Vous êtes sur la page 1sur 9

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 33, NO.

1, FEBRUARY 2003

33

Fuzzy Analytic Network Process and its Application to the Development of Decision Support Systems
Ludmil Mikhailov and Madan G. Singh, Fellow, IEEE
AbstractIn this paper we propose a fuzzy extension of the analytic network process (ANP) that uses uncertain human preferences as input information in the decision-making process. Instead of the classical Eigenvector prioritization method, employed in the prioritization stage of the ANP, a new fuzzy preference programming method, which obtains crisp priorities from inconsistent interval and fuzzy judgments is applied. The resulting fuzzy ANP enhances the potential of the ANP for dealing with imprecise and uncertain human comparison judgments. It allows for multiple representations of uncertain human preferences, as crisp, interval, and fuzzy judgments and can find a solution from incomplete sets of pairwise comparisons. An important feature of the proposed method is that it measures the inconsistency of the uncertain human preferences by an appropriate consistency index. A prototype decision support system realizing the proposed method is developed, and its performance is illustrated by examples. Index TermsAnalytic network process, decision support systems, fuzzy programming, multiple criteria decision making.

I. INTRODUCTION HE analytic hierarchy process (AHP) [1] is widely used for multi-criteria decision-making and has successfully been applied to many practical decision-making problems. The basic assumption of the AHP is that the decision-making problem can be decomposed in a linear top-to-bottom form as a hierarchy, where the upper levels are functionally independent from all lower levels, and the elements in each level are also independent. However, many decision-making problems cannot be structured hierarchically, or there would exist strong interactions and dependencies between inter-level and/or intra-level elements. In order to overcome this limitation, Saaty proposes a supermatrix approach [2] and extends the AHP to problems with dependencies and feedback. The resulting analytic network process (ANP) generalizes the AHP and provides a framework for dealing with decision-making problems within which assumptions about dependencies between criteria and alternatives are unnecessary. Similarly to the AHP, the priorities in the ANP (the weights of criteria and the scores of alternatives) are assessed indirectly from pairwise comparison judgments. The pairwise comparison process assumes that the decision maker can compare any two and , and provide a numerical value of the elements ratio of their importance.
Manuscript received September 30, 2002; revised October 22, 2002 and December 17, 2002. This paper was recommended by Guest Editors K. W. Hipel and N. P. L. Cassaigne. The authors are with the Computation Department, the University of Manchester Institute of Science and Technology (UMIST), Manchester M60 1QD, U.K. (e-mail: l.mikhailov@ co.umist.ac.uk). Digital Object Identifier 10.1109/TSMCC.2003.809354

However, in many cases the preference model of the human decision maker is uncertain, and it is relatively difficult for the decision maker to provide exact numerical values for the comparison ratios. The decision makers could be uncertain about their own level of preference, due to incomplete information or knowledge, complexity and uncertainty within the decision environment, or a lack of an appropriate measurement units and scale. A natural way to cope with uncertain judgments is to express the comparison ratios as intervals, or fuzzy sets, which incorporate the vagueness of human thinking. When comparing two could be represented elements, the uncertain numerical ratio . Then an appropriate as an interval or a fuzzy ratio about prioritization procedure can be applied to derive local priorities, approximately satisfying the uncertain judgments. Some existing interval and fuzzy prioritization methods are briefly discussed in Section II. Although there have been many attempts to modify the AHP so that it can deal with uncertain human judgments, the ANP still copes with crisp comparison ratios only. The main reason is that the aggregation procedure in the AHP is very simple and can be performed on interval or fuzzy local priorities, whereas the supermatrix priority derivation process in the ANP requires complex matrix operations on real numbers. However, all known interval and fuzzy prioritization methods derive interval or fuzzy local priorities, which cannot be further used in the matrix calculations of the ANP. A new fuzzy preference programming (FPP) method was recently proposed by Mikhailov and Singh [3] for deriving crisp weights and scores from interval pairwise comparison judgments. The FPP method represents the prioritization process as an intersection of fuzzy sets, describing the uncertain humans preferences and determines a crisp priority vector, maximizing the degree of decision-makers satisfaction. The method transforms the prioritization problem into a fuzzy linear programming problem that is solved as a standard linear program. The interval FPP method can easily be modified for deriving crisp priorities from fuzzy or crisp comparison ratios. Linear and nonlinear modifications of the FPP method for deriving priorities from fuzzy pairwise comparisons are described in [4]. The approach, presented in Section IV of this paper, transforms the initial fuzzy judgments into a series of interval judgments, and employs the interval FPP method. A simple aggregation procedure is used for obtaining crisp overall values of the local priorities. As the crisp comparison judgments can be considered as interval judgments with equal lower and upper bounds, the FPP method can straightforwardly be applied for deriving priori-

