Sei sulla pagina 1di 10

Expert Systems with Applications 37 (2010) 3094–3103

Contents lists available at ScienceDirect

Expert Systems with Applications


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

The hybrid planning algorithm for the distribution center operation using tabu
search and decomposed optimization
Young Hoon Lee *, Soon Geol Kwon
Department of Information and Industrial Engineering, Yonsei University, 134 Shinchon-Dong, Seodaemun-Gu, SEOUL 120-749, Republic of Korea

a r t i c l e i n f o a b s t r a c t

Keywords: Distribution center operation planning problem to determine the facility locations and the distribution
Distribution center operation planning plans is studied in this paper under the single sourcing constraint, in which each customer can be served
Decomposition by a single supplier. Distribution centers can be installed or changed easily using the outsourcing with a
Heuristic reasonable setup cost, to meet the dynamically changing demand of customers, compared with the plants
Tabu search
which are usually run for the long time period with a heavy setup cost. The objective is to minimize the
sum of transportation, the facility setup and the product handling cost. A mixed integer programming
model for the distribution center operation planning is suggested. For the purpose of practical implemen-
tation, the hybrid heuristic based on the tabu search and the decomposed optimization is proposed with
the priority rule designed by using the so-called Unit Cost Ratio (UCR). The performances of the heuristic
algorithms are evaluated by the comparison with the optimal solution or the best solutions found within
the computational time limit by CPLEX.
Ó 2009 Elsevier Ltd. All rights reserved.

1. Introduction chain. As the demand of the customers is dynamically changing


and the customer satisfaction becomes critical, distribution center
Supply chain network is the logistic network which consists of operation may play a major role in the successful operations. While
facilities, customers, products in the procedure of the planning, the plants are usually operated for the long time period at a coop-
coordination, controlling inventory and distribution. Goods flow erate level, the distribution centers are not necessarily permanent,
through several stages from suppliers to customers, and each stage but can be operated based on the outsourcing with a cost of setup
may consist of many facilities (Sabri & Beamon, 2000). Manage- for the installment and cancellation if needed. Proper management
ment of such logistic networks and business processes has on the distribution centers can be the efficient and profitable pol-
emerged as a major topic in Operations Research (Lee & Billington, icies in supply chain management.
1993). Various logistical functions such as inventory control, distri- The typical example can be found in the electronic industry. One
bution and transportation planning have to be coordinated of the major companies manufacturing the electrical component in
through the networks. It is getting more important to design the Korea has suffered from the coordination of the production and dis-
supply chain networks efficiently due to increasing competitive- tribution. A number of domestic plants are running for the techno-
ness introduced by the market globalization (Thomas & Griffin, logically main processes, and many other plants are also running in
1996). Firms are obliged to maintain high levels of customer ser- China, Malaysia, Indonesia and Thailand. Final goods are supplied to
vice satisfaction while they are forced to reduce cost and maintain the PC or cellular phone assemblers located in Korea, Japan and Chi-
profit margins at the same time (Altiparmak, Gen, & Lin, 2006). It is na. Some sub-assembly components not finished may be trans-
quite common nowadays to see manufacturers and retailers such ferred to other plants where the corresponding manufacturing
as Proctor & Gamble and Wall-Mart joining efforts to efficiently processes cannot be operated because of the technological critical-
handle the flow of products and to coordinate the production ity. Several distribution centers are usually operated for the purpose
and supply chain network system (Amiri, 2006). of the in-time supply of proper final goods for the assemblers, in
Distribution centers in the supply chain network are the inter- which allocations or classifications of the components are per-
mediate facilities connecting the suppliers of plants and the cus- formed, or inventories are held to avoid the starvation of the supply.
tomers consuming products. Appropriate design on the locations This paper concerns the supply chain design with emphasis on
and the size of the distribution centers is the most comprehensive the decisions for the distribution centers and the resultant product
decision problems that need to be optimized for the whole supply transportation routings with a minimal cost. The main reason of
developing heuristics is to obtain the practical convenience in
* Corresponding author. Tel.: +82 2 2123 4813; fax: +82 2 364 7807. terms of the computational time required to get solutions. Another
E-mail address: youngh@yonsei.ac.kr (Y.H. Lee). reason to prefer the heuristics, as suggested using the priority rules

0957-4174/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2009.09.020
Y.H. Lee, S.G. Kwon / Expert Systems with Applications 37 (2010) 3094–3103 3095

