An Enhanced Analytical Hierarchical Process for Group Decision Victor B. Krenga, * and Chao-Yi Wub a. Graduate School of Information Management, National Cheng-Kung University, Tainan, Taiwan, ROC 70101 b. Department of Information Management, Southern Taiwan University of Technology, Tainan Hsin, Taiwan, ROC 710 Abstract The Analytical Hierarchical Process (AHP) is a widely used multi-criteria decision making approach and has been successfully applied in different areas. Nevertheless, AHP still has limitations. Based on fuzzy set theory and fuzzy integral, an extension model of AHP is proposed to resolve rank reverse problem, decision uncertainty, and group decision making. Furthermore, for dealing with decision uncertainty, an innovative concept of assurance level is introduced. Supply chain management strategy is used to illustrate the proposed model, which demonstrates the feasibility of this model as well. Keywords: Analytical Hierarchical Process; Group decision making; Fuzzy sets; Fuzzy integral 1. Introduction Analytic Hierarchy Process (AHP) method is one of the most popular Multi-Criteria Decision Making (MCDM) approaches. Since proposed by Saaty (1980), there have been many successful applications in different areas (Vargas, 1990). However, there are many discussions about the limitations in literature (Bryson and Mobolurin, 1994; Ramanathan and Ganesh, 1994; Tavana et al., 1996; Zahir, 1999), especially for the rank reversal problem (Zahir, 1999) and application in group decision making (GDM) (Aczel and Saaty, 1983; Ramanathan and Ganesh, 1994; Van Den Honert and Lootsma, 1996; Van Den Honert, 1998; Forman and Peniwati, 1998;Zahir, 1999; Xu, 2000). Another issue in decision-making is uncertainty, which includes vagueness and ambiguity (Klir and Floger, 1988). Vagueness is associated with the difficulty of making sharp or precise distinctions of boundaries. On the other hand, ambiguity concerns one-to-many relations, in which it is hard to differentiate from two or more alternatives. Many papers extended Saaty’s AHP to fuzzy AHP to propose useful approaches to solve the rank reversal problem (Xu and Zhai, 1992; Mon et al., 1994; Mohanty and Singh, 1994; Lee et al., 2001; Weck et al., 1997). However, the decision uncertainty and the GDM circumstance have not been consider. Within the scope of GDM, there are two kinds of approaches to aggregate individual preferences into the group consensus. One attempts to reach a consensus through intensive discussions by spending a great deal of time among DMs (Bard and Sousk, 1990; Liberatore et al., 1992; Madu and Kuei, 1995; Tavana et al., 1996), where Group Decision Support System (GDSS) is a powerful tool via information technology. The other one, called the decision aggregation model, is to directly synthesize individual judgments into one common preference. Geometric mean approach (GMM) and weighted arithmetic mean approach (WAMM) are the two most popular methods among available literatures (Aczel and Saaty, 1983; Ramanathan and Ganesh, 1994; Van Den Honert, 1998; Van Den Honert and Lootsma, 1996; Forman and Peniwati, 1998; Xu, 2000). However, these models seem to oversimplify the issue of decisions aggregation and overlook the decision uncertainty. The fuzzy integral (Sugeo, 1977) provides a nonlinear function to integrate evidences from different information sources and has been successfully applied to fulfill various objectives (Chen and Chiou, 1999; Pham and Yan, 1996; Tahani and Keller, 1990). In a GDM problem, each DM can be treated as single information source to provide his/her judgment or preference; and the fuzzy integral, can, then, be used to synthesize all the evidences. Accordingly, this study proposes an extension model of AHP based on fuzzy set theory and fuzzy integral to overcome rank reversal, decision uncertainty, and group decision aggregation. 