1094-6977/03$17.00 2003 IEEE

34

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 33, NO. 1, FEBRUARY 2003

ties from crisp comparison matrices [5]. An extensive comparison study is carried out by Mikhailov and Singh [6], where the FPP method is compared to the most popular crisp prioritization methods, using a large set of randomly generated inconsistent pairwise comparison matrices. This study shows that the FPP method outperforms some of the existing prioritization methods, especially when the decision-makers preferences are strongly inconsistent. The application of the FPP method to deriving priorities from interval and fuzzy comparison judgments eliminates some of the drawbacks of existing prioritization methods. This new approach does not require a construction of comparison matrices and can derive priorities from an incomplete set of judgments. The fuzzy extension of the method is invariant to the specific form of the fuzzy sets used to represent the judgments, and can be applied when the decision-maker provides different types of comparison judgments. Moreover, the FPP method provides an appropriate index for measuring the inconsistency of the uncertain human judgments. The main purpose of this paper is to apply the FPP method to the ANP for deriving priorities from different types of uncertain ratio scale judgments, thus increasing the potential of the ANP for decision-making under uncertainty. A prototype decision support system that facilitates the application of the fuzzy ANP is developed and its performance is illustrated. II. DERIVING PRIORITIES FROM UNCERTAIN JUDGMENTS The problem of deriving priorities from uncertain judgments has been studied by many researchers and different solution methods have been proposed. These methods can be classified in two main groups, according to the representation of the uncertainty used. The first group of methods represents the uncertain judgments as interval ratios, while the second uses fuzzy sets or fuzzy numbers. A. Interval Judgments Saaty and Vargas [7] represent the uncertain comparison raas intervals , where and are the tios lower and the upper bounds of the corresponding uncertain judgment, and construct a number of interval comparison matrices of the type

from the upper triangular part of the reciprocal interval matrix (1) as linear constraints on the priorities and formulates the prioritization process as a linear programming model. Arbels prreference programming method identifies a convex feasible region that encloses all priority vectors , associated with the initial set of interval judgments