in the followings, is that the practitioners want to get solutions Surveys for location and allocation problem can be found in Ai-
with predictability. Most optimal solutions from a mathematical kens (1985), Brandeau and Chiu (1989), Avella et al. (1998). For
programming, although it optimizes the total costs for the whole more recent studies, Pirkul and Jayaraman (1998) have investi-
supply chain networks, may suffer from the drastic change in the gated capacitated facilities (plant and warehouse) location and de-
decisions on each center even with the small change of the data gi- mand distribution to satisfy demands for several products. They
ven. For example, one distribution center may change the products formulated the supply chain model using the mixed integer pro-
handling period by period although the customer demand is gramming and developed heuristic approach based on Lagrangean
changing by the small amount since it may lead to the optimal cost relaxation. Syarif et al. (2002) suggested the mixed integer pro-
reduction overall. One of the factors which need to be considered gramming for the logistic chain network problem. The designing
in the supply chain operation, although not appeared in the objec- issue involves the choice of the facilities (plants and distribution
tive functions, is the easiness in implementation, in other words, centers) to open, and the distribution network design to meet
predictability. Managers of the distribution centers want to have the demand with a minimal cost. They have proposed, for the solu-
solutions or decisions which are similar to the previous ones if tion method, the spanning tree-based genetic algorithm. The PLOT
the situations of the nearby facilities, for example, the upstream (production, logistics, outbound, transportation design system) is
plants and the downstream customers, do not change much. Heu- explained as a decision making process of selecting the set of facil-
ristics are developed by considering the data of the nearby facilities ities in Jararaman and Ross (2003), considering cross-docking sys-
through the priority rules. tems based on a simulated annealing approach. A holistic approach
The suggested model can be defined by following assumptions is suggested in the plant location problem in Bhattacharya, Sarkar,
and characteristics. and Mukherjee (2004). Manzini, Gamberi, and Regattieri (2006)
has suggested the mixed integer programming model in the design
(1) There are multi-stages, multi products on a single period. of the distribution logistic network. More recently Amiri (2006)
Supply chain network consists of plants, distribution centers studied a distribution network design problem in the supply chain
and customers. Each stage may consist of many facilities and system that involves locating plants and facilities, in which the set
each customer has demand of only one type of product. of operational facilities and their capacities are determined using
(2) The costs considered in this model are the distribution cen- Lagrangean relaxation. Construction of the efficient logistics net-
ter setup cost, transportation cost and product handling cost. work is always a hot issue, and the many kinds of evaluation heu-
The cost varies depending on the product family because of ristics are suggested. Venkateswaran and Son (2004) have utilized
the handling process. the distributed simulation method in the design and evaluation of
(3) The single sourcing constraint: each customer can be served the supply chains. Development and application of heuristics are
by only one distribution center. If a customer has to be suggested in determination of the inventory location of a distribu-
served by more than one distribution center, orders can then tion warehouse.
be split into several units satisfying single sourcing con- The meta-heuristic has been used in many models of supply
straint. Single sourcing constraint condition is applied to chain management. Vergana, Khouja, and Michalewicz (2002) used
the distribution center and customer stage in the supply the evolutionary algorithm to solve the replenishment problem.
chain network. Ross (2000) suggested the problem of reconfiguring an existing dis-
tribution network. Jararaman and Ross (2003) have expanded the
The objective of this problem is simultaneously to determine problem of Ross (2000) consisting of one manufacturing plant, mul-
distribution center locations among the candidate facility set, tiple distribution center and retailers. They have developed a good
and the product distribution plan to meet customer demands, with heuristics using simulated annealing for the problem of distribution
a minimal total cost of distribution center setup cost, product han- center location and allocation of retailer and vehicles. Syarif et al.
dling cost and transportation cost. (2002) have suggested a heuristics using a genetic algorithm for
the supply chain network design problem. Altiparmak et al.
(2006) studied a supply chain network design problem and pro-
2. Literature review posed the solution approach based on the genetic algorithm to find
solutions for the multi-objective supply chain network design.
The supply chain management had been studied to perform an Our model differs from the past study in a sense that decisions
efficient physical logistic control coordinating corporation’s differ- on the distribution centers are emphasized which includes the
ent functions such as purchasing, manufacturing, distribution and locations, and the products in operations for the open facility,
sales, etc. (Oliver & Webber, 1982). The optimization models have and the corresponding transportation routes to deliver the prod-
been studied by Efroymson and Ray (1966), Atkins and Shriver ucts. In our study, the fixed cost due to the products is considered
(1968), Khumawala (1972). Most of their researches focused on as well as the facility opening fixed cost, while it is imposed when
constructing models that are relevant to various conditions and the facility is open and established, but independent of kinds of the
developing heuristic algorithms which can be implemented in products to be handled. In practice, the fixed cost of operation may
practice. Problems can be classified by criteria such as single-prod- be charged on which products are handled, since some loading/
uct/single-stage, single-product/multi-stage, multi-product/single- unloading equipment or space requirement may differ by the
stage, multi-product/multi-stage models as shown in Table 1. products. Single sourcing constraint is assumed without loss of

Table 1
Classification of previous research.

Product Stage Literature


Single-product Single-stage Baumol and Wolfe (1958), Kuehn and Hamburger (1963), Efroymson and Ray (1966), Atkins and Shriver (1968), Akinc and
Khumawala (1977), Syarif and Gen (2002)
Single-product Multi-stage Klingman, Randolph, and Fuller (1976), Kim and Kim (1999), Syarif, Yun, and Gen (2002), Altiparmak et al. (2006), Amiri (2006)
Multi-product Single-stage Geoffrion and Graves (1974), Erenguc, Simpson, and Vakharia (1999), Ko, Yang, Jang, and Park (2000)
Multi-product Multi-stage Cohen, Lee, and Moon (1987), Chandra and Fisher (1994), Jayaraman and Pirkul (2001), Vergana et al. (2002), Lee and Lee (2002),
Jararaman and Ross (2003), Gen and Syarif (2005)
3096 Y.H. Lee, S.G. Kwon / Expert Systems with Applications 37 (2010) 3094–3103

