011-0161 An Integrated Inventory Control and Facility Location System with Capacity Constraints: A Multi-Objective Evolutionary Approach

Shu-Hsien Liao1, Chia-Lin Hsieh2,a

1

Processor of Department of Management Sciences and Decision Making, Tamkang University, No. 151, Yingjuan Road, Danshuei Jen, Taipei 251, Taiwan, ROC E-mail: [email protected]

Lecturer of Department of Industrial Management and Enterprise Information, Aletheia University, No. 26, Chenli Street, Danshuei Jen, Taipei 251, Taiwan, ROC E-mail: [email protected]

POMS 20th Annual Conference Orlando, Florida U.S.A. May 1 to May 4, 2009 Abstract A supply chain network system is to provide an optimal platform for efficient and effective supply chain management (SCM). SCM usually involves multiple and conflicting objectives such as cost, customer service levels (volume fill rate), and flexibility (responsiveness). In this research, a multi-objective capacitated location-inventory distribution network system is formulated which integrates the effects of facility location, distribution, and inventory issues and includes conflicting objectives such as cost, customer service level (order fill rate) and flexibility (responsiveness level). This model allows
a

Also PHD research student, Department of Management Sciences and Decision Making, Tamkang University, Taiwan, ROC

determining the optimal locations of distribution centers (DCs) and the assignment of buyers to DCs to find the set of Pareto optimal solutions. The study highlights the possibility of a hybrid evolutionary approach based on the elitist NSGA-II algorithm in understanding seemingly non-intuitive model performance. Sensitivity analysis is investigated to understand the model performance and to illustrate how parameter changes influence its output. Keywords: Integrated supply chain design, Multiobjective evolutionary algorithm, Tradeoff analysis. 1. Introduction Businesses are undergoing a revolution in terms of implementing new strategies and technologies in response to the challenges and customer demands. Recently, two generic strategies for supply chain design emerged: efficiency and responsiveness. Efficiency aims to reduce operational costs; responsiveness, on the other hand, is designed to react quickly to satisfy customer demands and save costs. In traditional distribution systems, minimizing costs or maximizing profits as a single objective is often the focus. However, very few distribution network systems are single objective problems. Multi-objective formulation has to be considered whose solutions will be a set of Pareto alternatives representing the tradeoffs among different objectives. Recently, Daskin et al. (2002) introduced a joint location-inventory model with risk pooling (LMRP) that incorporates inventory costs at the distribution centers (DCs) into the

location problem. LMRP assumes uncapacitated DCs which is usually not practical. Capacity limitation may affect the number and locations of the facilities, the inventory that can be stored at the facilities and consequently the order frequency as well as the assignment of buyers to the facilities. Our model builds upon the initial LMRP model with some differences. First, a capacitated version of the similar model is established. Second, customer responsiveness and volume fill rate are incorporated as two extra performance metrics to make our contribution. With these considerations, we present a capacitated Multi-Objective Location-Inventory Problem (MOLIP) which results in a Mixed-Integer Non-Linear Programming (MINLP) formulation. Evolutionary optimization algorithms are known to be efficient-solving and easy-adaptive, especially those where traditional methods failed to provide good solutions (e.g. MINLP). Recently, multiobjective evolutionary algorithms (MOEAs) have become prevailing since the pioneering work by Schaffer (1985). There are many efficient MOEAs that are possible to find Pareto optimal solutions as well as widely distributed sets of solutions, NSGA-II (Deb et al.,2002) is one of the most successful approaches. In our study, the well-known NSGA-II algorithm and a heuristic assignment procedure are incorporated that help to approximate the Pareto frontier for optimizing MOLIP. The paper is organized as follows. Section 2 presents an overview on location inventory problems. Section 3 details the problem formulation. Section 4 proposes a hybrid genetic