(2) Then the method finds successively all vertices of by implementing a technique based on the pivoting operation in linear programming. However, if the interval judgments are inconsistent, i.e., there exist no such priorities, satisfying simultaneously all interval constraints in (2), then the feasible region is an empty set and the preference programming method cannot obtain a solution, as discussed by Kress [10]. The above problem can be resolved by the FPP method [3], described briefly in Section III. This method extends the feasible region (2) into a convex nonempty fuzzy feasible region and derives crisp priorities, approximately satisfying the inconsistent interval judgments. B. Fuzzy Judgments Some researchers [11][15] represent uncertain judgments as . Van Laarhoven and Pedrycz fuzzy sets or fuzzy numbers [16] modify the AHP to a fuzzy hierarchical analysis, using comparison matrices with triangular fuzzy numbers. They obby applying a fuzzy tain fuzzy priorities , version of the Logarithmic least squares method. Wagenknecht and Hartmann [17] also employ the , least squares method and calculate fuzzy priorities , which approximate the fuzzy ratios so that . The obtained fuzzy priorities are represented as (LR) fuzzy numbers. In the same paper the authors propose another approach where the fuzzy priorities are calculated from a fuzzy comparison matrix using the geometric mean. The approach proposed by Buckley [12] is rather similar to the second approach of Wagenknecht and Hartmann. Buckley represents the uncertain judgments by fuzzy trapezoidal numbers, claiming that they are more easily understood and used by experts. The prioritization process is also based on the geometric mean. The derived fuzzy local priorities are then combined in Saatys hierarchy to compute the final fuzzy scores, which are further compared by fuzzy ranking. All known interval and fuzzy prioritization methods derive interval or fuzzy priorities, so they cannot directly be used in the ANP. The FPP method, however, is able to obtain crisp priorities from interval and fuzzy judgments, so it can easily be applied within the ANP for decision-making under uncertainty. III. FUZZY PREFERENCE PROGRAMMING METHOD The main idea behind the FPP method [3], proposed initially for solving inconsistent interval problems, is to define fuzzy sets which represent the decision makers satisfaction with the

(1) They suppose that the priorities, corresponding to the interval matrices are also intervals and formulate the prioritization problem as determining the bounds of the variation of the priorities. To solve the problem, the authors randomly sample values from the judgment intervals, compute the eigenvectors of the resulting crisp matrices, and construct confidence intervals for each of the components. Obviously, such an approach is rather intractable from a computational point of view. An original approach to handle interval preferences is proposed by Arbel [8], [9]. He interprets the interval judgments

MIKHAILOV AND SINGH: FUZZY ANALYTIC NETWORK PROCESS

35

fulfillment of the interval judgments. Introducing some tolerance parameters, Arbels feasible region (2) is transformed into a nonempty fuzzy feasible area. Then the prioritization problem is formulated as a maximin fuzzy programming problem, which finds a crisp priority vector, maximizing the overall decisionmakers satisfaction with the final solution. Consider the interval judgments from the upper triangular part of the reciprocal pairwise comparison matrix (1). When those judgments are consistent, there are many priority vectors, whose elements satisfy the inequalities

The solution to the prioritization problem by the FPP method is based on two additional assumptions [3]. The first one requires the existence of nonempty fuzzy feasible area on the simplex hyperplane (8) The membership function of the fuzzy feasible area is expressed as the intersection of all interval membership functions (7), i.e.

(3) In the inconsistent cases, however, there is no such priority satisfying all interval judgments sivector multaneously. But it is reasonable to try to find a vector that satisfies all judgments as well as possible. This means that a good enough solution has to satisfy all judgments approximately, keeping the violations close to zero, or (4) where the symbol denotes the statement fuzzy less or equal to. In order to represent the prioritization problem as a linear function of all decision variables, the FPP method transforms the double-side inequalities (4) into a set of single-side linear fuzzy inequalities

(9) The second assumption of the FPP method specifies a selection rule, which determines a priority vector having the highest degree of membership in the aggregated membership function (9). It can easily be proved [3] that is a convex set, so there is that has a maximum degree always a priority vector in of membership .

(10) The maximin prioritization problem (10) can be represented as the following fuzzy programming problem: maximize subject to