generality: any order from the customer can be met by at most one The objective is to minimize the sum of the fixed setup cost,
supplier. Applying product handling cost and single sourcing con- product handling cost, and transportation cost. Fixed setup cost
straint condition can make the problem more realistic. is incurred when the distribution center is determined to open
and the product handling cost is charged when products are
3. Problem statement delivered through a distribution center. Transportation cost is
charged proportional to the transportation quantity from plants
The proposed model is formulated using a mixed integer to distribution centers, and distribution centers to customers.
programming. Constraints (2) imply that the total quantity of transported prod-
ucts from plants to distribution centers may not exceed the
3.1. Indices capacity of the plant. Constraints (3) show that the total quantity
of products transported by an open distribution center to satisfy
i: Index for plants customer’s demand may not exceed the capacity of the distribu-
j: Index for distribution centers tion center. Constraints (4) ensure that the demands of customers
k: Index for customers for each product have to be satisfied by some open distribution
l: Index for products centers. Constraints (5) represent that only the products deter-
mined to be handled at each distribution center are allowed to
be transported, and constraints (6) are balance equations repre-
3.2. Decision variables senting that total quantity of all products transported from plants
to distribution center is equal to that of all products transported
Xijl: the transportation quantity of product l from plant i to dis- from the distribution center to customers. Constraints (7) and (8)
tribution center j. ensure that each customer is served by only one distribution cen-
Yjkl: the transportation quantity of product l from distribution ter by assumptions of a single sourcing constraint. Constraints (9)
center j to customer k. and (10) are the integrality and non-negativity restrictions
ZPjl: 1 if the distribution center j handles product l, and 0 respectively.
otherwise.
ZDj: 1 if the distribution center j is open, and 0 otherwise.
ZSjk: 1 if the distribution center j provides services to the cus- 4. Hybrid heuristic procedure
tomer k, and 0 otherwise.
The usage of the exact mathematical programming in the prac-
3.3. Data tical operation has a couple of drawbacks: the computational
tediousness and the low predictability for the solutions. Although
CPi: the capacity of plant i the computational ability of computers has increased with a great
CDj: the capacity of distribution center j deal, and a number of efficient and powerful software which can
DMkl: the demand of product l for customer k handle a large size programming are provided in a market, the
TPDij: the unit transportation cost from plant i to distribution computational time needed to solve the problem of hundreds of
center j products, hundreds of facility candidates, thousands of customers,
TDCjk: the unit transportation cost from distribution center j to and tens of plants is still cumbersome. In the actual business prac-
customer k tice, decisions on the operations for the distribution centers are
FDj: the setup cost for opening distribution center j evaluated for many scenarios with many kinds of possible data
FPl: the setup cost for handling product l to reach an acceptable solution. The computational time with more
M: Big number than 30 min to get the solution for the model is not desirable in a
practice. Another reason to prefer the heuristic to the optimized
approach is that the solution obtained is easy to expect and to
3.4. Formulation implement. The optimal solution has a tendency to give a totally
different one only with a minor change of the data such as the
X XX XXX transportation cost or the customer demand. Since the decisions
Min FDj ZDj þ FPl ZPjl þ TPDij X ijl
j j l i j l
on the distribution center operations have to be implemented
XXX based on the current operations, drastic changes on the distribu-
þ TDC jk Y jkl ð1Þ
tion centers installment, the handling products, or the transporta-
j k l
XX tion routes therefore may lead to an operational mess. Managerial
s:t: X ijl 6 CPi 8i ð2Þ manpower may ask the evolutionary solutions which can be in-
j l
XX ferred from a current operation and data change.
Y jkl 6 CDj  ZDj 8j ð3Þ Heuristics are constructed using the decomposition of the net-
k l works to gain in the computational efficiency, and the meta-heu-
X
Y jkl P DMkl 8k; l ð4Þ ristics, tabu search, to use the neighbor solutions which can be
j obtained by applying the priority rules of data of the adjacent facil-
X
Y jkl 6 M  ZPjl 8j; l ð5Þ ities and routes to each decision point. The heuristic is processed
XX k
XX by the iterative search of the neighbor solutions starting from
X ijl ¼ Y jkl 8j ð6Þ the initial solution, while checking the tabu list to avoid returning
i
X l k l to the neighbor solution which has visited recently, and in each
ZSjk ¼ 1 8k ð7Þ iteration the neighborhood is generated through the exchange
j and the delete/add of the distribution center and the transporta-
Y jkl 6 M  ZSjk 8j; k; l ð8Þ tion arc which are selected by the priority rule representing the
ZDj ; ZPjl ; ZSjk 2 f0; 1g 8j; k; l ð9Þ average unit cost for the operation.
The original model is decomposed into two sub problems,
X ijl ; Y jkl P 0 8i; j; k; l ð10Þ
the demand decomposition and the supply decomposition. The
Y.H. Lee, S.G. Kwon / Expert Systems with Applications 37 (2010) 3094–3103 3097

demand decomposition is applied in the distribution center and Step 1.2. Calculate UCR-Fj for each closed distribution center.
customer stage, to determine the facility locations and distribution Step 1.3. Sort the distribution center in an increasing order
plans. The supply decomposition is applied in the plant and distri- according to UCR-Fj.
bution center stage to determine the distribution plans. A solution Step 2: Demand decomposition solution
is found in two steps: the demand decomposition the first and the Step 2.1. Select distribution centers sequentially according
supply decomposition the second sequentially. In the demand to the sorted list.
decomposition the priority index is used to select the distribution Step 2.2. Allocate customer’s demands randomly to the
center, the corresponding transportation routes and quantity, and selected distribution centers until the capacity is allowed.
in the supply decomposition the optimal allocation is found for Step 2.3. If all demands are allocated to distribution centers,
the logistics quantity between the plants and the distribution cen- then stop.
ters which are already determined in the demand decomposition. Else, then go to Step 2.1.
Starting from the initial solution the tabu search mechanism works Step 3: Supply decomposition solution
through the generation of the neighbor solution, in which the solu- Step 3.1. Generate supply decomposition solution by using
tions from the demand decomposition and from the supply decom- CPLEX, and combine with the demand decomposition solu-
position are combined. tion to complete an initial solution.
In the demand decomposition the distribution centers and the
transportation arcs are changed or deleted/added to form a neigh-
bor solution using the priority index with a tabu control mecha-
nism. Two kinds of decisions are to be made: the selection of the
distribution centers to be open, and the distribution plan which in-
cludes the product item to be handled in the open distribution cen-
ter, and the transportation route to be delivered to the customers.
In the selection of the distribution centers or the arcs for the trans-
portation, the priority index are used, the Unit Cost Ratio (UCR) in a
greedy manner. UCR is designed to include the most important cost
portion which may give a critical influence in the decisions so that
it works efficiently to select the distribution centers or the trans-
portation arcs.
Three kinds of priority indices are suggested: UCR-F, UCR-O, and
UCR-S, which are used in the generation of the initial solution, in
the distribution center change, and in the arc change of the neigh-
bor solution generation respectively. Whenever the new distribu-
tion center and the new arc are selected for the neighbor
solution, the tabu list is checked to discourage the recently selected
ones not to be trapped in the local optimum.
With the decisions on the distribution centers and the custom-
ers in the demand decomposition, the optimal solution for the
plants and the distribution centers is computed using CPLEX, as
the supply decomposition. The supply decomposition model is rel-
atively small, which makes it possible to get the optimal solution in
a negligible computational time even for the practical size prob-
lem, although it runs as many times as the number of tabu search
iterations. The solutions found in the demand decomposition and
in the supply decomposition are combined to be a complete solu-
tion for the original model in each iteration. Overall procedure of
the suggested heuristic method is shown in Fig. 1.