algorithm with a heuristic solution procedure for MOLIP. Section 5 illustrates and analyzes computational results using a base-line model instance and finally, conclusions with some directions are provided in section 6. 2. Literature Review Research on integrated location-inventory distribution network systems is relatively new. Nozick and Turnquist (2001) proposed a joint location-inventory model to consider both cost and service responsiveness trade-offs based on the uncapacitated facility location problem. The analysis demonstrated an approximate linear safety-stock cost function in the framework and proposed a Lagrangean-based scheme. Miranda and Garrido (2004) studied an MINLP model to incorporate inventory decisions into typical facility location models to solve the distribution network problem by incorporating a stochastic demand and the risk pooling phenomenon. A heuristic solution approach, based on Lagrangian relaxation and the sub-gradient method was presented. Sabri and Beamon (2000) presented an integrated multi-objective multi-echelon stochastic model that simultaneously addresses strategic and operational planning decisions by developing an integrated two sub-module model. Similarly, Gaur and Ravindran (2006) studied a bi-criteria optimization model to represent the inventory aggregation problem under risk pooling to find out the tradeoffs in costs and responsiveness. Daskin et al. (2002) and Shen et al. (2003) present a LMRP model that incorporated safety stock placement into a location problem for a two-stage network. There are several

variations of the LMRP model. Ozsen (2004) presents a capacitated version of LMRP which determines the ordering policy at the DCs so that the inventory aggregation does not exceed DC capacities. A Lagrangian relaxation algorithm was applied to solve this problem. Shen & Daskin (2005) extended LMRP to include the customer service component and proposed a nonlinear multi-objective model including both the cost and service objectives. They developed a weighting method and an efficient GA-based heuristic solution approach for quick and meaningful evaluation of cost/service trade-offs. From the survey, some innovative research aspects that are noteworthy have been incorporated in our research work as follows: Multi-objective location inventory problem. Very few researches have addressed this problem. A multiobjective formulation should be required to provide the tradeoffs of Pareto optimal alternatives among total costs and customer service. Multi-objective evolutionary algorithms (MOEAs). Most reviewed research works focused on traditional optimization techniques but few have performed successfully and efficiently. In the contrast, MOEAs have been successful developed for various optimization problems and enable the possibility for the proposed MOLIP. 3. Mathematical Formulation 3.1. Problem Description and Model Assumption Suppliers and distributors in general, route their products either through DCs. Consider the problem of configuring a supply chain distribution network system, where a single

supplier and a set of DCs are established and dispersed in a region to distribute various products to a set of buyers. DCs act as intermediate facilities between the supplier and the buyers and facilitate the product shipment between the two echelons, as depicted in Figure 1. The supplier wishes to determine the subsets of DCs to be opened and design the distribution strategy that will satisfy all DC capacities and buyers demands for the products. The problem is a single supplier, multiple-products, two-echelon supply chain network design problem.

Figure 1 Two-echelon distribution network problem General basic assumptions are used when modeling our problem. It is assumed that all the products are produced by a single supplier and one specific product for a buyer should be shipped from a single DC. There are no reverse flows and no in-transit inventory, or pipeline inventory is considered. All demands of the buyers are uncertain and the storage capacities at the supplier are unlimited but capacitated at open DCs. Further assumptions will be stated when we illustrate the mathematical models. 3.2. Mathematical Models Our problem is formulated as a multi-objective mixed-integer non-linear programming

model so as to explore the tradeoff between facility location, distribution, and inventory costs on one hand and the level of customer service on the other. The mathematical notation and formulation are as follows: Indices: i is an index set for buyers ( i I ). j is an index set of potential DCs ( j J ).
k is an index set for product classifications ( k K ).

Decision Variables: Q k is the aggregate economic order quantity for DC j for product wi k shipped from the supplier.
1, Yj = 0, if DC j is open , otherwise

1, if DC j serves buyer i for shipping product k X kji = otherwise 0,

Model Parameters: u j is the capacity of DC j. d ik is the mean demand rate (daily) for product k at buyer i. jk is the average lead time (daily) for product k to be shipped to DC j from the supplier. ik is the standard deviation of daily demands for product k at buyer i. is the number of working days per year. D k is the total expected annual demand for product wj k going through DC j. f j is the fixed annual facility operating cost of locating at DC site j.
h k is the aggregated inventory unit holding cost at DC j for product k. o k is the ordering j j

cost at DC j for product k per order (including fixed transportation cost)). tc k is the unit ji variable transportation cost for shipping product k from DC j to buyer i. rc k is the unit j variable production and transportation cost for shipping product k from the supplier to DC j.
h ji is the distance between DC j and buyer i. D max is the maximal covering distance, that is,

buyers

within

this

distance

to

an

open

DC

are

considered

well

satisfied.

 i = {j J | h ji D max

is the set of DCs that could attend buyer i within D max .