(5) The above set of form as fuzzy constraints is given in a matrix (11) (6) , . where the matrix represents a fuzzy The -th row of (6), denoted by linear constraint and might be defined by a linear membership function of the type (7) is a tolerance parameter, defined by the decision where maker, which corresponds to the admissible interval of ap, proximate satisfaction of the crisp inequality . The membership function (7) represents the decision-makers satisfaction with the fulfilment of the single-side constraints (5). is equal to zero when the corresponding The value of is strongly violated. The degree of crisp constraint membership takes values between zero and one when the crisp constraint is approximately satisfied and it is greater than one if the constraint is fully satisfied. The optimal solution to the above problem is a vector , whose first component represents the priority vector that maximizes the degree of membership in the fuzzy feasible area, whereas, the second one gives the value of the maximum . The value of measures the degree degree, of satisfaction and is a natural indicator for the inconsistency of the decision-makers judgments, so it is called a consistency index. When the human interval judgments are consistent, is greater than or equal to one. For inconsistent judgments, takes a value between one and zero, the consistency index depending on the degree of inconsistency and the values of the tolerance parameters . It can be shown that the corresponding represents a point in the feasible area maximizing solution are equal to . for which some of the ratios The FPP method transforms the prioritization problem into a linear program (11) that can easily be solved by the standard simplex technique. For a practical implementation of this method, all tolerance parameters can be set equal, because the decision-makers usually have no preferences about their specific pairwise comparison judgments. It should be noted, that

36

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 33, NO. 1, FEBRUARY 2003

the values of the tolerance parameters if they are equal, do not affect the value of the maximizing solution .

These considerations suggest that the value of can be used as a weighting factor of the solutions, so we can obtain aggregated values of the priorities by a weighted sum of the type

IV. EXTENDING THE FPP METHOD TO FUZZY JUDGMENTS Kaufmann and Gupta [18] introduce the concept of -cuts to decompose fuzzy numbers into a number of intervals, adequately representing the initial fuzzy sets. Other researchers further employ this concept in the decision-making process to transform the overall weighted fuzzy performance matrix and rank alternatives, applying interval calculations [14], [15], [19]. The approach, proposed initially in [4] is rather different. It decomposes the fuzzy comparison ratios into intervals and finds priorities for each -level cut, which are further aggregated in crisp local and global priorities. This approach, which can easily be implemented in the ANP, is described below. A. Decomposing Fuzzy Judgments Suppose that the decision-maker can provide a set of fuzzy comparison judgments, represented or fuzzy numbers. as normal convex fuzzy sets The crisp sets of the ratios between the unknown crisp priorithat belong to the fuzzy judgment to degree ties , and are of are called -level sets (or simply -cuts) of . defined as can be repreUsing this concept, each fuzzy judgment , , sented as a sequence of interval sets . For , the where corresponding -level interval represents the support of the , while is the core of . Since the fuzzy judgment fuzzy judgments are normal convex fuzzy sets or triangular , the -level sets are closed fuzzy numbers , intervals, such that , where and are the lower and the upper bounds of the corresponding intervals. By applying -cuts, the initial set of fuzzy comparisons can be converted into a series of interval sets . B. Aggregation of the Priorities Decomposing the fuzzy comparison judgments into a series of interval judgments by -cuts and applying the FPP method, we can obtain a sequence of crisp priorities, corresponding to , each -cut level , . The value of could be considered as a degree of confidence by the decision-maker regarding his comparison judgments [4]. A small value of yields a construction of interval judgments, having large spreads, which indicates a high level of uncertainty and less trustworthy assessments. Increasing the level of the -cuts, we get smaller spreads of the intervals, whose upper and lower bounds have a greater degree of membership in the initial fuzzy sets. When the fuzzy judgments are represented as trianproduces a set gular fuzzy numbers, the cutting at level of the of crisp judgments, equal to the most possible value . fuzzy comparison ratios (12)

where

denotes the aggregated weight of the -th priority. V. FUZZY ANALYTIC HIERARCHY PROCESS