4.1. Initial solution generation

It is important that initial solution has to be generated so that it


maintains feasibility such as the single sourcing constraint. In an
initial solution search heuristic, the distribution center is selected
based on the priority index, UCR-F, which may indicate the unit
setup or installation cost per unit capacity. The distribution center
with a smaller value of the UCR-F is preferred to be open, and cus-
tomers to serve are selected randomly as long as its capacity is al-
lowed. The priority index UCR-Fj of the distribution center j is
defined as follows:

FDj
UCR  F j ¼ ð11Þ
CDj
Heuristic procedure of generating initial solution is as follows:

Step 1: Initialization and computation of the priority index


Step 1.1. Set all distribution centers closed
Fig. 1. Overall procedure of the heuristic.
3098 Y.H. Lee, S.G. Kwon / Expert Systems with Applications 37 (2010) 3094–3103

4.2. Neighborhood generation: demand decomposition Step 3.1.2 Sort closed distribution centers in an increas-
ing order of UCR-Oj.
In a neighborhood generation of the demand decomposition, Step 3.1.3 Select closed distribution centers sequentially
four types of neighborhood generation methods are suggested: according to the sorted list and check feasibility by com-
The distribution center (DC, hereafter) exchange, DC add, arc ex- paring capacities of both selected open and close distri-
change and arc add. bution center.
Step 3.1.4 If it is feasible, then exchange distribution cen-
4.2.1. DC exchange ters and stop procedure.
DC exchange can generate the neighborhood by replacing the Else, If there exist a closed distribution centers which
open distribution center with the closed distribution center, or vice has not been selected before, then go to Step 3.1.3.
versa. For the DC exchange, the priority index UCR-O is used. Unit Step 3.2 If there exist an open distribution centers which
operational cost in UCR-O is defined to be the sum of the fixed set- were not selected before, then go to Step 3.1 and repeat steps.
up cost for the open DC, and the unit transportation cost of plants Else, stop procedure since DC exchange is not possible.
to distribution centers. Distribution center with smaller value of The example of DC exchange method is shown in Fig. 2, in which
UCR-O is preferred to be open. In an exchange procedure, the open distribution center 3 is replaced with the distribution center 1.
distribution center with higher UCR-O is determined to be closed,
and replaced with the closed distribution center with the lower 4.2.2. DC add
UCR-O to be open. Let NPj denote the number of plants from which In a neighborhood generation procedure, the closed distribution
the products are transported to distribution center j, and let NCj center can be open without any replacement with the open distri-
denote the number of customers served by distribution center j. bution center facility. It may increase the opening fixed setup cost
Let Qj be the transportation quantity passing through the distribu- with a trade off of a decrease of the transportation cost. The distri-
tion center j. UCR-O of DC j is defined as follows: bution center to be open can be selected randomly considering the
P P capacity of replaced distribution centers. If there does not exist a
i TPDij k TDC jk FDj
UCR  Oj ¼ þ þ ð12Þ distribution center which results in a cost reduction, the adding
NP j NC j Qj
DC is then discouraged. To find an optimal condition for the num-
The first term in UCR-Oj implies the average transportation cost per ber of open distribution centers, one closed distribution center is
unit of products delivered to distribution center j, and the second selected and added into the open distribution center sets for each
term does the average transportation cost per unit of products iteration of neighborhood generation, whose procedure is repeated
delivered form distribution center j, and last term does the unit until a termination criteria is applied.
fixed setup cost for transported quantities through distribution cen-
PP
ter j. It is noticeable that Qj is equal to the k l Y jkl in the mathe- 4.2.3. Arc exchange
matical formulation of Section 3. Heuristic procedure using UCR-O Arc exchange method can be used in generation of neighbor-
of generating neighborhood solution by changing distribution cen- hood by exchanging two arcs in the demand decomposition stage.
ter sets is as follows: It may give an reduction in the transportation cost and the product
handling cost by changing distribution centers serving certain set
Step 1: Setting initial condition. of customers. In an arc exchange, the priority index UCR-S is used.
Step 1.1 Classify distribution centers according to whether In an exchange procedure, two arcs are selected randomly and the
open or not. sum of UCR-S for selected arc is calculated. If the sum of UCR-S val-
Step 2: Calculation of UCR-Oj. ues after exchanging is smaller than the one before exchanging,
Step 2.1 Calculate UCR-Oj for each open distribution center. then the arc exchange is encouraged.
Step 2.2 Sort open distribution centers in a decreasing order Let NCj denote the number of customers served by the distribu-
of UCR-Oj. tion center j, and let CPjl denote the number of customers with a
Step 3: Distribution center exchange. demand of product l served by distribution center j. Let QAjk be
Step 3.1 Select open distribution centers sequentially the transportation quantity from the distribution center j to the
according to the sorted list. customer k. UCR-S of arc (j, k) is defined as follows:
Step 3.1.1 Calculate UCR-Oj for closed distribution centers P
FDj l ðFP l =NCP jl Þ
according to transported quantity of selected distribution UCR  Sjk ¼ TDC jk þ þ ð13Þ
NC j  QAjk QAjk
center in Step 2.1.

Fig. 2. Neighborhood generation by DC exchange method.