To begin modeling this problem, we also assume that the daily demand for product k at
2 each buyer i is independent and normally distributed, i.e. N( dik , ik ). Furthermore, at any site

of DC j, we assume a continuous review inventory policy (Qj, rj) to meet a stochastic demand pattern. Also, we consider that the supplier takes an average lead time jk (in days) for

shipping product k from the supplier to DC j so as to fulfill an order. From basic inventory theory (Eppen, 1979) considering centralized inventory system (e.g. VMI), if the demands

at each buyer are uncorrelated, then the aggregate demand of product k during lead time jk
k k at the DC j is normally distributed with a mean of jk d wj , where d wj = dik X ji , and a variance
iI

of

k j

i I

2 ik

ji

, the total amount of safety stock for product k at any DC j with risk

k 2 pooling is z1 j ik X ji where 1- is referred to the level of service for the system and iI

z1 is the standard normal value with P ( z z 1 ) = 1 . In our proposed model, the total cost of the system can be decomposed into the following items: (i) facility cost (FC), which is the cost of setting up the DCs, (ii) transportation cost (TC), which is the cost of transporting products from suppliers to the demand regions via specific DCs, (iii) operating cost (OC), which is the cost of running the DCs, (iv) cycle stock cost (CC), which is the cost of maintaining working inventory at the DCs, and (v) safety stock cost (SC), which is the cost of holding sufficient inventory at DCs in order to provide specific service level to their buyers. Hence, it can be represented as total cost function Z1 as follows.

Z 1 = f j Yj +
jJ

(rc + tc ) dik X
kK jJ iI k j k ji

k ji

kK jJ

k j

k dwj k Qwj

+ h
k K jJ k j

k Q wj

Y jk + h k z1 j
kK jJ

k j

iI

2 ik

Xk ji

(1)

k Based on Z1, the optimal order quantity Q wj* for product k at each DC j can be obtained

k through differentiating eq. (1) in terms of Qwj , for each DC j and each product k, and equaling

k to zero to minimize the total supply chain cost. We can obtain Q wj =

k 2 o jk D wj h k j

for

open DC j, k.

k In this case, there is not any capacity constraint for the order quantities Qwj

k since we assume the storage capacity at the supplier is unlimited. Thus, replacing Q wj* in the

third and fourth terms of Z1 in eq. (1), we can obtain a non-linear cost function of Z1. In the following, we propose an innovative mathematical model for the Multi-Objective Location-Inventory Problem (MOLIP). Min Z1 =

kK jJ

f
+

Y jk +
kK jJ
k j

iI
k ji

k ji

Xk + ji

kK

D
j J

k j

i I

ik

k ji

k K j J

i I

k ji

(2) (3)

Max

Z 2 = d ik X k ji kK iI
Z 3 = d ik X kK iI j i
k ji

kK iI

ik

Max s.t.

k K i I

ik

Xk ji

(4) (5) (6) (7) (8)

X
jJ

k ji

i I ; k K i I ; j J ; k K

Xk Yj ji

kK iI

ik

Xk + ji
kK

iI

k ji

X k u j Yj ji
j J ;

j J

X k {0,1}, ji

i I ;

k K

Y j {0,1} ,

j J

(9)