1 2. Fuzzy set theory and assurance level in AHP TFN from the fuzzy set theory is used to derive the preferences of DMs for solving rank reversal problem and improving the ambiguity. The intervals among membership functions of the TFNs in this study are almost equal, the rank reversal problem can, then, be solved. In addition, the ambiguity is improved by that the preferences of DMs are presented in TFN, not a real number. Besides, the assurance level is introduced in this study to enhance the AHP model. While filling out the questionnaires, DMs are not always confident to all the pairwise comparisons, which causes vagueness. Therefore, this study adds one more question to each pairwise comparison to indicate the assurance level with linguistic status of high, medium, and low. Such assurance level can, then, be used to reform the vagueness. The steps to apply TFN and assurance level to Saaty’s AHP approach are described as follows: (1) DMs make the pairwise comparisons according to the hierarchical structure of the problem. As suggested by Saaty, DMs use the seventeen scales, 1/9, 1/8, …, 1, 2, …, 8, 9, to demonstrate their preferences. In addition to the conventional AHP, DMs need to indicate the assurance level to each pairwise comparison with linguistic status of high, medium, and low. (2) The pairwise comparisons by seventeen scales are transferred to TFNs. Let a~ijk be the fuzzy pairwise comparison of criterion i over criterion j from DM k. According to the fuzzy number converting method proposed by Chen and Hwang (1992), the membership functions for the TFNs used in this study are listed as follows: ~ 1 9 : (0, 0, 0.05), ~ 1 3 : (0.35, 0.4, 0.45), ~ 5 : (0.7, 0.75, 0.8), ~ 1 8 :(0, 0.05, 0.1), ~ 1 2 : (0.4, 0.45, 0.5), ~ 6 : (0.775, 0.825, 0.875), ~ 1 7 :(0.05, 0.1, 0.15), ~ 1 : (0.45, 0.5, 0.55), ~ 7 : (0.85, 0.9, 0.95), ~ 1 6 : (0.125, 0.175, 0.225), ~ 2 : (0.5, 0.55, 0.6), ~ 8 : (0.9, 0.95, 1), ~ 1 5 : (0.2, 0.25, 0.3) ~ 3 : (0.55, 0.6, 0.65), ~ 9 : (0.95, 1, 1), ~ 1 4 : (0.275, 0.325, 0.375), ~ 4 : (0.625, 0.675, 0.725). ~ (3) Develop the pairwise comparison matrices. Let A ck be the pairwise comparison matrix from DM k based on criteria/object c, then 2 ~ 1 ~ a ~ ~ A ck = ( a cijk ) = c 21k ~ a cn1k where a~cijk = 1 a~c12k ~ 1 a~cn 2 k ~ 1 ~ a c1nk ~ a~c 2 nk 1 a = c12k ~ ~ 1 1 a c1nk a~c12k ~ 1 ~ 1 a c 2 nk a~c1nk a~c 2 nk , ~ 1 ~ acijk . (4) In order to deal with the uncertainty of assurance, the fuzzy ranking method proposed by Baldwin and Guilds (Chen and Hwang, 1992) is adopted to measure the fuzzy ranking ~ value of a TFN. Each a~cijk in the pairwise comparison matrix will be compared with 1 to derive its fuzzy ranking value. The calculation of fuzzy ranking value is as follows: ~ Supposed that a~cijk {( x m , acijk ( x m ))} and 1 {(x1 , 1 ( x1 ))}, then, the fuzzy ranking value of a~cijk , which is denoted by O cijk O cijk sup {min[ a ( x m ), 1 ( x1 ), P ( x m , x1 )]} , cijk m1 xm , x1 where P m1 The O cijk , is defined as ( x m ) 0.5 ( x1 ) 0.5 , x m x1 , ( x ) 2 ( x ) 2 , 1 m (1) assurance level is low, assurance level is medium, assurance level is high. will be, further, transformed to fit in the interval (0,9), named O cijk , to preserve he same scale with Saaty’s AHP, and be recognized as the representative value ~ for the relative pairwise comparison. The matrix Ack will be defined as ( ̂ O cijk ) and has the form, ˆ O 1 ~ ˆ Oc 21k Ack = ˆ Ocn1k c12k 1 ˆ O cn 2 k (5) The values of ̂ O cijk ˆ O c1nk ˆ O c 2 nk , where ̂ O 1 ̂ O . cijk cjik 1 can be treated as the preference of criterion i over criterion j from ~ DM k. Therefore, the preferences for criterion i in pairwise comparison matrix Ack , 3 denoted as wcik, can be obtained by eigenvector, where wik is grouped into matrix, Wck, according to each DM. Wck = the preference matrix from DM k based on criteria/objective c, = (wcik) = (wc1k, wc2k, …, wcnk). 