Location Optimization of Multiple Distribution Centers under Fuzzy Environment Yong Wang Ph.D. Student School of Transportation Engineering, Southeast University Nanjing, China, 210096 Email: yongwx6@uw.edu Xiaolei Ma Ph.D. Student Department of Civil and Environmental Engineering, University of Washington, Seattle, WA, 981952700 Email: xiaolm@uw.edu Yunteng Lao Ph.D. Student Department of Civil and Environmental Engineering, University of Washington, Seattle, WA, 981952700 Email: laoy@uw.edu Yinhai Wang (Corresponding Author) Professor Box 352700 Department of Civil and Environmental Engineering, University of Washington Seattle, WA 98195-2700 Tel: (206) 616-2696 Fax: (206) 543-1543 Email: yinhai@u.washington.edu Haijun Mao Professor School of Transportation Engineering, Southeast University Nanjing, China, 210096 Email: maohaijun@seu.edu.cn Wang, Ma, Lao, Wang, Mao 1 ABSTRACT Locating distribution centers optimally is a crucial and systematic task for logistic operators and decision makers. Optimally located distribution centers can significantly improve the logistic system’s efficiency and reduce its operational cost. However, it is not an easy task to optimize distribution center locations and previous studies focused primarily on location optimization of a single distribution center. With the growing logistics demands, multiple distribution centers become necessary to meet customers’ requirements, but few studies have tackled the Multiple Distribution Centers Location (MDCL) problems. This paper presents a comprehensive algorithm to address the MDCL problem. Fuzzy integration and clustering approach using the Improved Axiomatic Fuzzy Set (AFS) theory is developed for location clustering based on multiple hierarchical evaluation criteria. Then, Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) is applied for evaluating and selecting the best candidate for each cluster. Sensitivity analysis is also conducted to assess the influence of each criterion in the location planning decision procedure. Results from a case study in Guiyang, China, reveals that the approach developed in this study outperforms other similar algorithms for multiple distribution center location selections. This new method may easily be extended to address location planning of other types of facilities, including hospitals, fire stations, and schools. Keywords: Multiple distribution centers; Location selection; Clustering algorithm; Axiomatic Fuzzy Set; TOPSIS. Wang, Ma, Lao, Wang, Mao 2 INTRODUCTION Distribution centers play a vital role in transportation and logistics system. With the growing urban freight movements, a single distribution center in a large region cannot efficiently accommodate the growing demands from customers due to lengthened transit time and high warehouse handling cost, especially for Just-In-Time and Cold Chain deliveries. Therefore, systems with multiple distribution centers become highly desirable in practice and locating these centers optimally is a critical issue faced by logistics system operators and transportation planners (Yang et al. 2007). A method for optimal location selection of multiple distribution centers helps balance production and consumption, which further improves the efficiency in logistic systems, optimizes distribution system network, and alleviates urban traffic congestion. However, developing such a method is not a straightforward task because many factors and criteria need to take into account in the decision-making procedure (Sun et al. 2008; Syam 2002). For example, distance to the shared production facility, local traffic pattern, economic condition for each potential distribution center, all need to be considered. These factors require certain domain knowledge to justify and should be adequately quantified and incorporated into the multiple locations selection problem. It is worth mentioning that a general solution to the problem of multiple distribution center locations selection (MDCL) also benefits the planning of other types of multiple facility locations such as hospitals, fire stations, and schools. Most previous studies focused on selecting single distribution center location rather than simultaneously selecting multiple distribution center locations. The two most classical quantitative frameworks for single facility location selection are the cost minimization approach (Hansen et al. 1992; Lee 1993; Tyagi and Das 1995) and the profit maximization approach (Hakimi and Kuo 1991). Yang et al. (2007) constructed a mathematical model to minimize the total relevant cost, including setup cost, turnover cost, and transportation cost with each customer’s demand. Sun et al. (2008) presented a bi-level mixed-integer programming model to seek for the optimal distribution center location, where the upper-level model minimized the planners’ cost, and the lower-level model minimized the customers’ cost. Hande et al. (2011) also formulated a bi-level mixed-integer nonlinear programming model to deal with a competitive facility location problem based on the game theory, which maximized the profit and the attractiveness of each facility simultaneously. MDCL problem needs to take into account a wealth of influential factors such as resource availability and local economic conditions. These factors are traditionally measured by human perception, and are considered fairly difficult to process by computers in an automatic fashion. To tackle the vagueness of the selection criteria, Fuzzy set theory was used to evaluate the numerous attributes of potential facilities’ locations, and has been utilized to handle ambiguous scenarios in the decision-making procedure by many researchers ((12), (13), (14), (15)). One of the main merits for fuzzy set theory is that it is capable of not only considering uncertainty in human’s cognitive process, but also evaluating different subjective attributes by linguistic terms ((16), (17)). Fuzzy set theory was then developed and applied in the facility location selection process, since valuation criteria for each alternative location are often defined in linguistic terms, e.g. “Low”, “Medium”, ”High”, ”Poor”, “Fair”, “Good”, etc. (13). This suggests that Fuzzy set theory is well suitable to transform these linguistic attributes into fuzzy relations, fuzzy numbers, and fuzzy inference systems (18). Thus, there are many facility location selection studies Wang, Ma, Lao, Wang, Mao 3 employed fuzzy set theory to interpret evaluation criteria, such as labor costs, resource availability, and freight