In its present form, the ANP employs the classical Eigenvector method for deriving crisp local priorities from crisp comparison matrices [2], [20]. The obtained priorities are further combined into matrices of real numbers, and complex matrix calculations are performed to find the global priorities of the alternatives. It is essential to emphasize that the ANP cannot be applied if the elements of the matrices are intervals or fuzzy numbers. As already noted, all known interval and fuzzy prioritization methods, which are suitable for representing the uncertainty in the decision-making process, derive interval or fuzzy local priorities. However, the proposed FPP method, which derives crisp priorities from interval and fuzzy judgments, can be easily applied to increase the capabilities of the ANP for dealing with inconsistent and uncertain judgments. The process of applying the fuzzy ANP (FANP) that combines the FPP and the ANP comprise of the following main steps: Step 1) Regarding the decision goal, configure a hierarchical or network structure including clusters, criteria, subcriteria, lower elements and alternatives. Step 2) Identify the dependences among all components of the previous structure and list them in a table in order to define the impacts between each. Step 3) Construct pairwise comparison matrices of the components with crisp, interval, or fuzzy ratio judgments. Step 4) Derive relative importance weights (local priorities) from each matrix using the FPP method. Step 5) Fill the supermatrix with the elicited weights in order to form an unweighted supermatrix. Step 6) Adjust the supermatrix to column stochastic so that the sum of the elements in each column is equal to one (producing a weighted supermatrix). Step 7) Limit the weighted supermatrix by raising it to a sufficiently large power (where is an arbitrarily large number) until it converges into a stable supermatrix (also called a limiting or steady-state supermatrix). Step 8) Aggregate the weights of criteria and the scores of alternatives into final priorities by multiplying the scores by the weights of the control criteria. Steps 3 and 4 are explained in Sections III and IV of this paper, whereas all the other steps are discussed in detail in [2], and illustrated in Sections VI and VII.

MIKHAILOV AND SINGH: FUZZY ANALYTIC NETWORK PROCESS

37

Fig. 1. Structure of the FDSS.

VI. DEVELOPMENT OF A FUZZY DECISION SUPPORT SYSTEM BASED ON THE FANP We have adopted both Excel and MATLAB as the development environments for prototyping a fuzzy decision support system (FDSS), because these environments provide powerful numerical functions, optimization procedures, good visualization capabilities and programming interfaces. In addition, the MATLAB environment allows for an easy communication with Excel, as well as with other applications. The architecture of our prototype system is composed of two main components: macros of Excel worksheets and M-files of MATLAB (see Fig. 1). The problem statement, the initial configuration of the FDSS, the manipulation of input and final matrices, and the synthesis of the final priorities are all done in the Excel workspace. MATLAB is used for implementing the optimization procedure of the FPP method and for deriving local priorities, as well as for some complex matrix operations, such as raising the supermatrix to powers. Some Excel macros are also developed for facilitating the users interfaces of the FDSS system. The FDSS workbook consists of two worksheets named Input Matrix and Supermatrix, which realize most functions and solve the main problems of implementing the FANP. The worksheets also offer flexibility to the decision-maker when manipulating matrices and graphics, and synthesizing the final scores of alternatives. A. Input Matrix Worksheet This worksheet is specially designed to derive weights by the FPP. Crisp, interval, triangular, and trapezoidal judgments are the four acceptable types of pairwise comparison judgments in the FDSS prototype. The pairwise judgments are entered by the user or loaded from an input file. The system converts each set of pairwise comparisons into a MATLAB readable format and forms a ma, where is the number of the judgments, and trix represents the type of the judgments. After obtaining acceptable formats for all input matrices, the following steps for deriving weights are carried out. Step 1) Enter the pairwise comparison matrix FM into the bordered area of the Input Matrix Worksheet. Step 2) Set the tolerance parameter (only one parameter is required for all judgments, as discussed in Section III), and execute the deriving weights procedure. The set of the derived weights and the corresponding consistent index are visualized on the screen. If the

Fig. 2.

Flowchart of the deriving weights procedure.