where

k ji = (rc k + tc k ) d ik j ji

jk = 2 o k h k j j
2 kji = ( z1 ) 2 jk ik

Dik = d ik

Eqs. (2)-(4) gives the objectives. While eq. (2) of Z1 is to minimize the total cost, eq. (3) of Z2 and eq. (4) of Z3 give the objectives referred to maximizing customer service by two performance measurements: (i) volume fill rate (VFR), defined as the satisfied fraction of total demands without shortage; (ii) responsiveness level (RL), the percentage of fulfilled demand volume within specified coverage distance Dmax. Eq. (5) restricts a buyer to be served by a single DC if possible. Eq. (6) stipulates that buyers can only be assigned to open DCs. Eq. (7) are the maximal capacity restrictions on the opened DCs to enable the capability of holding sufficient inventory for every product that flows through the DC, and also the part of safety stock so as to maintain the specified service level. Eq. (8)-(9) are binary constraints. The proposed MOLIP model would not only determine the DC locations, the assignment of buyers to DCs, but also find out endogenously both the optimal order quantities and safety-stock levels at DCs. Since two of the three objective functions (Z1 and Z3) are nonlinear, the formulation results in an intractable multi-objective MINLP model. 4. Problem Solving Methodology 4.1 Background Multiobjective optimization problems give rise to a set of Pareto-optimal solutions, none

of which can be said to be better than other in all objectives. Unlike most traditional optimization approaches, evolutionary algorithms (EAs) work with a population of solutions and thus are likely candidates for finding multiple Pareto-optimal solutions simultaneously (Michalewicz 1996; Coello Coello 1999). There are primarily two goals tasks to be achieved in a multiobjective EA (MOEA): (i) convergence to the Pareto-optimal set, and (ii) maintenance of population diversity in the Pareto-optimal set. Most MOEAs work with the domination concept. For a problem having more than one objective function (say fj, j =1,..,K and K > 1), any two solutions X1 and X2 have one of two possibilities, one dominates the other or none dominates the other. One solution X1 is said to dominate the other solution X2, if both the following condition are true: (i) the solution X1 is not worse than X2 in all objectives; (ii) the solution X1 is strictly better than X2 in at least one objective. If any of the above conditions is violated, the solution X1 does not dominate X2. There are many MOEAs allowing the attainment of these results, but the NSGA-II algorithm (Deb et al., 2002) has proved to be quite efficient in many different applications. NSGA-II is an improved version of a search algorithm originally proposed by Srinvas and Deb (1994). With respect to its predecessor, the new algorithm uses an improved multi-objective non-dominated sorting method that requires a significantly smaller number of comparisons (O(LK2) with respect to the O(LK3) of conventional non-dominated sorting, where K is the number of objectives, and L is the population size).

For each solution, one has to determine how many solutions dominate it and the set of solutions to which it dominates. Thus, it ranks all solutions to form non-dominated fronts according to a non-dominated sorting process to classify the chromosomes into several fronts of nondominated solutions. To allow the diversification, NSGA-II also estimates the solution density surrounding a particular solution in the population by computing a crowding distance operator. During selection, a crowded-comparison operator considering both the non-domination rank of an individual and its crowding distance is used to select the offspring without lost good solutions (elitism), whereas crossover and mutation operators remain as usual. We summarized the NSGAII algorithm as shown in Figure 2.

Figure 2 Graphical representation of the NSGA-II algorithm 4.2 NAGAII-based Evolutionary Approach A NSGAII-based NAGAII-based Evolutionary Approach algorithm for MOLIP is implemented in this study and is detailed in Table 1. The algorithm starts by generating a random population P(1) of size L. For each chromosome in P(1) the algorithm evaluates its cost and coverage using the encoded solution expressions. Then, the algorithm applies

non-dominated sorting on P(1) and assigns to each chromosome the front to which it belongs. Next, the algorithm applies binary tournament selection (to form the crossover pool), crossover, and mutation operators to generate the children population C(1) of size L. Once initialized, the main body of the algorithm repeats for T generations. The algorithm applies non-dominated sorting to R(t), the resulting population from the union of parents P(t) and children C(t). The algorithm obtains the next generation population P(t+1) after selecting the L chromosomes from the first fronts of R(t). Next, it applies binary tournament selection, crossover, and mutation operators to generate the children C(t+1). Table 1 NAGAII-based algorithm for MOLIP.

