipatory decision
(choice of locations) must be made prior to have full information about random parameters of the problem (i.e. customer
demands) and short-term decisions are available as recourse actions once the uncertainty has been revealed. The aim is
to determine a here-and-now decision which minimizes the total expected costs associated with both the long-term and
the short-term decisions.
The general formulation of a two-stage stochastic program with xed recourse is:

min

z ct x En fmin qxyjW hx  Txx; y 2 Yg

subject to : Ax b x 2 X

21

where the rst-stage decisons are represented by the n1  1 vector x. Corresponding to x are known data vectors c, b and the
matrix A of sizes n1  1, m1  1 and m1  n1, respectively. And the second-stage decisions are represented by the n2  1 vector y. In the second stage, a number of random events x 2 X may realize. For a given realization of x, the second-stage data
q(x), h(x) and T(x) become known, where q(x) is n2  1, h(x) is m2  1 and T(x) is m2  n2. n is the vector formed by the
components q, h and T, and En is the expected value respect to n. W is called the recourse matrix which is xed (it means, do
not depend on the random variable).
Under this structure, we will consider as a priori decisions the distribution center location and the service choice on the
two levels of the supply network (variables Zj, Aij, Bjk), while the a posteriori decisions will be the ows between facilities
(variables X sij ; Y sjk ). Note that the selection of the service choice (that is, the transportation channel) implicitly determines the
distribution center-plant assignments and the distribution center-customer assignments.
We are considering uncertainty only at the customers demands, that means that our unique random parameter is the
right hand side h(x); the matrices T and W are xed.
In general, we have no guarantee that the relatively complete recourse (RCR) property is satised. This property would imply that every network design we get at stage one is guaranteed to be feasible in the second one (Birge and Louveaux, 1997).
This would allow us to use any well developed solution method. In order to guaranty the RCR property within realistic
assumptions we introduce a ctitious distribution center with zero xed cost for opening and sufcient capacity to absorb
any customer demand. Transportation costs from the ctitious distribution center to customers will be large enough so that
a solution containing this ctitious distribution center is highly penalized. Implemented model is adjusted accordingly.
4. Solution method
First of all, it should be noted that as we are tackling a bi-objective problem, its solution will be a Pareto front possible
design alternatives representing the tradeoff between both objectives rather than a unique solution. In general, solutions to

826

Y. Cardona-Valds et al. / Transportation Research Part C 19 (2011) 821832

stochastic multi-objective optimization problems are obtained in two phases. In one of them, the stochastic problem is transformed into its equivalent deterministic problem, and in the other one, some of the existing techniques in multi-objective
programming is applied to obtain efcient solutions, which involves transforming the multi-objective problem into a problem with only one objective function (Caballero et al., 2004). That is, the solution of such problems involves a double transformation: transforming the multi-objective problem into a problem with only one objective function and the stochastic
problem into its equivalent deterministic problem.
Depending on the order in which the transformations are carried out, it is possible to classify the existing techniques for
the solution of this kind of problems (Ben-Abdelazis, 1992; Caballero et al., 2004). The multi-objective approach refers to the
set of techniques that solve the problem by considering rst the transformation of the stochastic multi-objective problem
into its equivalent multi-objective deterministic problem. The stochastic approach refers to those techniques that transform
the stochastic multi-objective problem into a stochastic problem with only one objective function.
The solution approach that we propose for solving SCFCLP-TC can be classied as a stochastic approach. As it is illustrated
in Fig. 1 our approach is an iterative algorithm consisting on a Benders decomposition embedded in the -constraint method,
which is a well known scalarization technique to transform a multi-objective problem into a single-objective one (Ehrgott,
2005). The values a and b specify the range of variation for the objective function f2.
In the following sub-section we explain the details.
4.1. Transforming SCFCLP-TC into a single-objective problem by the -constraint method
In the -constraint method one objective function is selected as main objective function and the others are transformed in
constraints as shown below.
In general, the multi-objective optimization problem with p objectives is dened as

minf1 x; . . . ; fp x
x2X

and is substituted by the -constraint

min
x2X

f j x

subject to : f k x 6 k ;

k 1; . . . ; p; kj

22

where  = (1,2,. . .,p) is the vector of right hand side of all -constraint.
The procedure to select the objective functions that will be treated as constraints and to assign values for the k parameters is problem-dependent. The transformation selected for implementing the -constraint method to SFCLP-TC has the

Fig. 1. Flow diagram showing the solution approach proposed for the stochastic multi-objective problem.