3. Aggregating individual pairwise comparison matrices by fuzzy integral In this study, the fuzzy integral, which is one of the fuzzy measure approaches, is employed to fuse DMs’ preferences about each pairwise comparison matrix, not the final evaluations to alternatives. Fuzzy measure assigns the value to each alternative subset from the universal set, which signify the degree of evidence or belief that a particular element belongs to the subset. Using the notation of fuzzy measure, Sugeno (1977) defined the concept of fuzzy integral as a non-linear function over measurable sets. The necessary equations are listed as follows. g(AB) = g(A) + g(B) + g(A)g(B), for some >-1. +1 = (2) n (1 g i ) . (3) i 1 n e = max [min( h( x i ), g( Ai ))] , where Ai={x1,x2, … , xi}. (4) i1 The values of g(Ai) can be recursively determined as g(A1) = g({x1})=g1, g(Ai) = gi+ g(Ai-1)+gig(Ai-1) for 1 < i n. (5) Both the explications of h(xi) and gi are the critical issues while implementing fuzzy integral. According to the available literature, h(xi) is represented by the objective evidence; and gi is represented by the weights of information sources while applying fuzzy integral to information fusion (Tahani and Keller, 1990; Pham and Yan, 1996; Chen and Chiou, 1999).In this study, the fuzzy integral is utilized to integrate the judgments of DMs to each pairwise comparison, where h(xi) is interpreted as the DM’s judgment to a pairwise comparison and gi as the assurance level of each DM as well. The details about h(xi) and gi will be discussed as follows. 3.1 Explications of h(xi) for decision fusion With respect to decision fusion, each DM can be regarded as an individual information source. The preference matrix of DM i, Wck, which is obtained by the method proposed in section 2, are treated as the evidence from information source k; moreover, the boundary of wik, 0 wik 1, is identical to the constraint of h(xi). This study, therefore, considers wcik as h(xck)i to reflect the expertise/judgment of DM k, where h(xck)i = wcik = the preference to criterion i from DM k. 3.2 Explications of gi for decision fusion 4 In decision aggregation issue, gi is introduced as the assurance level of DM i to one pairwise comparison matrix, instead of using the degree of importance from single information source. The target of decision aggregation in this study is the pairwise comparison matrix, not the final evaluations to alternatives; thus, each pairwise comparison matrix has an assurance level. Accordingly, each DM can signify his different assurance levels to judgments about various criteria. This let DMs’ assurances to different criteria are completely considered. ~ The assurance levels of DM k to pairwise comparison matrix c ( A ck ), g ck , are computed by the following steps. ~ Step 1: Indicate the assurance level to each pairwise comparison in A ck by each DM with linguistic status of high, medium, and low. Step 2: Transfer the linguistic status into the assurance level, ccijk, with the scale from 0 to 1. In this study, 1 is used to represent the high assurance level, 0.5 as the medium assurance level, and 0 as the low assurance level, where ccijk: the assurance level for the pairwise comparison of criterion i over criterion ~ j in A ck . Step 3: Group these assurance levels of DM k into the assurance matrices, Cck, where Cck = the assurance matrix from DM k based on one criterion/objective, 1 c = (ccijk)= c 21k ccn1k cc12k 1 ccn 2 k cc1nk cc 2 nk . 