1: Randomly generate P(1) 2: Evaluate P(1) 3: Nondominated sort P(1) 4: Generate C(1) form P(1), apply binary tournament selection, crossover, and mutation. 5: Evaluate C(1) 6: while t T do 7: R(t) =P(t) C(t) 8: Nondominated sort R(t) 9: Sort R(t) using n {see Definition 1 below} 10: Select P(t+1) from the first L chromosome of R(t) 11: Generate C(t+1) from P(t+1), apply binary tournament selection, crossover, and mutation 12: Mutate C(t+1) 13: Evaluate C(t+1) 14: t t + 1 15: end while

In the selection process of the next generation, chromosome fitness depends on the evaluation of the decoded solution in the objective functions and its comparison with other chromosomes. The non-domination sorting updates a tentative set of Pareto optimal solutions by ranking a population according to non-domination. After that, each individual p in the population is given two attributes: (i) non-domination rank in the optimization objectives

(p.rank); (ii) local crowding distance in the objectives space directions (p.distance). If both chromosomes are at the same non-domination rank, the one with fewer chromosomes around in the front is preferred. Thus, a partial order (n ) defined in Definition 1 is used to guide the selection process of the algorithms to decide among two chromosomes which one is fitter. Definition 1. Let p, qR(t) be chromosomes in population R(t). p is said better fitted than q (p n q), either if (p.rank < q.rank) or ((p.rank = q.rank) and (p.distance > q.distance)). Definition 1 states that a higher non-domination level chromosome is always preferred. If chromosomes are at the same level, the one with fewer chromosomes around is preferred. 4.3 Solution Encoding Each solution of MOLIP is encoded in a binary string of length m = |I|, where the j-th position indicates if DC j is open (value of 1) or closed (value of 0). This solution encoding is shown in Figure 3. This binary encoding only considers if a given DC j is open or closed (variables Yj). A solution of MOLIP also involves the assignment of buyers to open DCs (variables Xji ). This assignment is performed by a procedure that tries to minimize cost without deteriorating coverage and capacity.

Figure 3

Solution encoding of the MOLIP problem

The capacity constraints in MOLIP limits the amount of buyers demands that can be assigned to each candidate DC. A greedy heuristics is used to obtain the buyer-DC assignments where the buyers are sorted in the descending order of their demand flows and assign them in the sorted order to the DC according to the following greedy rules: Rule 1: If the buyer i is covered (there are DCs within a coverage distance), it is assigned to the DC with sufficient capacity (if exists) which can serve it with the minimal difference between the remaining capacity of an open DC j and the demand flow of the buyer i through DC j. That is, a DC is assigned as full as possible. Rule 2: If the buyer i cannot be covered or there is no successful assignment from its coverage set, it is assigned to the candidate DC (with sufficient capacity) that increases the total cost by the least amount, regardless of its distance to the DC if possible. These assignment rules seem fairly reasonable for the practical cases if the facility cost of opening a specific DC is much more expensive than distribution costs. Reducing the cost of wasting DC capacities then becomes an important issue as compared to the distribution costs. However, this procedure cannot guarantee that each buyer can be assigned to a specific DC due to the fact of capacity limitations on the DCs. However, this procedure cannot guarantee that each buyer can be assigned to a specific DC due to the fact of DCs capacity limitations. 5. Model Approach and Experimentation 5.1 Base-Line Problem

There is no MOLIP instances in the public domain nor available in the literature to serve for benchmarking. For this reason, a random instance base-line problem using a problem generator was developed by taking as reference the size of a Gamma.com supply chain network with 15 DCs and 50 buyers. The potential DC locations are randomly generated within a square of 100 distance units of width. Once established, these DCs would serve a total of 50 clusters of buyers, and their mean demands of all clusters are given in advance. Other key input parameters of the model are given in Table 2. For simplicity, Euclidean distance is used for measuring distribution distances. The company intended to determine the number of opening DCs needed for order assignments but the capacity limitation of DCs will affect the assignment results of buyers. The managers also need to evaluate tradeoffs among three criteria: total cost (TC), volume fill rate (VFR) and responsiveness level (RL). To obtain
* the approximate Pareto front PF , we attempted to solve the specified problem using the

proposed hybrid evolutionary approach. Table 2

Parameters
Annual cost of operating a DC j Annual holding cost at DC j for product k Unit ordering cost at DC j for product k Capacity of DC j Unit variable transportation cost Unit production and shipping cost for product k from the supplier to DC j Maximal covering distance Lead time (daily) Working days per year Average daily demand for product k at buyer i Standard deviation of daily demands Standard normal value (service level =0.95)