Y.H. Lee, S.G. Kwon / Expert Systems with Applications 37 (2010) 3094–3103 3099

UCR-Sjk is the sum of three terms: the unit transportation cost, the The example of exchanging arc method is shown in Fig. 3,
setup cost of distribution center j divided by the number of used where arc (2, 4) and arc (4, 2) are exchanged by using UCR-S.
arcs, and the product handling cost divided by the number of used
arcs serving product l through the distribution center j. It is notice- 4.2.4. Arc add
P
able that QAjk is equal to the l Y jkl in the mathematical formulation. As in DC add method, the arc can be added with deleting the
Heuristic procedure using UCR-S of generating a neighborhood current arc connected to the distribution center serving more than
solution by changing arc sets is as follows: two customers to diversify the neighborhood solution. The exam-
ple of neighborhood generation by adding arc method is shown
Step 1: Setting initial conditions. in Fig. 4, where the arc (4, 2) is deleted and the arc (1, 2) is added.
Step 1.1 Classify arcs according to whether used or not.
Step 2: Calculation of UCR-Sjk.
4.2.5. Tabu procedure
Step 2.1 Calculate UCR-Sjk for each used arc.
Tabu search procedure is iterated with the tabu attributes and
Step 2.2 Sort used arc in decreasing order according to UCR-
the tabu list. In a DC exchange method, two kinds of tabu list are
Sjk.
operated; one is for the distribution center to be deleted, the other
Step 3: Arc exchange.
for the distribution center to be joined. Whenever the distribution
Step 3.1 Select used arcs sequentially according to sorted
center is selected to be deleted or to be joined in the neighbor
list.
search, the corresponding tabu list is checked if the decision is
Step 3.1.1 Select another used arcs sequentially among
proper. In an arc exchange, tabu list is used to discourage selecting
remained used arc sets.
arcs that are exchanged quite recently. The tabu list for the arc ex-
Step 3.1.2 Assuming that selected two used arcs are
change consists of the pair of the arcs which are interchanged re-
exchanged, calculate UCR-Sjk for each exchanged arc.
cently. In an overall heuristic procedure, tabu list and attributes
Step 3.1.3 Check feasibility by comparing capacities of
are applied to each neighborhood generation step for demand
distribution centers that are used by selected arcs.
decomposition.
If it is feasible, then go to Step 3.1.4.
Else, go to Step 3.1.1 and repeat steps.
4.3. Neighborhood generation: supply decomposition
Step 3.1.4 If the sum of UCR-Sjk that is calculated in Step
3.1.2 is smaller than sum of UCR-Sjk that is calculated in
In the demand decomposition the decisions are made on the
Step 2, then exchange two arcs and stop procedure.
selection of the distribution centers (ZDj), and the distribution
Else, go to Step 3.1.1 and repeat steps.
plans (Yjkl, ZPjl, ZSjk). The remaining decisions to be made in the
Step 3.2 If there are used arcs not selected before, then go to
supply decomposition are on the transportation quantity from
Step 3.1.3.
the plant to the open distribution center for each product. The

Fig. 3. Neighborhood generation by arc exchange method.

Fig. 4. Neighborhood generation by arc add method.


3100 Y.H. Lee, S.G. Kwon / Expert Systems with Applications 37 (2010) 3094–3103

solution in the supply decomposition is found optimally using The objective is to minimize the sum of transportation cost in-
CPLEX since the problem becomes the simpler linear programming curred for the logistics between the plants and the distribution
P
model. Let DMDjl ¼ k Y jkl , the demand quantity to be transported center. Constraints (15) and (16) are on the distribution capacity
through the distribution center j of the product l, which can be and on the logistics balance.
computed from the solution obtained in the demand decomposi- After the solution of the supply decomposition is found, the
tion. The mathematical model to be solved in the supply decompo- complete solution combined with the demand decomposition is
sition is as follows: specified. Compared with the current best at the iteration, the
search process is repeated until the termination rule is applied.
4.3.1. Formulation
XXX
Min TPDij X ijl ð14Þ 5. Computational experiments and results
i j l
XX
s:t: X ijl 6 CPi 8i ð15Þ Several instances were generated for evaluation with respect to
j l the number of plants, distribution centers, customers, products,
X
X ijl ¼ DMDjl 8j; l ð16Þ and are shown in Table 2. Small size problems are designed so that
i the performances of the heuristics can be evaluated with the opti-
X ijl P 0 8i; j; l ð17Þ mal solutions. Large size problems are also designed to compare
the performances with the heuristics’. Although the optimal solu-
tions are not available within the 1800 s of computational time,

Table 2
Experimental design for various cases.

Case Small size example Large size example


Plant Distribution center Customer Product Plant Distribution center Customer Product
1 3 5 10 3 7 15 40 3
2 3 5 10 4 7 15 40 4
3 3 5 10 5 7 15 40 5
4 3 7 15 3 7 15 50 3
5 3 7 15 4 7 15 50 4
6 3 7 15 5 7 15 50 5
7 5 10 20 3 10 20 60 3
8 5 10 20 4 10 20 60 4
9 5 10 20 5 10 20 60 5
10 5 10 30 3 10 20 80 3
11 5 10 30 4 10 20 80 4
12 5 10 30 5 10 20 80 5

Table 3
Comparison of performances for the heuristics to the optimal solution for the small size instances.