Fig. 3. Holarchy example: forecasting the turnaround in economic stagnation.

user needs more sets of weights regarding different values, he/she should confirm the results already obtained in the waiting table first, and then redo this step without changing the value. Step 3) The desired overall weights, regarding the control criteria are calculated by executing the aggregation procedure. The cooperation between MATLAB and EXCEL in the process of deriving priorities is shown as a flow diagram in Fig. 2. B. Supermatrix Worksheet This worksheet is designed to raise a weighted supermatrix to powers and to obtain a limiting supermatrix. The following five-step procedure is carried out. Step 1: Decide the size of the supermatrix. Step 2: Configure an unweighted supermatrix. The clusters and elements included in this supermatrix are determined by feedback within the ANP structure. Step 3: Fill in the supermatrix with the derived priorities of elements regarding their control elements on the top. Assign a zero for those elements having no influence on others.

38

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 33, NO. 1, FEBRUARY 2003

TABLE I IDENTIFYING INFLUENCES IN THE NETWORK STRUCTURE

Fig. 4.

Relative weights with respect to monetary policy.

Step 4: Adjust the unweighted supermatrix by the weights of the clusters to make it column stochastic. Thus, a weighted supermatrix is obtained. Step 5: Raise the weighted supermatrix to large powers until all elements in each row converge. The obtained limiting supermatrix is used for the calculation of final global priorities. The limiting priorities of the supermatrix are calculated according to whether it is reductable, primitive or cyclic. These

TABLE II INTERVAL RELATIVE INFLUENCE OF MONETARY POLICY ON PERIODS OF ADJUSTMENT

MIKHAILOV AND SINGH: FUZZY ANALYTIC NETWORK PROCESS

39

Fig. 5.

Limiting supermatrix for interval judgments.

properties of the supermatrices are explained in detail in [2], together with corresponding conditions for their evaluation. VII. NUMERICAL ILLUSTRATION In this section, the concepts of the FANP and the FDSS prototype, discussed in Sections IIVI are illustrated and clarified via the case study Turnaround of the US economy, given by Saaty in [2].

TABLE III COMPARING THE RESULTS OF THREE METHODS

B. Solving the Problem by the FDSS A. Turnaround of the US Economy Problem This case study concerns a governments need to forecast the strength of an economic recovery and, in particular, whether the duration of an economic turnaround will be about three, six, twelve, or twenty-four months. The goal is impacted by two primary criteria: conventional adjustment and economic restructuring. The first criterion is decomposed into six sub-criteria: consumer spending, investment spending, exports, confidence, fiscal policy, and monetary policy. The last criterion is further divided into three sub-criteria: financial sector, defense posture, and global competition. However, with respect to a specific period of time, the conventional adjustment and economic restructuring have different influences on the turnaround. Hence, there is a feedback relationship between the bottom level and the first level of hierarchy, shown in Fig. 3. The solution to this holarchy example by the FDSS prototype follows the eight main steps, discussed in Section V of the paper. The first and second steps identify the structure of the problem and the relationships among all components together with their influences (see Table I). Case 1: Interval Judgments: All interval judgments are generated by modifying the crisp pairwise comparison matrices, given in [2]. The original crisp matrices are converted into interval judgments by multiplying by 0.8 and 1.2 the values of the elements in their upper triangular parts. The overall number of comparisons matrices for this example is 15, as indicated in Table I. For example, the relative importance of targeted time periods for monetary policy to drive a turnaround is shown in Table II. After entering the upper triangular interval judgments, the Input Matrix Worksheet calculates the corresponding priorities

40

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 33, NO. 1, FEBRUARY 2003

Fig. 6. Eleven sets of derived weights and the aggregation results.

and the value of the consistency index. The values obtained by the FPP method for the matrix, shown in Table II are , , , , , as illustrated in Fig. 4. From here we can conclude that with respect to the monetary policy, a three month period is the most likely length of a turnaround, because it has the highest weight of 0.6545, followed by a six month period. The relative weights for all other comparison matrices are calculated in a similar manner. The build supermatrix procedure defines the size and the inner structure of the supermatrix identified in Table I. The weights derived by the FPP method in the previous step are stored in the corresponding cells to form an unweighted supermatrix, while zeros are assigned to each element having no influence. The unweighted supermatrix is normalized to become column stochastic and is then raised to a sufficiently large , all power until it converges. In this specific case, for columns turn out to be identical. Hence, the limiting supermatrix is in a steady and convergent state. The limiting outcome is shown in Fig. 5. Finally, the global scores of the alternatives are synthesized from the elements of the limiting supermatrix. The period of the U.S. economic recovery is estimated from the desirability indices, which have been generated in the first column of the limiting supermatrix, coupled with the midpoints of the different time periods [2]. The predicted period of the turnaround thus obtained by the FDSS is 10.327 months (see Table III).