regulation (19). Based on fuzzy set theory, numerous systematic analysis methods are adopted into the potential location ranking process. Chou proposed an integrated short-term and long-term multiple-criteria decision-making approach to solve location selection problems (20). Wey and Chang presented a hybrid analytical hierarchy process (AHP) and data envelopment analysis (DEA) location ranking method (21). Yu et al. (22) ranked and evaluated the potential location plans for multiple urban transit hubs by using an AHP-based framework and fuzzy set theory. Among all location ranking algorithms, the TOPSIS algorithm proposed by Negi is most widely used and has been improved by many researchers (23). Chu presented a fuzzy TOPSIS model under a group of decisions to solve the facility location selection problem (19). Awasthi et al. (15) also used the fuzzy TOPSIS approach to evaluate and select the best urban distribution center. However, these methods can only be used to select single facility location, and most of them are hard to expend to tackle the MDCL problem. To fill this research gap and provide a potential solution for the MDCL problem, a novel approach for multiple distribution centers location selection is proposed. The remaining part of this paper is organized as follows: Comprehensive evaluation index system is first established by using fuzzy integration and clustering algorithms, which groups several potential locations into different clusters. Then fuzzy TOPSIS method is applied for selecting the alternative distribution centers location within each cluster. Next, a case study of MDCL in Guiyang, China is presented to demonstrate the effectiveness of our method, followed by a comparison among different relevant algorithms. Last and but not least, a sensitivity analysis is conducted to examine impacts incurred by changing selection criteria weights. Finally, conclusions are given at the end of this paper. METHDOLOGY The study methodology contains three parts: 1) major criteria and sub-criteria are established to form a hierarchical analysis structure; 2) the linguistics variables used for location evaluation are then defined and transformed into fuzzy numbers for further assessments; and 3) a location planning framework for urban multiple distribution centers is finally developed based on the fuzzificated evaluation criteria. Hierarchical analysis structure for Location Planning The location selection criteria are used for evaluating potential urban distribution centers location. As shown in Figure 1, six major criteria ( Ci ) and sixteen sub-criteria ( Cij ) are finalized to determine the locations for implementing urban distribution centers. These criteria are obtained from previous studies (e.g. Chen, 2001; Heizer and Render, 2004; Stevenson 2005) and are considered important after the discussions with several transportation experts and members in the local transportation agencies. Wang, Ma, Lao, Wang, Mao 4 Geological condition C11 Natural environment C1 Hydrology condition C12 Landform condition C13 Road facilities C21 Transportation C2 Accessibility C22 Connectivity to Multi-modal transport C23 Business environment C3 Proximity to Customers C31 The degree of close to logistics center C32 Resource availability C41 Candidate land C4 Sustainable freight regulations C42 Land costs C43 Water supply C51 Supply condition C5 Power supply C52 Gas supply C53 Environmental impact C6 Drainage C61 Solid castoff disposal C62 l s er nt ce n io ut ib tr is D e pl ti ul M FIGURE 1 Hierarchical Analysis Structure for Locations Selection of Multiple Distribution Centers These experts gave their linguistic evaluations on each sub-criterion. Based on these evaluations, each potential location can be evaluated and graded using linguistic terms by the experts. The scores of an upper level criterion can then be calculated from these sub-criteria evaluations. The above sub-criteria structure can be divided into two categories: cost criteria and benefit criteria. The criterion C43 , criterion C61 and criterion C62 fall into the cost criteria, where an alternative location with higher cost is less likely to be selected as the final location. The remaining sub-criteria are grouped as benefit criteria, where an alternative location with higher benefit criteria is more likely to become the final location. Linguistic Variables Fuzzification and Related definitions Linguistic Variables Fuzzification As the first step, natural languages are converted into numerical inputs using fuzzy set theory. Adopted in this study is the triangle fuzzy transformation, in which a triangular fuzzy number is Wang, Ma, Lao, Wang, Mao 5 represented as a triplet n = (n1 , n2 , n3 ) (15, 25), and the membership function µn ( x) can be calculated based on the triangular fuzzy number n as follows: 0, x−n 1 , n2 − n1 µ n ( x) = n3 − x , n3 − n2 0, x ≤ n1 , n1 ≤ x ≤ n2 , (1) n2 ≤ x ≤ n3 , x ≥ n3 Where n1 , n2 , n3 are real numbers, x at n2 gives the maximal grade of µn ( x) , and x at n1 and n3 gives the minimal grade of µn ( x) . In the fuzzy set theory, each linguistic term is transformed into a fuzzy number. Due to the uncertain nature of the MDCL problem, criteria ratings and alternative ratings are used as the linguistic variables. In the current study, as shown in Table 1, we apply a scale of 1-9 (VL-VH) to rate the criteria and (VP-VG) to rate the alternative locations. Table 1 presents the linguistic variables and fuzzy numbers for each criterion and alternative. TABLE 1 Linguistic Terms for Criteria and Alternative Ratings Linguistic term Abbreviation Fuzzy numbers Very Low (Very Poor) VL (VP) (0.56,0.62,0.65) B.VL&L (B.VP&P) (0.62,0.65,0.68) L (P) (0.65,0.68,0.73) B.L&M (B.P&F) (0.72,0.75,0.78) M (F) (0.75,0.78,0.83) B.M&H (B.F&G) (0.80,0.83,0.90) H (G) (0.85,0.88,0.93) B.H&VH (B.G&VG) (0.90,0.93,0.96) VH (VG) (0.95,0.98,1) Between Very Low and Low (Between Very Poor and Poor) Low (Poor) Between Low and Medium (Between Poor and Fair) Medium (Fair) Between Medium and High (Between Fair and Good) High (Good) Between High and Very High (Between Good and Very Good) Very High (Very Good) Related Definitions To address the MDCL selection problem, several related definitions are needed and presented as follows. Wang, Ma, Lao, Wang, Mao 6 Definition 1. Notations for the parameters are defined as follows: D{Di i = 1, 2, , m ' } denotes the decision makers who evaluate the alternative distribution center locations; C 1{Ct1 t = 1, 2, , r} denotes the major criteria for location