1 1 Step 4: Compute the maximum eigenvalue, ack , of each assurance matrix, where ack : the maximum eigenvalues of the assurance matrix Cck. Since the interval of ack changes with the size of matrix, the ack has, therefore, to be transferred to a relative value, ack , by equation (6). ack = [ ack -min( ack )]/[max( ack )-min( ack )]. (6) The maximum of ack , max( ack ), occurs when all the elements from assurance matrix are 1, except for those in the diagonal. And the ack reaches its minimum, min( ack ), 5 while all the elements of assurance matrix are 0, except for those in the diagonal. Step 5: Normalize those ack s from all DMs based on criteria/objective c to demonstrate the relative strengths of assurances among DMs. Then, the normalized ack , g ck , is ~ regarded as the assurance level to pairwise comparison matrix A ck . g ck = ack / a cj . (7) all DMs 4. An extension model of AHP based on fuzzy set theory and fuzzy integral The steps of the AHP-based decision aggregation model based on fuzzy set theory and fuzzy integral are summarized as follows: (1) Construct a hierarchical AHP structure for the decision problem. (2) Collect the judgments from DMs by the AHP type questionnaires of pairwise comparisons. In addition, DMs have to fill out the assurance level for each comparison with high, medium, and low. In order to limit the amount of questions been answered, DMs only need to reply to those with the high or low assurance. (3) Derive preference matrixes and assurance levels of all DMs for all pairwise comparison matrices. ~ The following steps are for the A ck , and will repeat for all pairwise comparison matrices. (3.1) Find individual preference matrices, h(xck)i, of DMs by the steps in section 2. (3.2) Calculate the assurance levels of DMs, g ck , by the steps listed in section 3.2. (3.3) Calculate g(Aci) about criteria i. (i) Use equation (3) to find c. (ii) Rearrange g ck in the decreasing order of h(xck)i for criteria i. (iii)Calculate g(Aci) by equation (5). (3.4) Fuse the evaluations from DMs to one criterion/objective by fuzzy integral. Based on equation (4), the various preferences from DMs can be integrated into a single preference of each criterion. (4) Synthesize the weights of all criteria through the AHP hierarchical structure to obtain the final weights of all alternatives. (5) Choose the alternative with the highest value as the decision. 5. An illustrative example The problem to identify a Supply Chain Management (SCM) strategy is used, here, to illustrate this model. To integrate and perform both logistics and manufacturing activities 6 effectively is the main objective of SCM. According to Pagh and Cooper (1998), postponement and speculation strategies both offer opportunities to achieve a timely and cost-effective delivery by rearranging the conventional production and logistics structures. In addition, four strategies have been further developed accordingly, which are Full Speculation Strategy (FSS), Manufacturing Postponement Ptrategy (MPS), Logistics Postponement Strategy (LPS), and Full Postponement Strategy (FPS). In this study, the framework of SCM strategy selection is also based on the study of Pagh and Cooper (1998), which is shown in figure 1. Since there are numerous computations for such decision, only those data, which are related to the criteria in the first level of hierarchical AHP structure, are listed for illustration. Best strategy of SCM Market Manufacturing 因素 Product Cooperation Demand Logistic Product Life-cycle Production Modulization Information Power technology uncertainty capability type stage sharing Full Speculation Strategy Figure 1 Manufacturing Postponement Strategy Logistics Postponement Strategy Full Postponement Strategy The hierarchical structure for decision of SCM strategy. The pairwise comparisons from DM 1, including those relate to the criteria in the first level of hierarchical AHP structure, are listed in table 2. ~ ~ Accord to the data in table 2, the matrices, Ac1 and Ac1 , are as follows: ~ 1 ~ 1 5 ~ Ac1 = ~ 1 / 6 ~ 1 / 3 ~ ~ ~ 5 6 3 ~ ~ ~ 1 1 / 4 1 / 3 , ~ ~ ~ 4 1 4 ~ ~ ~ 3 1/ 4 1 6.457 6.436 4.881 1 2.377 1 2.679 3.242 ~ Ac1 = . 