Case 2: Triangular Fuzzy Judgments: To form triangular fuzzy judgments, all crisp values of the comparison matrices, given in [2], have been converted into triangular fuzzy numbers. The lower and upper bounds of the fuzzy judgments coincide with the interval bounds from Case 1, whereas the means of the fuzzy sets are equal to the initial crisp judgments. As in the previous interval case, the Input Matrix Worksheet is used to elicit weights from each triangular fuzzy matrix. Decomposing each fuzzy matrix by eleven equally distanced -cuts, the program obtains eleven sets of weights, which are then aggregated by applying (12). The overall weights of the time periods regarding the controlling criterion monetary policy are shown in Fig. 6. Proceeding with all other steps, as in Case 1, the FDSS finally . From this matrix obtains the limiting supermatrix for the predicted period of the turnaround is calculated, which is 10.209 months (see Table III). C. Comparing the Results to the Classical ANP The results of the Interval FANP and Triangular FANP are compared with Saatys ANP results [2] in Table III. Although the desirability indices derived from these methods are slightly different, the ranking of the recovery periods, the overall weights and the predicted turnaround periods are rather consistent for all methods. As it can be observed from Table III, twenty-four and three months are the most likely time recovery periods, and the meaningful turnaround obtained by each method is about ten months.

MIKHAILOV AND SINGH: FUZZY ANALYTIC NETWORK PROCESS

41

When compared with the classical ANP, some additional advantages of the FANP can be clearly identified. a) It better models the ambiguity and imprecision associated with the pairwise comparison process. b) It successfully derives priorities from both consistent and inconsistent judgments. c) It is cognitively less demanding for the decision makers. d) It is an adequate reflection of the decision-makers attitude toward risk and their degree of confidence in the subjective assessments. VIII. CONCLUSIONS In this paper an attempt is made to extend the ANP so that it can deal with uncertain decision-making problems. A fuzzy modification of the ANP is proposed that handles multiple forms of uncertain pairwise comparison judgments. Another contribution of this study is the development of a working prototype of a fuzzy decision support system, realising the proposed fuzzy modification of the ANP. Although the poorly structured nature of the ANP results in many difficulties in modeling decision making processes, the constructed FDSS provides a user-friendly implementation of the proposed approach. ACKNOWLEDGMENT L. Mikhailov wishes to express his special thanks to the late Professor M. G. Singh, who initiated research in this area. Without his valuable assistance, guidance and support the realization of this project would not have been possible. The author is also very grateful to his research student Y. Chen for his efforts in developing the FDSS software. REFERENCES
[1] T. L. Saaty, Fundamentals of Decision Making and Priority Theory With the Analytic Hierarchy Process. Pittsburgh, PA: RWS, 1994. [2] , Decision Making With Dependence and Feedback: The Analytic Network Process. Pittsburgh, PA: RWS, 1996. [3] L. Mikhailov and M. G. Singh, Fuzzy assessment of priorities with application to the competitive bidding, J. Dec. Syst., vol. 8, no. 1, pp. 1128, 1999. [4] L. Mikhailov, Deriving priorities from fuzzy pairwise comparison judgments, Fuzzy Sets Syst., vol. 134, no. 3, pp. 365385, 2003. [5] , A fuzzy programming method for deriving priorities in the analytic hierarchy process, J. Oper. Res. Soc., vol. 51, pp. 341349, 2000. [6] L. Mikhailov and M. G. Singh, Comparison analysis of methods for deriving priorities in the in the AHP, in Proc. IEEE Conf. Systems, Man, Cybernetics, Tokyo, Japan, Oct. 1999, pp. 10371042. [7] T. L. Saaty and L. G. Vargas, Uncertainty and rank order in the analytic hierarchy process, Eur. J. Oper. Res., vol. 32, pp. 107117, 1987. [8] A. Arbel, Approximate articulation of preference and priority derivation, Eur. J. Oper. Res., vol. 43, pp. 317126, 1989. [9] A. Arbel and L. G. Vargas, Preference simulation and preference programming: Robustness issues in priority derivation, Eur. J. Oper. Res., vol. 69, pp. 200209, 1993. [10] M. Kress, Approximate articulation of preference and priorities derivationA comment, Eur. J. Oper. Res., vol. 52, pp. 382383, 1991.