selection of multiple distribution centers; C 2 {Ct2 t = 1, 2, , k} denotes the sub-criteria for location selection of multiple distribution centers; A{ Ai i = 1, 2, , n} is the potential location i of urban distribution centers; = Ah' { Ahi' i 1,= 2, , n ; h 1, 2, , c} is the number of potential locations within cluster ' = Wut2l (u 1,= 2,, m' ; t 1, 2, , r ) h; denotes the fuzzy numbers for sub-criterion l of major criterion t by decision maker u ; = Wut2' (u 1,= 2, , m ' ; t ' 1, 2, , k ) denotes the fuzzy numbers for sub-criterion t by decision maker u ; 2 = X uhit , m' ; h 1,= , n' ; t ' 1, 2, , k ) denotes the fuzzy numbers for alternative 1, 2,= 2,, c; i 1, 2,= ' (u location i under sub-criterion t ' by decision maker u within cluster h ; = xujt2 l (u 1, 2,= , m' ; j 1,= 2, , n; t 1, 2, , r ) denotes the fuzzy numbers for alternative location j under sub-criterion l of major criterion t by decision maker u ; p '{ ph' h = 1, 2, , c ' } represents each initial cluster, and c ' is the total number of initial clusters. p{ ph h = 1, 2, , c} represents each final cluster, and c is the total number of final clusters. Definition 2. X {x1 , x2 , , xn } ⊆ R n denotes the sample set of potential urban distribution centers = location; µ = {µ1 , µ2 , , µr } is the attribute set of X ; xtj = µt ( x j ) , t = 1, 2, , r , j = 1, 2, , n , where xtj denotes the membership function value of the sample x j with attribute µt ; The attribute µt can be further separated as mt ,1 , mt ,2 , , mt , st , thereby, the attribute set µ can be expressed as µ =M = {m1,1 , m1,2 , , m1,s1 , m2,1 , m2,2 , , m2,s2 , , mr ,1 , mr ,2 , , mr ,sr } , which is defined as the fuzzy attribute (concept) set. Definition 3. Let X , M be sets, 2 M is the power set of M , τ:X × X → 2 M , if τ satisfies the following conditions (26), AX 1:∀( x1 , x2 ) ∈ X × X ,τ ( x1 , x2 ) ⊆ τ ( x1 , x1 ) AX 2:∀( x1 , x 2 ),( x2 , x3 ) ∈ X × X ,τ ( x1 , x2 ) ∩ τ ( x2 , x3 ) ⊆ τ ( x1 , x3 ) Wang, Ma, Lao, Wang, Mao 7 ( M ,τ , X ) is called an Axiomatic Fuzzy Set (AFS) structure, X is called a sample set, M is called an attribute (concept) set, τ is called structure. Definition 4. Let R be the binary relation of the sample set X (27), when x, y ∈ X , x ≠ y, R satisfies the following conditions, (1) If ( x, y ) ∈ R , then ( x, x) ∈ R ; (2) If ( x, x) ∉ R and ( y, y ) ∈ R , then ( y, x) ∈ R ; (3) If ( x, y ) ∈ R and ( y, z ) ∈ R , then ( x, z ) ∈ R ; (4) If ( x, x) ∈ R and ( y, y ) ∈ R , then ( x, y ) ∈ R or ( y, x) ∈ R ; R is called sub-preference relation, the corresponding concept of sub-preference relation is called simple concept, otherwise R is called complex concept. Definition 5. Let m be the simple fuzzy concept, and m ∈τ ( x, y ) , ρ m:X → R + =[0, ∞) (28), (1) ρ m ( x) = 0 ⇔ ( x, x) ≠ Rm , x ∈ X ; (2) ( x, y ) ∈ Rm ⇒ ρ m ( x) ≥ ρ m ( y ), x, y ∈ X ; ρ m is called membership function of simple fuzzy concept m . Definition 6. Let η(x j ) be the simple fuzzy concept set, that is, η ( x= or st ;j 1, 2, , n} , let ζ A be the fuzzy description of each 2, , r; s ' 1 or = j )={mt , s ' |t 1,= j sample Aj , then, ζ Aj =argmin{ ∏ mt , s' | mt , s' ∈η ( x j ),=t 1, 2, , r;= s ' 1 or or st ;=j 1, 2, , n} . Definition 7. Let X , M be two sets, and let ( M ,τ , x) be an AFS structure (29), A ⊆ X , B ⊆ M , then B(= A) { y | y ∈ ρ m ( x), m ⊆ B, ∀x ∈ A} , where B ( x) denotes the samples belonging to the degree of B is less than or equal to the sample x belonging to the degree of B . Definition 8. Let X be a sample set, and X ⊆ R n , ρ m is the membership function of simple fuzzy concept m , and m ⊆ B , then Lm ( x) is measure of sample x belonging to the simple fuzzy concept m , the Lm ( x) is defined as Lm ( x) = ∑ ∑ x∈m ( x ) ρ m ( x) ρ m ( x) x∈ X (2) Definition 9. Let the fuzzy concept B ⊆ M , the membership function µ B ( x) of fuzzy concept B is defined as Wang, Ma, Lao, Wang, Mao 8 = µ B ( x) inf ( Lm ( x)) ∈ [0,1] , x ∈ X (3) m∈B Definition 10. Let X be the sample set of potential urban distribution centers location, M is the simple fuzzy concept set of X , B ⊆ M , n is the sample size, the membership information entropy (30) function E ( B ) and the membership distribution coefficient function D( B) can be defined as E ( B ) = − ∑ ( µ B ( x) lg( µ B ( x))) (4) D( B) = −(( ∑ µ B ( x) n) lg( ∑ µ B ( x) n)) (5) x∈ X x∈X x∈X When E ( B ) becomes smaller, the membership degree of the sample x within concept B is more approaching to the two ends of the interval [0, 1], and the boundaries are clearer. When D( B) is smaller, the membership degree of the sample x within concept B is more approaching one end of the interval [0, 1], instead of both left and right ends. Thereby, in order to comprehensively evaluate the membership information entropy and distribution coefficient, the evaluation index V can be defined as: V = E ( B) D( B) , where when V is smaller, it is more reasonable that the concept B describes the sample X . The Proposed Framework for Location Planning There are three major steps included in this framework. Each evaluation sub-criterion should be properly mapped into the higher hierarchical criterion based on the fuzzy integration method in the first step, and next, clustering algorithm is undertaken to group the possible alternative locations into different clusters. The last step is to finalize the final location within each cluster by using the fuzzy TOPSIS approach. These steps are presented in details as follows: Fuzzy integration method based on sub-criteria The xujt2 and Wut2 are expressed respectively as triangular fuzzy numbers: xujt2 = (aujt2 , bujt2 , cujt2 ) , Wut2 = (hl2 , gl2 , kl2 ) . The comprehensive evaluation index from all of experts for location j under l l l l l l l major criterion t can be expressed as ' 1 Z= tj m 1 2 2 ⊗ ∑ (( xujt ⊗ Wut21 ) ⊕ ⊕ ( xujt ⊗ Wut2s )) ' 1 s m × ts u =1 (6) Where ⊗ denotes vector multiplication, and ⊕ denotes vector addition. Z tj1 denotes the comprehensive triangular fuzzy number to evaluate the location j under major criterion t , m' denotes the number of experts, ts denotes the number of sub-criteria is s under major criterion t . Let Y = (a, b, c) be a triangular fuzzy number, the representation of triangular fuzzy number 1 Y is P(Y ) = (a + 2b + c) ((13), (14) (19), (20), (31), (32),). Thereby, if we suppose that 4 Wang, Ma, Lao, Wang, Mao 9 Z tj1 =(Ttj1 ,Q1tj ,H1tj ) denotes the triangular fuzzy number, then the integrated membership function value µtj1 of the location j under major criterion t will be described as µtj1 = 1 1 (Ttj +2Q1tj +H1tj ) 4 (7) Clustering algorithm procedure The integrated membership function value µtj1 calculated from the fuzzy integration method is used as the input. Our next step is to conduct clustering analysis to divide the potential distribution centers into different clusters. Axiomatic fuzzy set (AFS) theory logic ((27),(28),(29)) has been proven as an effective approach to tackle human perception related clustering problem. Since the traditional AFS method is not appropriate to incorporate a large number of criteria (33), an improved AFS theory logic algorithm is detailed as follows: Step 1. According to Definition 2, the integrated membership function value µtj1 can be transformed into three values, they are expressed as ρ m , ρ m , ρ m respectively, and if we define t ,1 t ,2 t ,3 µt ( x j ) = µ ,and then, the membership function corresponding to each membership function 1 tj value µtj1 can be written as ρ mt ,1 ( x j ) = µt ( x j ) ρ mt ,2 ( x= ht ,1 − µt ( x j ) j) ρ mt ,3 ( x j ) = ht ,3 − µt ( x j ) − ht ,2 (8) ht ,1 = max {µt ( x1 ), µt ( x2 ), , µt ( xn )} µt ( x1 ) + µt ( x2 ) + + µt ( xn ) n = ht ,3 max{ µt ( x j ) − ht ,2 + ht ,2 } ht ,2 = 1≤i ≤ n = t 1,= 2, , r; n number of potential locations Step 2. Calculate the fuzzy attributes of each sample. Step 2.1 Calculate Lmt ,1 ( x j ) , Lmt ,2 ( x j ) , Lmt ,3 ( x j ) through Eq.(2) and the Step 1. Step 2.2 Define µmt ,1 ( x j )=Lmt ,1 ( x j ) , µmt ,2 ( x j )=Lmt ,2 ( x j ) , µmt ,3 ( x j )=Lmt ,3 ( x j ) , and the maximum of membership fuzzy values under simple fuzzy concept µt can expressed as: { } µm ( x j ) = ρ m ' ( x j )= max µmt ,1 ( x j ), µmt ,2 ( x j ), µmt ,3 ( x j ) t ,s ηt , s' ( x j ) = mt , s' t ,s' , t 1,= 2, , r , s ' 1 or 2 = or 3, j 1, 2, , n = Step 2.3 Respectively calculate the ratio of membership information entropy and distribution coefficient function, which corresponds to each sample attribute of step 2.2 as (9) Wang, Ma, Lao, Wang, Mao Vη Step { t ,s' 2.4 (xj ) 10 = E (ηt , s' ( x j )) D(ηt , s' ( x j )) η(x j ) Let be the (10) simple } fuzzy concept set, that is, ( x j ) | t 1,= 2, , r ; s ' 1 or 2 = = η ( x j ) η= or 3; j 1, 2, , n . t , s' Step 2.5 Select the smallest value µa ( x) which corresponds to attribute a ; Select the second smallest value µb ( x) which corresponds to attribute b ; they are defined as follows: µa ( x j ) = ( x j )' Let η ( x j )' be η= inf ( Lα ( x j )) ∈ [0,1] (11) α ∈η ( x j ) {η ( x ) − {a}} , and µ ( x ) is, j = µb ( x j ) b inf ( Lα ( x j )) ∈ [0,1] (12) α ∈η ( x j )' Step 2.6 Find the evaluation index Vη t ,s' evaluation index Vη t ' ,s' (xj ) (xj ) ≥ Vη t ,s' (xj ) ( x j )' attributes are expressed as η= (xj ) < Vη t ,s' (xj ) which corresponds to attribute a , and the t ,s' t ' ,s' t ' ,s' (xj ) which corresponds to attribute b , and compare Vη Step 2.7 If Vη until Vη j (xj ) and Vη t ' ,s' (xj ) ; , then eliminate attribute a , and the remaining sample {η ( x ) − {a}} , and return to step 2.5, continue the recursion j ; Step 2.8 Return the final remaining attributes η ( x j )' for each sample x j , the remaining attributes can be used to describe the sample x j , thereby, we get the fuzzy attributes of each sample via the above step 2.1-step 2.8. Step 3. Clustering procedure will be based on the fuzzy attributes of each sample Step 3.1 The fuzzy description of each sample Ai is ζ Ai , and we construct the fuzzy { sample relation rij = min µζ Ai Λζ A j } . Based on universe of discourse X = { A , A ,, A } , ( Ai ), µζ A Λζ A ( Aj ) i j 1 2 n Liu (1998) has demonstrated that the integer k exists to make ( M Λk )2 = M Λk , thereby(29), we can deduct the equivalence relation at universe of discourse X based on the fuzzy sample relation k matrix= R M = (rijk ) n×n . Λ Step 3.2 Calculate the initial clustering results based on the fuzzy equivalence matrix Denote the diagonal element of fuzzy sample relation matrix as riik . It can be verified as k r= rijk ≤ riik , thereby we can identify the different membership function values based on the ji diagonal elements and other elements in the matrix. They are expressed as α l (l = 1, 2, , g ) , and Wang, Ma, Lao, Wang, Mao 11 these values gradually increase according to the sequence of l . Find rijk =riik =α l in the fuzzy sample relation matrix; the corresponding samples can be grouped into one or multiple clusters, and the remaining sample’s relation where rijk > α l can be grouped as another cluster. This procedure will recursively continue until l = g . Step 3.3 Calculate the weight of each fuzzy description within clusters The samples are divided into different clusters according to different α l values, we suppose that the initial clusters are p , p , , p based on α l , calculate the weight of each fuzzy ' 1 description of ph' : ξ p ' h { = C p' , wp' h h } , ξ = {m x { ' 2 g , s' ' c' } | g ∈ {1,2, , r= } , s' 1 or 2 or 3, x ∈ ph' , ξ x denotes the fuzzy } { } attributes set of ph' , C p = m m ∈ ξ x , x ∈ ph' ,= wp wm m ∈ C p , wm = ' h ' h ' h ph' m ∑ m∈C 'm h p { } , ph' m = x x ∈ ph' , m ∈ ξ x , ph' where . is defined as the number of elements, wm ∈ wp' is the weight of m ∈ C p' . h h Step 3.4 Finalizing the clusters based on the weighted membership function For each sample x ∈ X , calculate the weighted membership function values ' f ( c , w,h ) ( x) = ∑ wm Lm ( x) , h = 1, 2, , c , in each initial cluster, where Lm ( x) is the measure of m∈C ph' initial sample, select f ( c , w,h* ) ( x) = arg max { f ( c , w,h ) ( x)} which corresponds to h* , then x ∈ ph' , and * 1≤ h ≤ c' finally { obtain the result of } clusters p1 , p2 , , pc , and let ξ x = mg ,s' | g ∈ {1, 2, , r= } , s ' 1 or 2 or 3, x ∈ ph , C p = {m m ∈ ξ x , x ∈ ph } . h Step 3.5 Calculate the clustering validity index According to different thresholds α ∈ [0,1] , we can select the optimal result from the clustering validity index. Some notations are introduced to cluster the validity indices, N is the number of samples, p1 , p2 , , pc are the cluster results, c is the number of clusters, m V = {v1 , v2 , , vc } is the set of sample center in each cluster, = vh = vh ∑ρ x∈ ph m ( x) n ph m ∈ C ph , 1 ≤ h ≤ c , where ρ m is the membership function of simple fuzzy concept m , and n ph is the number of samples in cluster h . Thereby, the clustering validity index Iα can be expressed as follows: c c × (c − 1)∑ ∑ Iα = ∑ h= 1 x∈ ph m∈M ρ m ( x) − vhm 2 2 c c 2 × α × ∑ ∑ ∑ vhm − vkm h =1 k =1,h ≠ k m∈M (13) Wang, Ma, Lao, Wang, Mao c c ∑ ∑ ∑ 12 vhm − vkm where RC = h =1 k =1,h ≠ k m∈M hk 2 describes the dispersion