827

Y. Cardona-Valds et al. / Transportation Research Part C 19 (2011) 821832

objective function f1 (cost) as objective function and the function f2 (time) as a constraint. There are two reasons for doing so.
The rst one is the physical meaning and importance of the objectives. Then we give higher priority to the economical objective than the customer service. The second reason is the suitability of the optimization. For example, the discretization of the
time parameters induces to integer values of objective f2. Therefore, the value of the parameter  can be changed by one unit
from one iteration to the next one. This possibility solves the problem of dening how to change the value of that parameter.
If objective function f1 were selected as the constraint, the rule to change the value of  should be more elaborated because of
the continuous value of this objective. Another consideration is the difference in the ranges for both objectives. While these
ranges depend on the instance data is expected that the range of the cost function is in the order of millions of monetary
units and the range of the time function is in the order of tenths of time units. Hence the number of iterations may be potentially high if objective f1 is transformed to a constraint.
A negative correlation between two objective functions has the effect that for the highest values of one criterion the other
exhibits the lowest values. Hence, a natural implementation is that once one criterion has been selected as main objective
function, the values of  for the objective function treated as constraint are reduced iteratively. To obtain the interval where
the parameter  varies we solve each single-objective deterministic problem. It is expected that the time objective takes its
highest value when the objective function f1 is optimized.
4.2. Solving the stochastic mono-objective problem by the L-shaped method
The L-shaped method is an extension of the Benders decomposition to take care of feasibility questions in Stochastic Programming. First of all, note that the general problem (21) established in sub-section 3.1 is equivalent to the so-called deterministic equivalent problem

z ct x Qx

min

23

subject to : Ax b x P 0
where Qx En Q x; nx and

Q x; nx minfqxt yjWy hx  Txx; y P 0g

24

where c, b, A the deterministic parameters and q(x), h(w), T(x) the stochastic ones, are dened just as before.
The L-shaped method consists of solving an approximation of the model presented in (23) and (24) by using an outer linearization of Qx. Two types of constraints are, in general, sequentially added: feasibility cuts and optimality cuts. The feasibility cuts determine fxjQx < 1g and the optimality cuts are linear approximations to Qx on its domain of niteness. It
proceeds as follows (Birge and Louveaux, 1997):
1. Set q: = t: = v: = 0. The value of h is set equal to 1 or to an appropriate approximation of Qx
ignored in the computation.
2. Set v: = v + 1. Solve the master problem

s2S ps Q x; w

and is

min
z ct x h
subject to : Ax b
Dj x P d j ;

j 1; . . . ; q;

Ej x h P ej ;
x P 0;

25

j 1; . . . ; t

h2R

P
3. Evaluate Qxv s2S ps Q xv ; xs .
v
4. If Qx 1 means that there is some

xs such that Qxv ; xs 1, add a feasibility cut

rt hxs  Txs x 6 0
Set q: = q + 1 and go back to step 2.
5. If Qxv < 1 it is possible to solve the s linear programming problems and then get the dual optimal solution ps to
form the subgradient

u

ps pts Ts 2 @Qxv

s2S

6. If Qxv 6 hv ;
 xv is the optimal solution, the approximation is exact and minimized.
Otherwise add an optimality cut to the master problem (25)

h P Qxv ut x  xv
Set t: = t + 1; Go back to step 2.