Case Type A Type B Type C Type D


Optimal Heuristic Gap (%) Optimal Heuristic Gap (%) Optimal Heuristic Gap (%) Optimal Heuristic Gap (%)
1 5434 5519 1.56 6969 7100 1.87 4856 4856 0.00 7804 7804 0.00
(1.48) (5.45) (1.06) (2.70) (1.20) (2.06) (0.66) (1.63)
2 6105 6105 0.00 8058 8058 0.00 5610 5610 0.00 7913 8045 1.67
(0.67) (1.74) (0.89) (3.64) (1.03) (1.26) (1.45) (1.25)
3 4449 4467 0.40 8135 8135 0.00 7195 7440 3.40 7161 7425 3.68
(0.55) (0.55) (0.59) (0.63) (3.03) (2.47) (0.61) (1.20)
4 6185 6185 0.00 8897 9188 3.27 7592 8040 5.90 9341 9341 0.00
(9.08) (2.19) (0.81) (3.97) (14.67) (1.64) (14.86) (3.31)
5 6401 6727 5.09 10117 10312 1.92 7391 7426 0.47 9358 9441 0.88
(163.25) (1.49) (37.13) (2.84) (61.81) (5.77) (3.31) (1.70)
6 6299 6481 2.88 8342 8391 0.58 9913 10028 1.16 8827 8863 0.40
(21.05) (2.13) (2.16) (3.33) (34.00) (1.66) (3.27) (1.30)
7 6684 6956 4.06 10254 10705 4.39 9175 9400 2.45 11043 11498 4.12
(20.06) (9.17) (825.64) (7.11) (78.02) (18.22) (10.28) (8.36)
8 8643 8900 2.97 12017 12283 2.21 9282 9696 4.46 10551 10761 1.99
(236.55) (13.06) (180.63) (4.26) (414.51) (13.96) (29.95) (2.97)
9 7986 8284 3.73 10722 10948 2.10 9072 9454 4.21 9130 9333 2.22
(280.38) (15.84) (90.56) (6.14) (807.99) (17.72) (91.78) (7.05)
10 13114 14716 12.21 14722 15843 7.61 12147 1323 9.92 18545 19383 4.51
(404.25) (22.69) (83.52) (10.17) (319.31) (51.27) (88.36) (25.14)
11 10400 10762 3.48 17554 17998 2.52 12940 13773 6.43 18263 18461 1.08
(310.23) (22.80) (667.89) (20.61) (*) (22.41) (96.58) (5.81)
12 14101 15552 10.31 14882 15628 5.01 20344 22254 9.38 17720 18270 3.10
(355.92) (29.98) (64.45) (9.89) (*) (40.72) (62.06) (30.95)
Average 3.89 2.62 3.98 1.97

(1) The computation time taken in seconds is shown in the parentheses.


(2) (*) indicates that the best solution found in the 1800 s time limit, while the optimal solution is not available.
(3) Gap(%) = (the heuristics value  CPLEX value)/CPLEX value  100.
Y.H. Lee, S.G. Kwon / Expert Systems with Applications 37 (2010) 3094–3103 3101

the best feasible solution found in the time limit given, is com- tice that two cases do not give an optimal solution within the com-
pared with the one obtained through the heuristics. The number putational time limit even for small size problems. In Table 4,
of facilities and the products are set for the small size problem so performances of the heuristics are compared with the best solution
that the number of constraints is at most 1750, and the number found in time limit using CPLEX for the large size problem.
of the real(integer) variables is at most 1750(360), respectively in The performances of the heuristics are 3.11% away in average
terms of the mathematical modeling discussed in Section 3. It from the optimal solution for the small size instances, and 3.95%
was expected to give an optimal solution within the 1800 s compu- away in average from the best solution for the large size instances.
tational time limit, but a couple of cases failed to give an optimal The worst case in the small size instances is 12.21% away from the
solution in the time given. Large size problems are designed in con- optimal solution in experiment, and the worst case in the large size
sideration of the actual practices with a case of at most 8600 con- instance is 15.06% away from the best solution. It is noticeable that
straints, 9000 real variables, and 1750 integer variables. the heuristic solution found in several tens of seconds happens to
Various instances in terms of the cost structure are also de- be better than the best solution found in 1800 s using CPLEX for the
signed to investigate how the performance of the heuristic works large size instances although it is rare (type A cases 8, 12, type B
for different cases. Instances with 4 types of cost structure are gen- cases 3,6,10). The computational time taken to get the solution
erated: type A instances have a similar weight on the data for the for heuristics is at most 260 s even for the large size instances,
setup cost, product handling cost and the transportation cost, which can be considered reasonable to be implemented in the ac-
which is randomly generated in the range [2, 10] as a standard; tual practices. As expected, the time required to get the optimal
type B instances have double the weight on the setup cost; type solution is increasing sharply as the instance size increases, while
C instances on the product handling cost; and type D instances the time needed to run the heuristics are linearly increasing.
on the transportation cost, respectively. The suggested heuristics The sensitivity by the cost structure can be observed from the
are programmed in MS C++ and CPLEX 7.0 library, and the solution results. Performances from the heuristics are not likely to be signif-
obtained by the suggested heuristic is compared with the optimal icant due to the types of cost structure although type C instances
solution or the best feasible solution obtained by ILOG CPLEX with have the most deviation from the optimal or the best solutions in
computational time limit of 30 min. All experiments were per- both the small and large size instances. The computational time
formed on a Pentium IV PC with 3.0GHz and 1GB RAM. In a sug- needed to run heuristics also does not give any significant differ-
gested heuristic, every iteration of tabu search is repeated until ences by the types.
solution is not improved consecutively 10 times for each neighbor-
hood generation procedure during 1000 iterations. 6. Conclusion and discussions
The size of the tabu list used in the DC change is 5, and the size
for the arc change is 20, since it is observed that they are proper In this paper, the distribution center operation planning model
and do not give a critical impact on the result through the experi- with multi product and multi-stage is studied under single sourc-
mentation. With a different seed of random number, experiments ing constraint and product handling cost structure. The objective is
are repeated 10 times and the best solution value is found. Com- to minimize total logistic costs that consist of transportation cost,
parison of the performances between the optimal solutions and facility setup cost and product handling cost. In the distribution
the heuristics for the small size instances is shown in Table 3. No- center operations the location selections and the product selec-

Table 4
Comparison of performances for the heuristics to the best solution found for the large size instances.