degree between clusters, which can be c × (c − 1) 2 used for merging different clusters, and c ∑∑ ∑ h= 1 x∈ ph m∈C ph ρ m ( x) − vhm 2 denotes the closeness of samples within each cluster when Iα is the smallest value. This means the clustering is the most distinct. Ranking within cluster The Fuzzy TOPSIS approach (cite) is used to evaluate the potential locations based on the selected 16 sub-criteria. The goal is to find the negative ideal solution, where the benefit criteria is minimized and the cost criteria is maximized, and the positive ideal solution, where the cost criteria is minimized and the benefit criteria is maximized ((15),(34)). From the previous definitions, Ah' is the number of potential locations with cluster h , thereby, the fuzzy numbers of sub-criteria and potential locations are defined as: W 2 =(wut 1 ,wut 2 ,wut 3 ) , X 2 =(a uit ,buit ,cuit ) , ' ut ' = (u 1, 2,= , m ; h 1,= 2,, c; i 1, 2, = , n ; t 1, 2, , k ) ' ' ' ' ' uhit ' ' ' ' respectively. The procedures of fuzzy TOPSIS approach are described as follows: Step 1. Calculate the aggregate fuzzy ratings for the criteria and location alternatives Let w t be the aggregated fuzzy weights for sub-criteria within each cluster, and w t is ' ' expressed as w t ' = ( wt ' 1 , wt ' 2 , wt ' 3 ) , among which, { } wt ' 1 = min wut ' 1 , wt 2 = 1 ' u ' m m' ∑w u =1 ut ' 2 { } , wt ' 3 = max wut ' 3 . u (14) Let xt 'i be the aggregated fuzzy rating of location alternatives within each cluster, and xt 'i = (ait ' , bit ' , cit ' ) is expressed as { } ' { } m a it ' = min a uit ' , bit = 1 ' ∑ buit , cit ' = max cuit ' . u u ' m ' u =1 (15) Step 2. Calculate the location decision matrix and fuzzy weight matrix Based on step 1, location decision matrix R and the fuzzy weight matrix W can be calculated as follows: x11 x 21 R = xk1 x12 x1n' x22 x2 n' , W = ( w , w , , w ) 1 2 k xk 2 xkn' Step 3. Calculate the positive ideal and negative ideal solution. (16) Wang, Ma, Lao, Wang, Mao 13 Let R ' be the normalized fuzzy decision matrix, which is given as follows: ' R ' = [rt 'i ]k ×n' , t ' = 1, 2, , k ; i = 1, 2, , n ; where, a −' a −' a −' rt 'i = t , t , t c ' b ' a ' it it it − ait ' ) (Normalized cost criteria) , at ' = min( i a ' b ' c ' * cit ' ) rt 'i = it* , it* , it* , ct ' = max( i c' c' c' t t t (Normalized benefit criteria) (17) (18) Step 4. Calculate the weighted normalized matrix V by multiplying the fuzzy decision matrix R ' with the fuzzy weight matrix W , the vt 'i is calculated as ( ) ' vt 'i = rt 'i (.) w t ' = vt 'i1 , vt 'i 2 , vt 'i 3 , t = 1,2, , k ; i = 1,2, , n ' (19) Step 5. Calculate the positive ideal and the negative ideal solution C * = ( v1* , v2* , , vk* ) ,where vt*' = max(vt 'i 3 ) , t ' = 1, 2, , k ; i = 1, 2, , n ' (20) C − = ( v1 , v2 , , vk ) , where vt− = min(vt i1 ) , t ' = 1, 2, , k ; i = 1, 2, , n ' (21) i ' i ' Step 6. Calculate the dimensional Euclidean distance. The dimensional Euclidean distance of positive ideal solution is given as 1/ 2 k = d i* ∑ (vt 'i − vt*' ) 2 ' t =1 , i = 1, 2, , n ' (22) The dimensional Euclidean distance of negative ideal solution is given as 1/ 2 k = d i- ∑ (vt 'i − vt−' ) 2 ' t =1 , i = 1, 2, , n ' (23) Step 7. Calculate the closeness to the ideal solution Let CAi be the ratio of the distance to the fuzzy negative ideal solution di- and the sum of fuzzy positive ideal solution and fuzzy negative ideal solution, and it represents the relative closeness as follow: CAi = di− , i = 1, 2, , n ' * − di + di The higher the value CAi is, the more ideal the alternative is for the final location. (24) Wang, Ma, Lao, Wang, Mao 14 IMPLEMENTATION AND COMPARISONS Data Source To illustrate the applicability of the proposed approach in multiple facilities location planning, a multiple distribution centers location selection problem in Guiyang, China was used as a case study. Guiyang city is the capital of Guizhou Province, and is the transportation hub and an important distributing center to the whole province, as it is located in the north-central of Guizhou Province. The logistics company owns a logistics center in Guiyang city, and the company needs to establish multiple distribution centers to extend the market. Eighteen alternative distribution centers have been chosen as shown in the Figure 2, which are expressed as A1 , A2 , A18 . FIGURE 2 Alternative Distribution Centers Scatter Diagram In order to accurately select the location of distribution centers, three top decision makers D = { D1 , D2 , D3 } are invited to evaluate the criteria and alternative locations. Two are from the transportation planning department in Guiyang city, and the third is a senior manager who operates the logistics center. All three have many years of planning experience, and are very familiar with the alternative location surrounding conditions. After interviewing three top decision makers, we obtained the weights for 16 sub-criteria and the preferences for alternative locations, as shown in Table 2 and Table 3. Wang, Ma, Lao, Wang, Mao 15 TABLE 2 Linguistic Assessments for the Sub-criteria Decision makers Sub-criteria D1 D2 D3 C11 C12 C13 B.P&G B.F&G F F B.F&G F F G G C21 C22 G VG B.F&G B.G&VG G G C23 C31 B.F&G B.G&VG B.G&VG G G B.F&G C32 B.F&G B.F&G B.G&VG C41 C42 B.P&F P B.P&F F B.F&G B.P&F C43 C51 G G B.F&G G G B.F&G C52 B.G&VG G VG C53 C61 B.F&G B.F&G F B.P&F F F C62 F F B.P&F TABLE 3(a) Linguistic Assessments for Alternatives by D1 Linguistic assessments for alternatives by D1 Criteria A1 C11 L A2 A3 A4 A6 A7 M M M A8 A9 L B.L&M L M B.L&M B.L&M M M L B.L&M B.L&M M B.M&H M B.L&M H C12 B.L&M L C13 B.VL&L VL C21 B.H&VH B.H&VH B.M&H H M H VH B.H&VH H B.M&H H H H C23 B.M&H B.M&H H B.H&VH M B.M&H M B.M&H M C31 B.M&H H B.H&VH VH H B.H&VH VH B.H&VH H VH M B.VL&L B.L&M B.L&M L B.L&M B.L&M B.L&M B.L&M M B.M&H M B.H&VH H B.H&VH VH B.H&VH VH VH B.H&VH B.H&VH H B.H&VH B.H&VH B.M&H M B.M&H B.M&H M M B.L&M C43 B.M&H C51 C52 C53 VH C61 B.M&H C11 B.H&VH B.M&H B.M&H M C42 Criteria H H B.M&H B.L&M C41 C62 H B.M&H M A10 B.H&VH B.M&H B.L&M B.M&H A11 A12 H B.H&VH B.H&VH C12 VH VH C13 VH VH VH C21 VL VL B.VL&L C22 VL B.VL&L VL C23 C31 A13 B.H&VH VH M B.VL&L M B.M&H L L B.VL&L B.VL&L B.M&H B.H&VH B.H&VH B.H&VH H B.M&H B.M&H B.H&VH H A18 B.M&H B.VL&L B.H&VH H A17 B.H&VH B.VL&L B.VL&L H B.L&M B.VL&L C42 C62 B.L&M B.VL&L B.VL&L H B.H&VH B.L&M M L B.L&M H B.VL&L L B.VL&L B.M&H L B.L&M B.VL&L B.H&VH C61 B.M&H VL B.M&H L H H M B.VL&L B.VL&L H B.M&H B.M&H L B.VL&L H H H VL C53 B.M&H M VL L H M B.VL&L VL B.M&H B.M&H B.H&VH VL VL B.M&H A16 B.M&H C52 L H B.M&H B.L&M VL