26

828

Y. Cardona-Valds et al. / Transportation Research Part C 19 (2011) 821832

In our implementation, two facts have been taken into account. First, we introduced a ctitious distribution center with
enough capacity and a zero xed cost to meet any customer demand. Thus, our problem has a feasible solution in any iteration. In other words we do not need to add feasibility cuts (i.e. step 4 is never visited) thanks to we have guarantied the
relatively complete recourse property.
Second, in our problem some of the variables in the rst stage are restricted to be binary, following approach of Laporte
and Louveaux (1993) we add the following cut in addition to the above optimality cut:
Set xi = 1, i 2 Sr and xi = 0, i R Sr be the rth feasible solution and Q x the corresponding second-stage value. Dene optimality cut as

h 6 Qxr  L

xi 

i2Sr

!
xi

 Qxr  LjSr j  1 L

27

iRSr

where jSrj is the cardinality of Sr and L is a lower bound on Qx which can be presented as

L minfhjAx b; 0 6 x 6 1yh; x satisfies 26 for j 1; . . . ; qg

28

x;h

Then, the set of optimality cuts (27) dened for all rst-stage feasible solutions is a valid sets of optimality cuts.
As an example consider the network of three plants (circles), three distribution centers (hexagons) and four customers
(rectangles) shown in Fig. 2. We assume that  2 [6, 17] and there are two demand scenarios each with probability 12. There

2780
5638

7402
T=6
15958
5644

14462

4727
6289
2864
8824

f1 = 726,152
f2 = 6

2675
9669
2780
5638
4727

5455
T = 12
15307
7591

15113

f1 = 578,550
f2 = 12

6289
2864
8824
2675
9669
2780
5638

7591
T = 17
15113
5455

15307

4727
6289

f1 = 492,933
f2 = 17

2864
8824
2675
9669

Fig. 2. Network congurations and ow assignment for T = 6, T = 12 and T = 17.

829

Y. Cardona-Valds et al. / Transportation Research Part C 19 (2011) 821832

are two transportation modes where mode one is illustrated with solid line and mode two with dashed line. As this example
is considering two scenarios, the demands are represented numerically on the left side of the rectangle. The up left number
corresponds to the high demand, and the down left number corresponds to the medium demand. The numbers up and down
the hexagons shows the ow from plant to distribution center for the high demand scenario, and medium demand scenario,
respectively.
The black hexagons indicate an opened distribution centers. As the Fig. 2 shows, the conguration for each point is the
same for the two scenarios involved and for each point the cost decreases whilw the respond time increases.
5. Some numerical results
Several experiments were performed in order to asses the proposed method as well as to study the impact of various
problem parameters on problem complexity.
Solution algorithm is implemented in C++ calling CPLEX 11.2 in a SunFire workstation with Sun Fire V440 processor, connected to 4 Processors 1602 Ghz, Ultra SPARC III with 1 MB cache and 8 GB memory.
5.1. Study on distribution center capacity and customer demand variability
In order to illustrate the impact of various problem parameters on problem complexity, we studied the variability in demand (through the probability assigned to each scenario) and the capacities of the distribution centers. The stochastic part of
the problem is solved by formulating the equivalent deterministic problem.
As summarized in Table 1 we generate ve congurations where the second column shows the number of plants (P), the
third the number of candidate distribution centers (DC), the fourth shows the number of customers (C), the fth is the number of transportation channels (TM), and the sixth and last column is the number of demand scenarios (S).
To study the impact of the design problem parameters on the solution of stochastic multi-objective problem three sets of
demand realization probabilities were generated such that 15 instances were created and each conguration set pair were
replicated three times by generating random instances.
For each conguration three cases of demand variability are determined and associated demand realization probabilities
are presented in Table 2. Set I represents the low variability case, in set II each scenario is equally likely and set III has high
demand variability. The remaining parameters such as plant and distribution center capacities, transportation cost and expected demands kept xed throughout the cases.
The number of points in the Pareto front and the average solution time (in seconds) for the three sets of each group is
presented in Table 3.
As a result of our computational study on the demand variability, as it can be seen in Table 3, we observed that the variation on the demand realization probabilities does not have signicant effect neither on the number of points in the Pareto
fronts nor on the computational time to build it. Therefore, it can be concluded that the failure in considering a higher variation in the probability scenarios such as performed in this experiment would not signicantly inuence the results.
In another set of runs, we investigate the impact of capacity of potential of distribution center parameters on the solution
efciency behavior. We consider again for each conguration three sets (different distribution center capacity parameters),
with three instances each. Note that group one and three have the same number of customers and transportation modes, so
they share the information given at Table 4. Since the demand of each client needs to be completely satised as well as

Table 1
System congurations.
No.