Model parameters for the base-line problem

Value
U(900,1000) U(.2,.4) U(8,10) U(500,700) $5 U(1,3) 25 Km U(2,4) 260 U(60,80) U(2,4) 1.96

Through the GA approach, the base-line model (number of DCs = 15, number of buyers =

50) with product number (k=2) resulted in 765 binary variables and 815 constraints. With
* discrete number of alternatives in the population of PF , a multi-criteria decision process

based on comparisons of alternatives by the decision makers to assign weighted scores to calculate distance metrics from the ideal point is implemented to select the optimal solution. 5.2 Computational Results The proposed evolutionary approach is used for the base-line model. The input parameters are: population size=100; generation number =200; cloning = 20%; crossover rate = 80%; mutation rate varies from 5% to 10%. The approach program was coded in MATLAB. The hybrid GA allows the decision maker to rapidly find a set of Pareto optimal solutions that are large in number. The decision maker requires to determine weights by prior knowledge of objectives to generate the user defined optimal solution that is considered as the one which is nearest to the ideal point.
100 90 80 70 60 50 40 30 20 10 0 0 10 20 30 40 50 60 70 80 90 100

Figure 4

A graphical display of the optimal base-line solution

Figure 4 illustrates the optimal solution chosen among the alternatives with equal weights of three objectives in the base-line model that has the minimal total cost

$251,112,536, the maximal volume fill rate 71.97%, and the maximal customer responsive level 62.15%, respectively, where 9 out of 15 candidate DCs are required to open and aggregated. It is worth mentioning that most of the aggregated DCs could be assigned possibly to the buyers as close as to them within the maximal coverage. However, there are about 29.03% unassigned buyers ( ) which reveal percentage of the uncovered demands and would possibly result in sales losses. Also, there are 37.85% aggregated buyers () assigned to DCs farther than the coverage distance. 5.3 Hybrid GA Performance Evaluation The hybrid GA is evaluated with the illustrative example. Figure 5 illustrates a good evolution approach for generating the Pareto frontier after 200 generations in our experiment. The longer the GA is run, the more evenly the frontier is covered.
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1 0.5 0 2.5 3 3.5 4 x 10
8

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0 1 0.5
1 1.5 2 4.5

0 1 0.5

0.5

0.5

1.5

2.5

3.5 x 10
8

0.5

1.5

2.5

3.5 x 10
8

(a) First Generation (n=1) Figure 5

(b) Generation (n=50)

(c) Generation (n=200)

Evolutionary solutions of the base-line problem

To verify the convergence efficiency of the GA approach, Figure 6 plots out the 2D tradeoff curves of two objectives in MOLIP at specific generations. It is revealed that the population curve converges shortly after 10 generations, and after 50 generations, they are

nearly overlapped among themselves. Afterwards, no significant improvement is incurred.

Approximate Pareto Frontier (TC v.s. VFR)
1 0.9

Approximate Pareto Frontier (TC v.s. RL)

(pop = 100)
1 0.9

(pop = 100)

Volume Fill Rate

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 5 10 15 20 25 30 35 40 45

Responsiveness Level

0.8

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 5 10 15 20 25 30 35 40 45

gen =200 gen=50

gen =400 gen=10

gen=200 gen=50

gen=400 gen=10

Total Cost (x10 )

Total Cost (x10 7)

Figure 6

The convergence of the Pareto frontier

5.4 Experiments with sensitivity analysis 5.4.1 Sensitivity analysis with varying coverage

The increasing coverage distance of a DC to buyers can enhance the level of customer service and thereby increase responsiveness level. Table 3 summarizes the sensitivity analysis results with seven possibilities of coverage distances (e.g., 10, 15, .., 40 km). It has shown that the average total cost (TC) of the Pareto frontier decreases when the coverage distance increases from 10 to 20 and after that, the farther the coverage distance the larger TC incurred drastically. Thus, the coverage distance of 20 would result in the best case of total cost saving when other parameters are left unchanged. On the other hand, the average volume fill rate (VFR) decreases when the coverage distances increases from 10 up to 30 and increases again after that. Compromising TC and VFR, we recommend the ideal coverage distance 25 which results in the best tradeoff solution by choosing a smaller number of DCs (say 7.26 in our case) due to tradeoffs between savings in total costs and prevention of losing sales.