B.VL&L M L L B.L&M A15 VL M B.L&M A14 B.L&M B.VL&L L B.M&H C41 C51 L M B.VL&L L B.VL&L L C32 C43 H B.L&M VH H H B.L&M C22 C32 H B.VL&L A5 M L VL B.VL&L H B.H&VH VH B.L&M B.VL&L L B.H&VH VH L B.VL&L B.VL&L VL VL B.VL&L H B.H&VH B.H&VH VH B.L&M B.L&M B.VL&L VL B.H&VH M L VL B.VL&L VH H B.H&VH VH B.H&VH VH H L L B.L&M B.L&M B.VL&L L B.L&M B.VL&L B.VL&L M B.VL&L VH B.H&VH VH B.H&VH L B.H&VH VH B.H&VH VH Wang, Ma, Lao, Wang, Mao 16 TABLE 3(b) Linguistic Assessments for Alternatives by D2 Criteria C11 C12 Linguistic assessments for alternatives by D2 A1 M B.VL&L A2 A3 A4 A5 A6 A7 A8 B.L&M L B.VL&L L L B.L&M M B.VL&L L B.L&M B.L&M B.L&M M B.M&H H H B.M&H L C13 B.VL&L B.VL&L C21 C22 B.H&VH H B.M&H C23 M VH H H B.H&VH B.M&H B.L&M B.L&M L B.M&H B.M&H H M B.H&VH M B.M&H B.M&H A9 M B.L&M M B.M&H H VH VH B.M&H H M H M B.M&H H B.H&VH B.H&VH H H VH B.L&M M C32 VH B.H&VH H B.M&H B.H&VH B.H&VH B.H&VH C41 B.M&H L H B.L&M B.L&M B.VL&L B.L&M M B.M&H B.L&M H B.L&M M L B.L&M B.L&M B.VL&L L H VH M B.M&H H B.M&H B.M&H C31 C42 L C43 B.M&H C51 H C52 VH VH VH B.H&VH B.H&VH H C53 VH B.H&VH H B.H&VH B.M&H H C61 M H B.M&H M C62 B.M&H L M B.M&H Criteria A10 A11 C11 VH C12 B.H&VH C13 VH C21 B.VL&L C22 B.VL&L C23 VL B.VL&L C31 L C32 C41 B.H&VH B.VL&L M B.M&H M B.H&VH M B.M&H B.L&M L L H B.H&VH B.M&H VH VH VH B.M&H B.H&VH H B.H&VH VL VL B.VL&L VL B.VL&L VL VL VL A14 L B.L&M B.VL&L VL M B.H&VH B.H&VH B.M&H H B.H&VH C52 VL H B.M&H B.VL&L VL B.VL&L L A15 B.VL&L VH H L M A18 B.M&H M B.M&H B.M&H L B.VL&L B.VL&L B.VL&L B.VL&L VL B.L&M L L VL L VL VL VL B.VL&L L B.H&VH H B.H&VH VH B.H&VH H VH VH B.VL&L B.L&M L B.L&M VL B.VL&L B.L&M B.M&H M B.VL&L B.VL&L B.L&M B.VL&L L B.M&H B.M&H B.H&VH B.M&H M H H M M VL H A17 B.H&VH C51 H B.VL&L B.L&M C43 M B.L&M B.VL&L VH L B.L&M A16 A13 B.L&M C42 L B.M&H M M A12 B.L&M L L H VH B.VL&L B.VL&L VH B.H&VH B.H&VH B.H&VH L B.VL&L B.L&M L B.VL&L C53 B.VL&L VL L B.VL&L L B.VL&L L B.VL&L M C61 B.H&VH B.M&H H B.M&H H B.H&VH VH B.H&VH VH C62 B.M&H VH VH VH B.H&VH H B.H&VH VH B.H&VH TABLE 3(c) Linguistic Assessments for Alternatives by D3 Criteria C11 Linguistic assessments for alternatives by D3 A1 B.L&M A2 L C12 L B.L&M C13 L B.VL&L A3 B.VL&L L B.L&M C21 VH H VH C22 B.H&VH B.M&H H B.M&H B.H&VH C23 H A4 A5 A6 A7 A8 B.L&M M B.L&M M M L B.L&M M B.L&M B.VL&L B.L&M B.L&M M B.H&VH B.H&VH M H B.M&H B.M&H M H B.H&VH C31 M B.M&H H H C32 B.H&VH H B.M&H B.M&H C41 B.L&M B.L&M B.M&H H C42 L C43 H B.M&H M C51 H B.H&VH C52 VH B.H&VH C53 H C61 H C62 L VH B.M&H M Criteria A10 A11 C11 B.H&VH B.H&VH C12 C13 C21 C22 C23 VH B.H&VH VL B.VL&L VL VH B.H&VH B.VL&L VL VL L B.M&H VH B.H&VH B.H&VH M A12 H L B.L&M B.L&M B.VL&L B.H&VH B.M&H VH B.H&VH B.H&VH B.H&VH L B.L&M B.VL&L B.VL&L M B.L&M L L L L B.L&M B.L&M B.M&H M B.H&VH L B.H&VH B.M&H VH B.L&M B.VL&L VL B.VL&L B.VL&L VL B.VL&L VL B.VL&L B.M&H H M M VL M VH VH H B.M&H B.VL&L A16 B.M&H H H B.M&H M A15 M B.H&VH M B.M&H H A14 B.M&H B.H&VH H L B.VL&L M H VH B.H&VH H B.M&H A18 B.M&H M M M B.M&H L B.VL&L L B.VL&L B.VL&L L B.L&M L L VL VL B.VL&L B.M&H B.VL&L B.VL&L B.L&M C32 B.L&M B.VL&L B.L&M B.VL&L C41 M B.H&VH B.M&H B.H&VH C42 H B.H&VH B.M&H VH H B.H&VH B.H&VH VH C43 B.M&H B.M&H H M VH B.H&VH B.H&VH H C51 VL B.VL&L L L L B.L&M B.L&M B.VL&L B.VL&L B.VL&L M L L C52 B.VL&L C53 L C61 VH C62 B.M&H VL B.VL&L H B.H&VH VL B.VL&L B.L&M B.VL&L VL L B.H&VH B.VL&L VH B.M&H B.H&VH VH H H B.M&H B.H&VH VH B.L&M A17 B.M&H C31 L H B.M&H H H A13 B.M&H H H H VH M B.M&H M B.L&M VH H M B.M&H M B.H&VH B.L&M A9 VL M B.M&H L B.VL&L L L VL B.VL&L VL VH B.H&VH H VH B.H&VH L B.L&M B.VL&L VH B.H&VH B.H&VH VH B.L&M B.VL&L L B.H&VH VH Wang, Ma, Lao, Wang, Mao 17 Result Analysis According to Eq. (6) and Eq. (7), we can obtain the comprehensive evaluation values by integrating the sub-criteria into the major criteria. The evaluation values are presented in Table 4. TABLE 4 Aggregate Avaluation Matrix for Alternatives Aggregate evaluation for alternatives Criteria A1 A2 A3 A4 A5 A6 C1 C2 0.5788 0.8227 0.5727 0.7968 0.5709 0.8257 0.5721 0.8198 0.6270 0.7654 0.6234 0.7634 C3 C4 0.7598 0.6123 0.7616 0.6098 0.7639 0.6179 0.7653 0.6159 0.8167 0.5526 0.8230 0.5450 C5 0.8293 0.8293 0.8295 0.8379 0.7609 0.7665 C6 0.6547 0.6547 0.6615 0.6528 0.5844 0.5836 A7 A8 A9 A10 A11 A12 C1 C2 0.6424 0.7680 0.6409 0.7556 0.6378 0.7659 0.7864 0.5678 0.7781 0.5638 0.7739 0.5680 C3 C4 0.8178 0.5477 0.8244 0.5373 0.8244 0.5519 0.6248 0.6632 0.6173 0.6736 0.6248 0.6680 C5 C6 0.7615 0.5962 0.7711 0.5758 0.7667 0.6106 0.5713 0.7212 0.5718 0.7316 0.5709 0.7259 A13 A14 A15 A16 A17 A18 C1 C2 C3 0.7781 0.5638 0.6209 0.6767 0.6055 0.5641 0.6666 0.6105 0.5591 0.6857 0.6060 0.5651 0.6646 0.6095 0.5641 0.6763 0.6030 0.5646 C4 0.6878 0.7242 0.7256 0.7242 0.7242 0.7286 C5 C6 0.5716 0.7200 0.6182 0.7691 0.6166 0.7691 0.6182 0.7757 0.6169 0.7757 0.6182 0.7753 Following Step 1- Step 3 in the Clustering algorithm procedure, we can obtain the results as follows: ① When threshold α =0.6932 , Iα = 4.49 , there are two clusters: p1 = { A10 , A11 , A12 , A13 } A1 , A2 , A3 , A4 , A5 , A6 , p2 = A7 , A8 , A9 , A14 , A15 , A16 , A17 , A18 ② When threshold α =0.7119 , Iα = 3.87 , there are three clusters: p1 = { A10 , A11 , A12 , A13 } p2 = { A1 , A2 , A3 , A4 } p3 = { A5 , A6 , A7 , A8 , A9 , A14 , A15 , A16 , A17 , A18 } Wang, Ma, Lao, Wang, Mao 18 ③ When threshold α =0.7256 , Iα = 2.51 , there are four clusters: p1 = { A10 , A11 , A12 , A13 } p2 = { A1 , A2 , A3 , A4 } p3 = { A14 , A15 , A16 , A17 , A18 } p4 = { A5 , A6 , A7 , A8 , A9 } ④ When threshold α =0.7659 , Iα = 3.03 , there are five clusters: p1 = { A10 , A11 , A12 , A13 } p2 = { A1 , A2 , A3 , A4 } p3 = { A14 , A15 , A16 , A17 , A18 } p4 = { A5 , A8 , A9 } , p5 = { A6 , A7 } Comparing all of the results, we can find that I 0.7256 = 2.51 is the smallest, thus, the option with four clusters is the best choice. Then, we use TOPSIS approach to rank the alternative locations within each cluster. According to Eq. (14) –Eq. (24), the final ranking results are shown in Table 5 as follows. TABLE 5 Ranking Order for Alternative Distribution Centers Cluster p1 Index d * d - CAi Ranking order Cluster p3 A11 A12 A13 A14 A15 A16 A17 A18 2.978 3.011 3.105 2.942 3.160 3.285 3.216 3.225 3.208 3.021 2.886 2.912 3.015 2.912 2.755 2.845 2.786 2.736 0.503 0.489 0.484 0.506 0.480 0.456 0.469 A10 CA13 > CA10 > CA11 > CA12 Cluster p4 A1 A2 A3 A4 A5 A6 A7 d* 3.296 3.353 3.234 3.121 3.143 3.301 d- 2.905 2.899 2.972 3.075 2.931 CAi 0.468 0.464 0.479 0.496 0.483 Ranking order 0.460 CA14 > CA16 >CA17 > CA18 > CA15 Cluster p2 Index 0.463 CA4 > C A3 > CA1 > C A2 A8 A9 3.269 3.135 