Group

1
2
3
4
5

DC

TM

3
3
3
5
5

3
3
5
3
3

4
12
8
4
16

2
4
6
2
6

2
2
3
2
3

Table 2
Study sets for demand realization probabilities.
Set

Two scenarios

3
10

7
10

Three scenarios
4
10

2
10

4
10

II

1
2

1
2

1
3

1
3

1
3

III

9
10

1
10

2
10

6
10

2
10

830

Y. Cardona-Valds et al. / Transportation Research Part C 19 (2011) 821832

Table 3
Computational results about the demand variability on customers.
Group

1
2
3
4
5

Set I

Set II

Set III

No. of points

Time (s)

No. of points

Time (s)

No. of points

Time (s)

16
29
28.66.33
19.33
27.33

7.75
464.54
1543.34
9.52
7941.92

19.66
29
30.33
19.66
25

9.19
430.20
1577.12
10.66
12886.23

19
29.33
30.33
20
27.66

10.28
494.49
1609.09
13
16524.91

Table 4
Parameter variation for distribution center capacity.
Customers

Scenarios

Low capacity

Random capacity

High capacity

4
8
12
16

2
3
2
3

40,000
160,000
120,000
320,000

U(MWlow,
U(MWlow,
U(MWlow,
U(MWlow,

70,000
300,000
230,000
620,000

MWhigh)
MWhigh)
MWhigh)
MWhigh)

Table 5
Computational results about the capacity on distribution centers.
Group

Low capacity

1
2
3
4
5

U(MWlow, MWhigh)

High capacity

No. of points

Time (s)

No. of points

Time (s)

No. of points

Time (s)

17.33
29.66
21
20.66
30.66

6.74
308.58
35
9.97
537.69

19.66
29
30.33
19.66
25

9.19
430.20
1577.12
10.66
12886.23

13.33
28
20
17.66
30.66

5.84
112.72
35.07
8.04
538.50

Table 6
Instances for the L-shaped method.
No.

Group

1
2
3
4
5

DC

TM

3
3
3
3
3

3
3
3
3
3

4
4
4
4
6

2
2
4
4
2

2
3
2
3
2

supplied by a single distribution center, the capacity on each distribution center must be greater than the maximum demand. Therefore,

MW j 6 max Dk
k2K

8j 2 J

Each demand realization is uniform randomly generated, for the rst scenario we consider [1,000, 5,000], for the second one
[5,000, 10,000 ] and for the third one [10,000, 20,000 ]. Remember as well that the demand of each client needs to be completely satised as well as supplied by a single distribution center, so for the low capacity we consider as well the number of
scenarios involved as the number of customer that need to be satised. Then, the Low capacity is the maximum demand i.e.,
product between the number of customers and 10,000 or 20,000 (corresponding to the right value on the range interval of
the second or third scenarios) and the High capacity is twice the Low capacity less 10,000 or 20,000 as before.
For random capacity case, capacity of each distribution center is generated from a uniform distribution
P
[MWlow, MWhigh]. Where MWlow = maxk2KDk, MWhigh = DT + (DT  MWlow) = 2DT  MWlow, and DT k2K Dk . In order to
determine capacity parameters, we used the highest possible demand realizations. The computational results are summarized in Table 5.
Regarding the computational study on the distribution center capacities, it is important to note the following. As shown in
Table 5 the average number of points in the Pareto front obtained when capacity is generated through uniform distribution,
are greater than the number of points obtained when capacity is set to low or high value. However, as it can be seen in

Y. Cardona-Valds et al. / Transportation Research Part C 19 (2011) 821832

831

Table 7
Computational results about the L-shaped method.