1.551 5.309 1 5.633 1 3.242 4.881 2.770 The assurances matrix of DM 1 to the first level is 7 1 0 .5 1 1 1 0 .5 Cc1= 0 .5 0 .5 1 1 1 1 1 1 . 1 1 Thus, the DM 1s’ assurance level is derived as g c1 = 0.845. In addition, the decision aggregation process is shown in table 3. Then, FSS with the highest value (0.4050) is selected as the decision. Table 2 Examples of transformation from pairwise comparisons to the relative fuzzy ranking values. Target of comparison Assurance Fuzzy Ranking level Value ( Oijk ) High Medium 6.457 6.436 High 4.881 Medium 2.679 ~ 13 High 3.242 ~ 4 High 5.633 Product vs. Market ~ 15 High 2.377 Manufacturing vs. Market ~ 16 Medium 1.551 Cooperation vs. Market ~ 13 High 3.242 Medium 5.309 High 4.881 High 2.770 Market vs. Product Market vs. Manufacturing Market vs. Cooperation Product vs. Manufacturing Product vs. Cooperation Manufacturing vs. Cooperation Manufacturing vs. Product Cooperation vs. Product Cooperation vs. Manufacturing TFN ( a~ijk ) ~ 5 ~ 6 ~ 3 ~ 14 ~ 4 ~ 3 ~ 14 8 Table 3 The process of the decision fusion. Criterion Marketing Production Manufacturing Cooperation DM gi g(Ai) h(xi) min[g(Ai),h(xi)] e 1 0.845 0.845 0.335 0.335 0.335 4 0.582 0.938 0.314 0.314 2 0.683 0.984 0.306 0.306 3 0.757 1.000 0.203 0.203 3 0.757 0.757 0.264 0.264 4 0.582 0.901 0.249 0.249 2 0.683 0.972 0.207 0.207 1 0.845 1.000 0.189 0.189 2 0.683 0.683 0.300 0.300 3 0.757 0.926 0.271 0.271 1 0.845 0.993 0.245 0.245 4 0.582 1.000 0.208 0.208 3 0.757 0.757 0.263 0.263 1 0.845 0.966 0.232 0.232 4 0.582 0.989 0.229 0.229 2 0.683 1.000 0.187 0.187 0.264 0.300 0.263 6. Conclusion This study is proposed to overcome rank reversal problem, decision uncertainty, and group decision of AHP. A systematic methodology by using fuzzy set theory and fuzzy integral is demonstrated. In addition, the assurance level on each pairwise comparison is introduced to improve the quality of decision. The proposed model possesses the following features. (1) The pairwise comparisons have been converted to TFNs, and, then, coupled with assurance levels to further transform to become crisp fuzzy ranking values. Such transformation not only improves the rank reversal problem, but also deals with ambiguity and vagueness during decision-making process. (2) The aggregation of the proposed model occurs on individual pairwise comparison, not final preferences of alternatives. This makes DMs’ assurances, which are likely different for various criteria, be entirely deliberated in decision fusion. (3) While comparing with the average methods, such as GMM and WAMM, the proposed aggregation model based on fuzzy integral provides a nonlinear function to fuse 9 preferences from various DMs instead of using simplified mathematic average equation. Accordingly, such model can offer decision with better quality. 7. References Aczel, J., Saaty, T.L., 1983. Procedures for synthesizing ratio judgments, Journal of Mathematical Psychology 27 (1) 93-102. Bard, J.F., Sousk, S.F., 1990. A tradeoff analysis for rough terrain cargo handlers using the AHP: An example of group decision making, IEEE Transactions on Engineering Management 37 (3) 222-228. Basak, I., 1998. Probabilistic judgments specified partially in the Analytic Hierarchy Process, European Journal of Operational Research 108 (2) 153-164. Bryson, N., Mobolurin, A., 1994. An approach to using the Analytic Hierarchy Process for solving multiple criteria decision making