Case Type A Type B Type C Type D


Best Heuristic Gap (%) Best Heuristic Gap (%) Best Heuristic Gap (%) Best Heuristic Gap (%)
1 17,634 18,104 2.66 17,308 18,034 4.19 19,869 21,686 9.14 25,679 27,287 6.26
(27.30) (18.13) (21.98) (9.45)
2 13,882 14089 1.49 19,082 20,400 6.90 15,942 16,313 2.32 23,958 24,189 0.96
(30.08) (21.39) (23.86) (8.94)
3 15,012 15575 3.75 17,986 17,562 2.35 14,773 15,241 3.16 26,369 26,426 0.21
(16.58) (29.92) (25.86) (9.08)
4 19,032 19660 3.29 24,869 26,467 6.42 25,264 27,298 8.05 28,127 29,326 4.26
(89.97) (45.89) (27.95) (42.56)
5 20,083 21133 5.22 25,259 25,981 2.85 22,046 23,221 5.32 24,336 26,059 7.08
(36.94) (72.20) (42.38) (58.48)
6 19,920 20740 4.11 28,715 27,424 4.49 21,484 24,141 12.36 28,964 30,381 4.89
(61.13) (24.39) (27.33) (39.86)
7 24,715 25979 5.11 33,479 34,917 4.29 26,085 27,360 4.88 29,778 31,576 6.03
(99.84) (60.16) (225.69) (66.11)
8 24,798 24041 3.05 33,379 34,762 4.14 31,582 31,877 0.93 32,261 33,241 3.03
(39.92) (56.94) (62.61) (99.34)
9 23,462 24589 4.80 27,042 27,905 3.19 27,977 30,603 9.38 29,487 30,991 5.10
(117.95) (110.28) (188.28) (83.84)
10 28,083 29757 5.96 51,919 51,454 0.89 36,334 41,807 15.06 43,224 44,295 2.47
(129.09) (93.95) (131.01) (201.73)
11 29,265 31724 8.40 40,393 42,018 4.02 35,338 35,379 0.11 44,575 46,008 3.21
(86.74) (95.56) (178.14) (90.80)
12 36,084 35500 1.61 43,577 45,749 4.98 35,353 35,625 0.76 40,999 41,441 1.07
(86.56) (260.50) (118.56) (136.80)
Average 3.35 2.77 5.96 3.71

(1) The computation time taken in seconds is shown in the parentheses.


(2) The best solution found in the 1800 s computational time limit.
(3) Gap(%) = (the heuristics value  CPLEX value)/CPLEX value  100.
3102 Y.H. Lee, S.G. Kwon / Expert Systems with Applications 37 (2010) 3094–3103