Group

No. of points

Time (s)

1
2*
3
4*
5

15.66
17
18.30
26.66
16.33

43.87
12.66
1966.97
71.84
8624.05

The groups solved incorporating integer cut (27).

groups 3 and 5, the time to nd the Pareto front when the capacity is generated from a uniform distribution can be significantly higher with respect to the time for the low or high capacity.
Based on this experiment regarding to the impact of distribution center capacity on the computational efciency, we can
conclude that it has inuence in the number of points of the Pareto front that are found as well as in the computational time
required to nd them.

5.2. Performance of the solution method


In the second part of the study in order to study the stochastic multi-objective problem and test the methodology proposed in Fig. 1, we consider ve different congurations summarized in Table 6 and then we generate three repetition of
each conguration.
Table 7 reports the computational results: we show the number of points on the Pareto front as well as the computational
time to nd the Pareto front. Each entry in the table is the average for the three instances of each group.
Please note that two of the groups (indicated with an asterisk) were solved incorporating integer cut (27). It is important
to remark this because using L-shaped method with standard cut (26) it was not possible to determine the Pareto front in a
reasonable time of computation.

6. Conclusions and future works


In this paper we have addressed a problem of designing a supply chain where not only the total cost, but also the maximum time needed for shipping the product across the chain, has to be minimized. Locations for opening distribution centres
as well as transportation channels (from a set of candidates) have to be decided before the actual value of the customers
demand is realized.
A stochastic bi-objective mixed integer linear model is provided. In order to guarantee the relatively complete recourse
property when solving it as a two-stage stochastic model, some additional adjustments were required. We introduced a ctitious distribution center with sufcient capacity to absorb any customer demand. This avoided the addition of feasibility
cuts in the solution method.
In the algorithm that we designed to solve the problem, L-shaped method is embedded into the framework of -constraint
method, so it follows the so-called stochastic approach. That is, we rst transform the stochastic bi-objective problem into a
stochastic problem with only one objective function using the -constraint method and then, we solve the stochastic problem using the L-shaped method.
The experimental part of this work shows that even though it was possible to solve some groups of small instances using
the standard L-shaped method, we could solve instances with three or more scenarios only when we introduced the integer
cut.
Regarding to the study of the inuence of some parameters on the computational time we observed that the variability of
the demands does not affect. However, the number of points in the Pareto front as well as the computational time spent to
obtain them is highly inuenced by the capacity of the distribution centers.
The results obtained so far show that there are multiple future research avenues.
On the exact way we could introduce branch and bound techniques, as suggested by Laporte and Louveaux (1993), in
order to obtain more diversity (measured as the number of points) in the Pareto front. The algorithm they proposed differs
from the classical L-shaped method since rst-stage binary variables are taken into account. The fathoming rules are also
differs from what is usually done in branch and bound. In the sense that nodes are not necessarily fathomed when integrality
conditions are satised. We would expect that this methodology will allow to get more efcient point on the Pareto front.
Another extension of this work is the use of auxiliary cuts that may reduce the computational time. Laporte et al. (1990)
identify two valid lower bounds on h (the auxiliary variable used to approximate the recourse function) by relaxing the constraints of the available demand and available capacity on the function that computes for each n the maximum net revenue.
Or we could alternatively use some of the techniques for the accelerate Benders decomposition as done in the work of Santoso et al. (2005).