3.232 2.915 2.878 2.976 2.883 0.469 0.468 0.487 0.471 CA8 > CA5 > CA9 > CA6 > CA7 Wang, Ma, Lao, Wang, Mao 19 By comparing the CAi values in each cluster (Table 5), A13 , A4 , A14 , A8 are selected as the final urban distribution centers location for the logistics company. According to Eq. (13), we can calculate the dispersion degree between clusters. The ranking results of dispersion degree between clusters is RC34 > RC13 > RC12 > RC24 ; the higher the RChk between clusters is, the smaller the closeness between clusters is. Based on the demand for the economic development in each region, the logistics operators can also merge possible clusters, and finalize the number of distribution centers in each cluster. Algorithm Comparisons Chou et al. (2008) presented a new fuzzy multiple attributes decision-making approach based on the fuzzy set theory, with the added importance of each decision-maker (14). They applied the factor rating system and simple additive weighting approach to evaluate facility location alternatives. Li et al. (2011) presented another comprehensive methodology for the logistics center location selection(34). For comparison purposes, we implemented both algorithms into our case study in the same context with four distribution centers. The algorithm proposed by Chou generates the final locations as A4 , A10 , A11 , A13 , and Li’s algorithm results in the final locations as A10 , A13 , A14 , A16 . They are shown as follows: (a) (b) (c) FIGURE 3 The Results of Three Algorithms: [a] the results of Chou’s algorithm [b] the results of Li’s algorithm [c] the results of the proposed algorithm The proposed algorithm provides a novel approach to address the MDCL problem. The final distribution centers calculated by Chou’s approach (cite) and Li’s approach (cite) are located along one side of the logistics center, which may incur more traffic events near the distribution centers, and increase transportation costs for customers living far away from these distribution centers. The potential distribution centers by our proposed approach, however, are scattered more evenly around the logistics center, which adheres to reality more agreeably. Each attribute of alternative locations in our proposed approach is further split into several finer attributes Wang, Ma, Lao, Wang, Mao 20 during clustering, which gains the heterogeneity between alternative locations, but the algorithms developed by Li and Chou cannot take into account these factors. In summary, the proposed algorithm should be more suitable for MDCL problem in real life. Sensitivity Analysis To further investigate the impacts of different criteria weights on the location selection procedure, a sensitivity analysis is also needed. A total of 27 experiments were undertaken in each cluster. The sub-criteria C43 , C61 and C62 are under the cost criteria, while the remaining sub-criteria are all under benefit criteria. In each cluster experiments 1-9 have the weights of all criteria set equally to VL, B.VL&L, L, B.L&M, M, B.M&H, H, B.H&VH, VH; experiments 10-25 have the weight of one criterion set at the highest weight VH, with the remaining set at the lowest weight VL; experiment 26 have the cost criteria ( C43 , C61 , C62 ) set at the highest weight VH, with the remaining criteria set at the lowest weight VL; and in experiment 27 the weights of the cost criteria are set at the lowest weight VL, with the remaining criteria set at the highest weight VH. The results of the analysis can be observed in Figure 4(a)-(d). Wang, Ma, Lao, Wang, Mao 21 0.56 0.56 A13 A10 A11 A12 0.55 A4 A3 A1 A2 0.54 Closeness Coefficient Closeness Coefficient 0.54 0.53 0.52 0.51 0.52 0.5 0.48 0.5 0.46 0.49 0.48 0 5 10 15 Experiment Times 20 25 0.44 30 0 5 10 (a) 15 Experiment Times 20 25 30 (b) 0.51 0.54 A14 A16 A17 A18 A15 0.5 A8 A5 A9 A6 A7 0.53 0.52 Closeness Coefficient Closeness Coefficient 0.49 0.48 0.47 0.51 0.5 0.49 0.48 0.46 0.47 0.45 0.44 0.46 0.45 0 5 10 15 Experiment Times (c) 20 25 30 0 5 10 15 Experiment Times 20 25 30 (d) FIGURE 4 Results of Sensitivity Analysis: [a] results of sensitivity analysis in cluster P1 [b] results of sensitivity analysis in cluster P2 [c] results of sensitivity analysis in cluster P3 [d] results of sensitivity analysis in cluster P4 Results indicate that for cluster p1 , in the experiments 10, 20, 25 and 26, the final location has been changed to A10 due to CA10 > CA13 ; For cluster p2 , in the experiments 12, 20, 25 and 26, the location has been changed to A3 due to CA3 > CA4 ; For cluster p3 , in the experiments 14, 16, 20, 24, 25 and 26, the location has been changed to A16 due to CA16 > CA14 ; For cluster p4 , in the experiments 13, 19, 20, 24, 25 and 26, the location has been changed to A5 due to CA5 > CA8 . Through these experiments we can say that the location decision is relatively insensitive to benefit criteria weights, however they are sensitive to cost criteria weights. From Figures 7-10, we also observe that the cost criteria weights in clusters p3 and p4 are more sensitive than clusters p1 and p2 , which implies that decision makers need to assess the cost criteria weights more seriously than the benefit criteria for location planning. Wang, Ma, Lao, Wang, Mao 22 CONCLUSIONS This paper develops an innovative approach for selecting multiple distribution center locations under fuzzy environment. The hierarchical criteria analysis structure is initially constructed for planning distribution center locations. A fuzzy set approach is designed to convert the linguistic criteria ratings and alternative ratings into fuzzy numbers, and several relative definitions are also presented for algorithm procedures. Fuzzy integration algorithm is then proposed to synthesizing multiple lower hierarchical criteria into a higher hierarchical criterion followed by an improved AFS approach based fuzzy clustering algorithm to group similar alternative locations into several clusters, and furthermore, a fuzzy TOPSIS approach is integrated to seek the optimal distribution centers locations within clusters. The proposed method has been successfully applied to assisting logistics operators in selecting distribution centers for a logistics company in Guiyang city, China. With further comparisons with other location selection algorithms, our approach has been proven more effective for MDCL problems. A sensitivity analysis is conducted in capturing the impacts of different evaluation criteria weights. The results demonstrate that our approach performs very well as a cost-effective and decision-support tool for distribution center location planning strategies. Additionally, the approach can also be straightforwardly extended to solving other similar multiple facility locations problems for decision makers. ACKNOWLEDGMENT This research is supported by National Natural Science Foundation of China(51078087)and Ph.D. Programs Foundation of Ministry of Education of China (200802861017). Special thanks to Yunteng Lao and Carmen Kwan for their valuable suggestions and revisions. Wang, Ma, Lao, Wang, Mao 23 REFERENCES 1. Yang, L.I., et al., Logistics distribution centers location problem and algorithm under fuzzy environment. Journal of Computational and Applied Mathematics, Vol.208, No.2, 2007, pp: 303-315. 