tions to handle are to be determined in consideration of the cost. the lead time required to delivery from the order placement be-
The distribution center planning becomes operational, in other comes the core competitiveness in the electronic industry. Deci-
words, its decisions are to be made on monthly base or at most sions on the selection of the distribution centers in operation can
quarterly base, rather than strategic one which are performed on be made frequently, for example, monthly or at most quarterly.
yearly base or loner time periods. The suggested heuristics would The priority rule suggested can be used in many ways: while
be practical in the sense that the computational time needed to the distribution centers are operated according to the solution gi-
get the solutions are at most several minutes, and the priority rules ven, many other minor unexpected accidents have occurred. As a
used in the heuristics seems to be easily understandable, especially manager of the distribution center, the adapted solution using
for the practitioners. The priority rules are designed greedily using the priority rule can be available to meet the minor changes in
the nearby information such as the setup cost of its own distribu- the distribution center, or in the transportation arc, rather than
tion center, or the transportation operation cost of currently han- the full running of the whole supply chain model. The priority in-
dling product, rather than the global optimized allocation. But dex rule can be enhanced depending on the cost structure and
these myopic approaches may give a good solution with a merit importance of the factors considered. To our knowledge, tabu
of computational efficiency. Most priority rules from managers’ search among the meta-heuristics seems to fit to this model since
experiences and reasoning are used in actual practices, and its the priority rule used has a tendency of the staying near from the
implementations are much getting easier since managerial deci- current solution in developing the neighbor, hence getting out of
sions from the model can get the consensus from the practitioners. the local optimum is the critical way to reach the better solution
The model studied in this paper can be extended to the multi area. Some other meta-heuristics are under study for the various
level supply chain model, and to the multi-period model. The mod- supply chain networks.
el with more 3 levels are not much existent in the real business,
but in the supply chain the 4 or 5 level networks may be included
as a portion, in which this approach can be applied easily in the le- References
vel by level manner. The multi-period model has a couple of diffi-
culties: the computational heaviness and the inaccurate data. Aikens, C. H. (1985). Facility location models for distribution planning. European
Computational tediousness can be handled by using the hardware Journal of Operational Research, 22, 263–279.
Akinc, U., & Khumawala, B. M. (1977). An efficient branch and bound algorithm for
upgrade or the less number of tabu iterations with a tradeoff of the capacitated warehouse location problem. Management Science, 23, 585–594.
loss of the solution quality. More typical problems are in the data Altiparmak, F., Gen, M., & Lin, L. (2006). A genetic algorithm approach for multi-
such as the future demand forecasting. For the multi-period model objective optimization of supply chain networks. Computers & Industrial
Engineering, 51, 196–215.
the cost data and the demand data for the future periods needs to
Amiri, A. (2006). Designing a distribution network in a supply chain system:
be provided in advance, but they are obviously inaccurate. The Formulation and efficient solution procedure. European Journal of Operational
usability of the multi-period model is not verified still, not because Research, 171, 567–576.
of the computational ability or the quality of the solution, but be- Atkins, R. J., & Shriver, R. H. (1968). New approach to facility location. Harvard
Business Review, 70–79.
cause of the data availability. Avella, P., Benati, S., Cánovas Martinez, L., Dalby, K., Di Girolamo, D., Dimitrijevic, B.,,
The companies are facing the global market competition in et al. & Hultberg, T. H. (1998). Some personal views on the current state and the
which meeting the demand in time and reducing the cost incorpo- future of locational analysis. European Journal of Operational Research, 104,
269–287.
rated in management of the inventory and the stock facility are Bhattacharya, A., Sarkar, B., & Mukherjee, S. K. (2004). A new method for plant
critical competitiveness. In many cases the manufacturing plants location selection: A holistic approach. International Journal of Industrial
are located in the place the technology and the working resources Engineering – Theory, Applications and Practice, 11(4), 330–338.
Baumol, W. J., & Wolfe, P. (1958). A warehouse location problem. Operations
are available, and the distribution warehouse are located in an- Research, 6, 643–666.
other place close to the market and customers. Because of the Brandeau, M. L., & Chiu, S. S. (1989). An overview of representative problems in
changes in the demand of each markets the distribution centers location research. Management Science, 35, 645–674.
Chandra, P., & Fisher, M. L. (1994). Coordination of production and distribution
are need to meet the changing situation. The applications sug- planning. European Journal of Operational Research, 72, 503–517.
gested in this paper may fit to the manufacturing companies which Cohen, M. A., Lee, H. L., Moon, S. (1987). An integrated model for manufacturing and
operates several distribution centers near to customers: electronics distribution systems design. Working paper, 87-07-03, Department of Decision
Sciences, Wharton School, University of Pennsylvania.
component makers, electrical assemblers. Internet shopping com-
Efroymson, M. A., & Ray, T. L. (1966). Branch and bound algorithm for plant location.
panies or small parcel delivery services operating several subsidi- Operations Research, 14(3), 361–368.
ary facilities may gain benefits from this approaches, especially Erenguc, S. S., Simpson, N. C., & Vakharia, A. J. (1999). Integrated production/
in which the demands are fluctuating in time and in locations. distribution planning in supply chains: An invited review. European Journal of
Operational Research, 115, 219–236.
The practical decisions can be made in general for the long term Gen, M., & Syarif, A. (2005). Hybrid genetic algorithm for multi-time period
planning, for example, yearly or bi-yearly, since they may involve production/distribution planning. Computers & Industrial Engineering, 48,
the installations of the facilities such as the plants and distribu- 799–809.
Geoffrion, A. M., & Graves, G. W. (1974). Multicommodity distribution system design
tions centers. Hence the computation time needed to get the solu- by Benders decomposition. Management Science, 20(5), 822–844.
tions does not matter as long as the proper decisions are made, Jayaraman, V., & Pirkul, H. (2001). Planning and coordination of production and
where the optimized approach based on the mathematical pro- distribution facilities for multiple commodity. European Journal of Operational
Research, 133, 394–408.
gramming are preferred with the help of the commercial SW. But Jararaman, V., & Ross, A. (2003). A simulated annealing methodology to distribution
the market requirements on this industry have changed, and more network design and management. European Journal of Operational Research, 144,
frequent decisions are necessary to meet the dynamically changing 629–645.
Khumawala, B. M. (1972). An efficient branch and bound algorithm for the
demands. Most manufacturers want to keep their own plants be- warehouse location problem. Management Science, 18, 718–731.
cause of the technology intellectual properties, while the install- Ko, D. S., Yang, Y. C., Jang, Y. J., & Park, J. W. (2000). A study on integrated production
ment of them may require the long time based and heavy capital planning of distributed manufacturing system on supply chain. IE Interfaces, 13.
Kim, J., & Kim, Y. (1999). A decomposition approach to a multi-period vehicle
investments. The fluctuating demands due to the PC makers’ and
scheduling problem. The International Journal of Management Science, 27.
cellular phone assemblers’ orders have to be met from the products Klingman, D., Randolph, P. H., & Fuller, S. W. (1976). A cotton ginning problem.
through an owned distribution centers and an outsourced one as Operations Research, 24, 700–717.
well since the distribution centers can be built and closed easily Kuehn, A. A., & Hamburger, M. J. (1963). A heuristic program for locating
warehouses. Management Science, 9, 643–666.
with relatively small amount of setup costs. Another reason to give Lee, H., & Billington, C. (1993). Material management in decentralized supply
more flexibility on the decisions for the distribution centers is that chains. Operational Research, 41(5), 835–847.
Y.H. Lee, S.G. Kwon / Expert Systems with Applications 37 (2010) 3094–3103 3103

Lee, B. K., Lee, Y. H. (2002). Heuristic method to minimize logistics cost in supply Syarif, A., & Gen, M. (2002). Solving exclusionary side constrained transportation
chain. In Proceeding of the 2002 spring congress of Korean society of supply chain problem by using a hybrid spanning tree-based genetic algorithm. Journal of
management. Intelligent Manufacturing, 14, 389–399.
Manzini, R., Gamberi, M., & Regattieri, A. (2006). Applying mixed integer Syarif, A., Yun, Y. S., & Gen, M. (2002). Study on multi-stage logistic chain network:
programming to the design of distribution logistic network. International A spanning tree-based genetic algorithm approach. Computers & Industrial
Journal of industrial Engineering – Theory, Applications and practice, 13(2), 207–218. Engineering, 43, 299–314.
Oliver, R. K., Webber, M. D. (1982). Supply-chain management: Logistics catches up Thomas, D. J., & Griffin, P. M. (1996). Coordinated supply chain management.
with strategy (reprint from Outlook 1982). In: M. Christopher (Ed.), Logistics-the European Journal of Operational Research, 94, 1–115.
strategic issues (pp. 63–75). Venkateswaran, J., & Son, Y.-J. (2004). Design and development of a prototype
Pirkul, H., & Jayaraman, V. (1998). A multi-commodity, multi-plant, capacitated distributed simulation for evaluation of supply chains. International Journal
facility location problem: Formulation and efficient heuristic solution. of Industrial Engineering – Theory, Applications and Practice, 11(2), 151–
Computers & Operations Research, 25(10), 869–878. 159.
Ross, A. D. (2000). A two-phased approach to the supply network reconfiguration Vergana, F. E., Khouja, M., & Michalewicz, Z. (2002). An evolutionary algorithm for
problem. European Journal of Operational Research, 122, 18–30. optimizing material flow in supply chains. Computers and Industrial Engineering,
Sabri, E. H., & Beamon, B. M. (2000). A multi-objective approach to simultaneously 43, 407–421.
strategic and operational planning in supply chain design. Omega, 28, 581–598.

Potrebbero piacerti anche