Table 3 A sensitivity analysis with varying coverage distances of DCs

Dmax
TC (Ave.) VFR(Ave.) RL (Ave.) No. of DCs

10 $20.35 61.92% 27.28% 8.13

15 $20.18 59.67% 41.73% 7.89

20 $19.64 56.57% 43.65% 7.46

25 $19.81 55.17% 47.85% 7.26

30 $20.11 53.11% 49.96% 6.96

35 $24.22 58.32% 56.77% 7.68

40 $27.04 60.12% 58.94% 7.92

It is reasonable to believe that MOLIP is mainly dominated by the number of DCs. When the coverage distance is small (say 10), more DCs are chosen for better VFR and hence increases TC. Consequently, the larger distance leads to the reduction in total number of DCs and the subsequent savings in TC (up to 20 km). After that, TC increases again of a farther coverage distance for the requirement of a larger number of DCs. It is found that the number of opening DCs is likely the same at the 15 and 40 coverage instances respectively. However, the large 40 coverage instance incurred more TC ($27 million) than the smaller 15 coverage instance ($20 million only). The reason for this result may be because that the farther distance would increase more logistic costs in TC to retain the same VFR. 5.4.2 Sensitivity analysis of varying inventory cost

The longer the inventory were kept at DCs, the greater the saving in transportation cost, but the larger the inventory holding cost that was incurred. As such, we experimented on the model with varying unit inventory holding costs (e.g., 0.05, 0.1,.., 2), while keeping unit transportation cost constant. Figure 7 shows that as the unit inventory cost increases, TC will increase slightly since the number of open DCs is not explicitly affected by the unit inventory cost. That is, TC was insensitive to changes in unit inventory costs. For example, in the 25

coverage instance, the increase of the unit inventory holding cost in a significant manner from $0.05 to $2 has just little increment of just 8.9% over the minimal total costs. The analysis also depicted similar results in 10 and 40 coverage instances. The results also illustrate that the inventory costs are not significant enough to make additional DCs open but keep on accumulating the total supply chain costs linearly.

$32.00 $30.00 $28.00

T (a ra ) C ve ge

Dmax=10 Dmax=25 Dmax=40

$26.00 $24.00 $22.00 $20.00 $18.00 0.05 0.1 0.2

Unit holding cost

0.4

0.6

0.8

1.5

Figure 7 5.4.3

Sensitivity analysis with varying unit inventory holding cost

Sensitivity analysis of varying DC capacities

The capacity of a DC limits the amount of buyers demands be assigned to a DC. We determine DCs capacities in the following way. If a DC is open in the MOLIP solutions, we set the capacity of a possible DC location to the amount of the buyers demand that is multiplied by a capacity constant. We choose in our experiments this constant to be 70 which is the mean demand value of all buyers. We perform sensitivity analysis on different capacity settings by varying capacity multipliers. The performance of the procedure is implemented for eight capacity settings by setting capacity multipliers from 3 to 10 which illustrates different extents of capacity tightness.

Figure 8 depicted the trends of %TC, %VFR and %RL in MOLIP where three curves are nearly overlapped among themselves, which means that the effects of changing capacity tightness are similar for these problem performances. It shows that the number of DCs decreases as the capacity multiplier increases since the additional DC capacity enables more risk-pooling. Such risk-pooling effects would make an open DC facing increasing customer demands so as to modify its inventory policy and DC-buyer assignments. When the demands are increased enough at specific DCs with capacity limitation, the inventory costs will increase asymptotically which result in the increase of the total costs.
300% 280% 260% 240% 220% 200% 180% 160% 140% 120% 100% 3 4 5 6 7 8 9 10 %Diff in TC %Diff in VFR %Diff in RL