2. Sun, H.J., Z.Y. Gao, and H.J. Wu, A bi-level programming model and solution algorithm for the location of logistics distribution centers. Applied Mathematical Modelling, Vol .32, No.4, 2008, pp: 610-616. 3. Syam, S.S., A model and methodologies for the location problem with logistical components. Computers & Operations Research, Vol.29, No.9, 2002, pp: 1173-1193. 4. Hansen, P., E.D. Pedrosa, and C.C. Ribeiro, Location and Sizing of Offshore Platforms for Oil-Exploration. European Journal of Operational Research ,Vol. 58, No. 2, 1992, pp : 202-214. 5. Lee, C., The multiproduct warehouse location problem: applying a decomposition algorithm. International Journal of Physical Distribution & Logistics Management, Vol. 23, No. 6, 1993, pp: 3-13. 6. Tyagi, R. and C. Das, Manufacturer and warehouse selection for stable relationship in dynamic wholesaling and location problem. International Journal of Physical Distribution & Logistics Management, Vol. 25, No. 6, 1995, pp: 54-72. 7. Hakimi, S.L. and C.C. Kuo, On a General Network Location-Production-Allocation Problem. European Journal of Operational Research, Vol.55, No. 1, 1991, pp: 31-45. 8. Sathe, A. and E. Miller-Hooks, Optimizing location and relocation of response units in guarding critical facilities, Network Modeling, No.1923, 2005, pp:127-136. 9. Eksioglu, S.D., et al., Analyzing Impact of Intermodal Facilities on Design and Management of Biofuel Supply Chain. Transportation Research Record, No.2191,2010, pp:144-151. 10. Shiomi, Y., Y. Seto, and N. Uno, Model for Location of Medical Facility and Evaluation of Vulnerability and Accessibility of Road Network. Transportation Research Record, No. 2234, 2011, pp: 41-48. 11. Kucukaydin, H., N. Aras, and I.K. Altinel, Competitive facility location problem with attractiveness adjustment of the follower A bilevel programming model and its solution. European Journal of Operational Research, Vol. 208, No. 3, 2011, pp: 206-220. 12. Zadeh, L.A., Fuzzy Sets. Information and Control, Vol.8, No.3,1965, pp: 338-353. 13. Chen, C.T., A fuzzy approach to select the location of the distribution center. Fuzzy Sets and Systems, Vol.118, No. 1, 2001, pp: 65-73. Wang, Ma, Lao, Wang, Mao 24 14. Chou, S.Y., Y.H. Chang, and C.Y. Shen, A fuzzy simple additive weighting system under group decision-making for facility location selection with objective/subjective attributes. European Journal of Operational Research, Vol. 189, No.1, 2008, pp: 132-145. 15. Awasthi, A., S.S. Chauhan, and S.K. Goyal, A multi-criteria decision making approach for location planning for urban distribution centers under uncertainty. Mathematical and Computer Modelling, Vol. 53, No. 1-2, 2011,pp: 98-109. 16. Liang, G.-S. and M.-J.J. Wang, A fuzzy multi-criteria decision-making method for facility site selection. International Journal of Production Research, Vol. 29, No. 11, 1991, pp: 2313-2330. 17. Chu, T.C., Facility location selection using fuzzy TOPSIS under group decisions. International Journal of Uncertainty Fuzziness and Knowledge-Based Systems, Vol. 10, No. 6, 2002, pp: 687-701. 18. Lee, D. and D. Donnell, Analysis of nighttime drive behavior and pavement marking effects using fuzzy inference system. Journal of Computing in Civil Engineering, Vol. 21, No. 3, 2007, pp: 200-210. 19. Kahraman, C., D. Ruan, and I. Dogan, Fuzzy group decision-making for facility location selection. Information Sciences, Vol. 157, 2003, pp: 135-153. 20. Chou, C.C., Integrated Short-Term and Long-Term MCDM Model for Solving Location Selection Problems. Journal of Transportation Engineering-Asce, Vol. 135, No. 11, 2009, pp: 880-893. 21. Wey, W.M. and Y.H. Chang, A comparative location study for the joint development station of a mass rapid transit system: a case in Taichung City in Taiwan. Environment and Planning B-Planning & Design, Vol.36, No.4, 2009, pp: 573-587. 22. Yu, J., et al., Locating Urban Transit Hubs: Multicriteria Model and Case Study in China. Journal of Transportation Engineering-Asce, Vol.137, No.12, 2011, pp: 944-952. 23. Negi, D.S., Fuzzy Analysis and Optimization Thesis. Ph.D., Department of Industrial Engineering, 1984. 24. Chen, Z.P. and W. Yang, A new multiple attribute group decision making method in intuitionistic fuzzy setting. Applied Mathematical Modelling, Vol. 35, No. 9, 2011, pp: 4424-4437. 25. Buckley, J.J., Ranking Alternatives Using Fuzzy Numbers. Fuzzy Sets and Systems, Vol.15, No.1, 1985, pp: 21-31. 26. Liu, X.D., The fuzzy theory based on AFS algebras and AFS structure. Journal of Mathematical Analysis and Applications, Vol. 217, No. 2, 1998, pp: 459-478. 27. Zhang, Y.J., D.Q. Liang, and S.C. Tong, On AFS algebra-Part I. Information Sciences,Vol.167,No. 1-4,2004,pp: 263-286. Wang, Ma, Lao, Wang, Mao 25 28. Zhang, Y.J., D.Q. Liang, and S.C. Tong, On AFS algebra-Part II. Information Sciences, Vol.167, No. 1-4, 2004. pp: 287-303. 29. Liu, X.D., The fuzzy sets and systems based on AFS structure, EI algebra and EII algebra. Fuzzy Sets and Systems, Vol. 95, No. 2, 1998, pp: 179-188. 30. Shannon, C.E., A mathematical theory of communication. Mobile Computing and Communications Review, Vol. 5, No. 1, 2001, pp: 3-55. 31. Heilpern, S., Representation and application of fuzzy numbers. Fuzzy Sets and Systems, Vol. 91, No. 2, 1997,pp:259-268. 32. Li, R.J., Fuzzy method in group decision making. Computers & Mathematics with Applications, Vol. 38, No. 1, 1999, pp: 91-101. 33. Liu, X.D., W. Wang, and T. Chai, The fuzzy clustering analysis based on AFS theory. IEEE Transactions on Systems Man and Cybernetics Part B-Cybernetics, Vol. 35,No. 5, 2005, pp: 1013-1027. 34. Li, Y., X.D. Liu, and Y. Chen, Selection of logistics center location using Axiomatic Fuzzy Set and TOPSIS methodology in logistics management. Expert Systems with Applications, Vol. 38, No. 6, 2011, pp: 7901-7908.