832

Y. Cardona-Valds et al. / Transportation Research Part C 19 (2011) 821832

For an approximate solution we could study the problem according to the approaches for multi-objective stochastic optimization suggested (Caballero et al., 2004). On the multi-objective approach we would provide an deterministic equivalent
set, that can be provided for example using the chance constraint programming, and then for providing an approximate efcient set, we could make use of a multi-objective metaheuristic for nding the Pareto front. One of the main advantage of this
methodology is that it could be able to solve bigger instances.
Acknowledgement
Work reported in this paper was partially supported by the Mexican National Council of Science and Technology (Grants
Nos. 61903 and 61343).
References
Ballou, R.H., 1999. Business Logistics Management. Upper Saddle River, USA.
Ben-Abdelazis, F., 1992. Lefcacit en programmation multi-objectifs stochastique. Ph.D. thesis, Universit de Laval, Qubec, Canad.
Birge, J.R., Louveaux, F., 1997. Introduction to Stochastic Programming. Springer-Verlag, New York, USA.
Caballero, R., Cerda, E., Muoz, M., Rey, L., 2004. Stochastic approach versus multiobjective approach for obtaining efcient solutions in stochastic
multiobjective programming problems. European Journal of Operational Research 158 (1), 633648.
Chen, C., Wang, B., Lee, W., 2008. Multi-objective optimization for a multi-enterprise supply chain network. Industrial and Engineering Chemistry Research
42 (67), 18791889.
Chopra, S., Meindl, P., 2004. Supply Chain Management: Strategy, Planning and Operation. Prentice Hall, Upper Saddle River, USA.
De Toni, A., Tonchia, S., 2001. Performance measurement systems: models, characteristics and measures. International Journal of Operations & Production
Management 21 (12), 4670.
Ehrgott, M., 2005. Multicriteria Optimization. Springer-Verlag, Berlin.
Elhedhli, S., Gzara, F., 2008. Integrated design of supply chain networks with three echelons, multiple commodities and technology selection. IIE
Transactions 40 (1), 3144.
Erol, I., Ferrell Jr., W.G., 2004. A methodology to support decision making across the supply chain of an industrial distributor. International Journal of
Production Economics 89 (1), 119129.
Eskigun, E., Uzsoy, R., Preckel, P., Beaujon, G., Krishnan, S., Tew, J., 2005. Outbound supply chain network design with mode selection, lead times and
capacitated vehicle distribution centers. European Journal of Operational Research 165 (1), 182206.
Guilln, G., Mele, F.D., Bagajewicz, M.J., Espua, A., Puigjaner, L., 2005. Multiobjective supply chain design under uncertainty. Chemical Engineering Science
60 (6), 15351553.
Laporte, G., Louveaux, F.V., 1993. The integer L-shaped method for stochastic integer programs with complete recourse. Operation Research Letters 13 (3),
133142.
Laporte, G., Louveaux, F., Van Hamme, L., 1990. Exact solution of a stochastic location problem by an integer L-shaped algorithm. Technical report, NotreDame de la Paix, Sciences Economiques et Sociales, Namur, France.
Olivares, E., Gonzlez-Velarde, J., Rfos-Mercado, R., 2007. A metaheuristic algorithm for a bi-objective supply chain design problem. In: RED-M, pp. 58.
Sabri, E., Beamon, B., 2000. A multi-objective approach to simultaneous strategic and operational planning in supply chain design. The International Journal
of Management Science 28 (5), 581598.
Santoso, T., Ahmed, S., Goetschalckx, M., Shapiro, A., 2005. A stochastic programming approach for supply chain network design under uncertainty.
European Journal of Operational Research 167 (1), 96115.
Snyder, L.V., 2006. Facility location under uncertainty: a review. IIE Transactions 38 (7), 537554.
Truong, T., Azadivar, F., 2005. Optimal design methodologies for conguration of supply chains. International Journal of Production Research 43 (11), 2217
2236.
Tsiakis, P., Shah, N., Pantelides, C., 2001. Design of multi-echelon supply chain networks under demand uncertainty. Industrial and Engineering Chemistry
Research 40 (16), 35853604.
Van Landeghem, H., Vanmaele, H., 2002. Robust planning: a new paradigm for demand chain planning. Journal of Operation Management 20 (6), 769783.
Yu, C.-S., Li, H.-L., 2000. A robust optimization model for stochastic logistic problems. International Journal of Production Economics 64 (13), 385397.
Zhou, G., Min, H., Gen, M., 2003. A genetic algorithm approach to the bi-criteria allocation of customers to warehouses. International Journal of Production
Economics 86 (1), 3545.

You might also like