problems, European Journal of Operational Research 76 (3) 440-454. Chen, L.H., Chiou, T.W., 1999. A fuzzy credit-rating approach for commercial loans: A Taiwan case, Omega 27 (4) 407-419. Chen, S.J., Hwang, C.L., 1992. Fuzzy Multiple Attribute Decision Making, Springer-Verlag, Berlin. Forman, E., Peniwati, K., 1998. Aggregating individual judgments and priorities with the Analytic Hierarchy Process, European Journal of Operational Research 108 (1) 165-169. Klir, G.J., Floger, T.A., 1988. Fuzzy Sets, Uncertainty, and Information, Prentice-Hall. Lee, W.B., Lau, H., Liu, Z.Z., Tam, S., 2001. A fuzzy analytic hierarchy process approach in modular product design, Expert Systems 18 (1) 32-42. Leszczynski, K., Penczek, P., Grochulski, W., 1985. Sugeno’s fuzzy measure and fuzzy integral, Fuzzy Sets and Systems 75 (2) 147-158. Liberatore, M.J., Nydick, R.L., Sanches, P.M., 1992. The evaluation of research papers (On how to get an academic committee to agree on something), Interfaces 22 (2) 92-100. Madu, C.N., Kuei, C.H., 1995. Stability analyses of group decision making, Computers industry Engineering 28 (4) 881-892. Mohanty, B.K., Singh, N., 1994. Fuzzy relational equations in analytical hierarchy process, Fuzzy Sets and Systems 63 (1) 11-19. Mon, D.L., Cheng, C.H., Lin, J.C., 1994. Evaluation weapon system using fuzzy analytic hierarchy process based on entropy weight, Fuzzy Sets and Systems 62 (1) 127-134. Pagh, J.D., Cooper, M.C., 1998. Supply chain postponement and speculation strategies: How to choose the right strategy, Journal of Business Logistics 19 (2) 13-33. Pham, T.D., Yan, H., 1996. Information fusion by fuzzy integral, Proceeding 1996 Australian 10 New Zealand Conference on Intelligent Information Systems 18-20. Ramanathan, R., Ganesh, L.S., 1994. Group preference aggregation methods employed in AHP: an evaluation and an intrinsic process for deriving members’ weightages, European Journal of Operational Research 79 (2) 249-265. Saaty, T. L., 1980. The Analytic Hierarchy Process: Planning, Priority Setting, Resource Allocation, RWS Publications, Pittsburgh. Sugeo, M., 1977. Fuzzy measures and fuzzy integrals. In Gupta, M.M., Saridis, G.N., and Gaines, B.R. (ed.) A Survey in Fuzzy Automata and Decision Processes, North-Holland. Tahani, H., Keller, J.M., 1990. Information fusion in computer vision using the fuzzy integral, IEEE Transactions on Systems, Man, and Cybernetics 20 (3) 733-741. Tavana, M., Kennedy, D.T., Joglekar, P., 1996. A group decision support framework for consensus ranking of technical mangemer candidates, Omega 24 (5) 523-538. Van den Honert, R.C., 1998. Stochastic group preference modeling in the multiplicative AHP: A model of group consensus, European Journal of Operational Research 110 (1) 99-111. Van Den Honert, R.C., Lootsma, F.A., 1996. Group preference aggregation in the multiplicative AHP: the model of the group decision process and Pareto optimality, European Journal of Operational Research 96 (2) 363-370. Vargas, L.G., 1990. An overview of the analytic hierarchy process and its applications, European Journal of Operational Research 48 (1) 2-8. Weck, M., Klocke, F., Schell, H., Ruenauver, E., 1997. Evaluating alternative production cycles using the extended fuzzy AHP method, European Journal of Operational Research 100 (2) 351-366. Xu, R., Zahir, X, 1992. Extensions of the analytic hierarchy process in fuzzy environment, Fuzzy Sets and systems 52 (3) 251-257. Xu, Z., 2000. On consistency of the weighted geometric mean complex judgment matrix in AHP, European Journal of Operational Research 126 (3) 683-687. Zahir, S., 1999. Geometry of decision making and the vector space of formulation of the analytic hierarchy process, European Journal of Operational Research 112 (2) 373-396. 11