Capacity Multiplier

Figure 8 6. Concluding remarks

Sensitivity analysis with varying capacity tightness

This paper presented an MOLIP model for jointly examining the effects of facility location, distribution, and inventory related issues. The model determines the location of DCs and the assignment of buyers to DCs with three performance measurements: total costs, volume fill rate, and responsiveness level. The integrated system permits a comprehensive evaluation of the tradeoffs that exist among strategic and tactical issues of supply chain

management. MOLIP enables the decision maker to evaluate alternative solutions. The possibility of using MOEAs for solving MOLIP is also investigated. A NSGAII-based evolutionary approach for determining the best tradeoff solutions of a base-line problem is implemented on to obtain near-optimal solutions with reasonable computations. Computational analysis shows the efficacy of this evolutionary approach for the MOLIP in solving a practical-size problem with 50 buyers and 15 potential DCs. Hopefully, it could be a new efficient approach for such kinds of more difficult-to-solve problems. This proposed model can be extended in several ways. First, inclusion of other inventory decisions: frequency and size of the shipments from plants to the DCs and from DCs to the buyers based on different replenishment policies would be research directions that worth pursuing. Second, the inventory costs at the buyers has been explicitly modeled since our model only focuses on a centralized inventory policy. Third, we have not explicitly modeled the process of emergency shipments from the DCs to the retailers in the event of a stockout at the retailer. It is likely to include stockout and backorder costs in the model. From an analytical and computational solution procedure point of view, the proposed evolutionary approach points to a number of directions for future work. First, the future research should include multiobjective treatments to explicitly analyze the tradeoffs among costs, response time and demand fill rates. Second, the consideration of what-if scenarios involving changes in parameter values over time may be explored in the future to find out the

potential impact of the potential model. Third, the algorithm performance should be explored and evaluated on a variety of problems with varying sizes. Finally, the comparisons of the proposed evolutionary approach with other heuristics or algorithms such as Lagrangian relaxation, heuristics or soft computation approaches are worth investigating in the future. 7. References [1] C. Coello Coello, A comprehensive survey of evolutionary-based multiobjective optimization techniques, Knowledge and Information Systems, 1(3), pp. 269-308, 1999. [2] M. Daskin, C. Coullard, and Z. Shen, An inventory-location model: formulation, solution algorithm and computational results, Annals of Operations Research, 110, pp.83-106, 2002. [3] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, A fast and elitist multi-objective genetic algorithm: NSGA-II, IEEE Transaction on Evolutionary Computation, 6(2), pp. 181-197, 2002. [4] G. Eppen, Effects of centralization on expected costs in a multi-location newsboy problem, Management Science, 25(5), pp. 498501, 1979. [5] S. Gaur and, and A. Ravindran, A bi-criteria model for the inventory aggregation problem under risk pooling, Computers & Industrial Engineering, 51, pp. 482-501, 2006. [6] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution Programs, Springer-Verlag, Berlin, 1996.

[7]

P. Miranda and R. Garrido, Incorporating inventory control decisions into a strategic distribution network model with stochastic demand, Transportation Research Part E, 40, pp. 183-207, 2004.

[8]

L. Nozick and M. Turnquist, A two-echelon allocation and distribution center location analysis, Transportation Research Part E, 37, pp.425-441, 2001.

[9]

L. Ozsen, Location-Inventory Planning Models: Capacity Issues and Solution Algorithms, PhD Dissertation, Northwestern University, IL, 2004.

[10] E. Sabri and B. Beamon, A multi-objective approach to simultaneous strategic and operational planning in supply chain design, Omega, 28, pp. 581-598, 2000. [11] J. Schaffer, Multiple objective optimization with vector evaluated genetic algorithms, in Genetic Algorithms and their Applications, in Proc. of the First International Conference on Genetic Algorithms, Hillsdale, NJ, pp. 93-100, 1985. [12] Z. Shen, C. Coullard and M. Daskin, A Joint Location-Inventory Model, Transportation Science, 37, pp. 40-55, 2003. [13] Z. Shen and M. Daskin, Trade-offs between customer service and cost in integrated

supply chain design, Manufacturing and Service Operations Management, 7(3), pp. 188-207, 2005. [14] N. Srinivas and K. Deb, Multiobjective optimization using nondominated sorting in genetic algorithms, Evolutionary Computation, 2(3), pp. 221248, 1994.