[11] C. G. Boender, J. G. de Graan, and F. A. Lootsma, Multi-criteria decision analysis with fuzzy pairwise comparisons, Fuzzy Sets Syst., vol. 29, pp. 133143, 1989. [12] J. J. Buckley, Fuzzy hierarchical analysis, Fuzzy Sets Syst., vol. 17, pp. 233247, 1985. [13] D. Y. Chang, Applications of the extent analysis method on fuzzy AHP, Eur. J. Oper. Res., vol. 95, pp. 649655, 1996. [14] H. Deng, Multicriteria analysis with fuzzy pairwise comparisons, Int. J. Approx. Reasoning, vol. 21, no. 3, pp. 215231, 1999. [15] L. C. Leung and D. Cao, On consistency and ranking of alternatives in fuzzy AHP, Eur. J. Oper. Res., vol. 124, pp. 102113, 2000. [16] P. J. van Laarhoven and W. Pedrycz, A fuzzy extension of Saatys priority method, Fuzzy Sets Syst., vol. 11, pp. 229241, 1983. [17] M. W. Hartmann, On fuzzy rank ordering in polyoptimization, Fuzzy Sets Syst., vol. 11, pp. 253264, 1983. [18] A. Kaufmann and M. M. Gupta, Introduction to Fuzzy Arithmetic Theory and Application. New York: Van Nostrand, 1991. [19] M. Lee, H. Pham, and X. Zhang, Theory and methodology: A methodology for priority setting with application to software development process, Eur. J. Oper. Res., vol. 118, no. 2, pp. 375389, 1999. [20] L. Meade and J. Sarkis, Strategic analysis of logistics and supply chain management systems using the analytical network process, Transp. Res. E, vol. 34, no. , p. , 1998.

Ludmil Mikhailov received the M.Sc. degree in automatic control from the Technical University. Sofia, Bulgaria, and the Ph.D. degree in technical cybernetics from the Bulgarian Higher Certifying Commission, in 1976 and 1982, respectively. Since 1997, he has been teaching in the Department of Computation, University of Manchester Institute of Science and Technology (UMIST), Manchester, U.K. Before joining UMIST, he was an Associate Professor in the Institute of Control and System Research at the Bulgarian Academy of Sciences. He is the author of about 70 technical papers in peer-reviewed journals and international conferences, and he holds two patents in the area of systems and control. His current research interests include multiple criteria decision analysis, fuzzy logic systems, decision-making under uncertainty, and intelligent decision support systems.

Madan G. Singh (SM82F89) received the Ph.D. degree from Cambridge University, Cambridge, U.K., in 1974, the Doctorate dEtat degree from the University of Toulouse, Toulouse, France, in 1978, and an Honorary Doctorate degree from the University of Waterloo, Waterloo, ON, Canada in 1996. He was a Co-Founder and Director of Knowledge Support Systems (KSS) Group, plc, Manchester, U.K., and one of the leading experts in the world in systems engineering and in decision support technology. Prior to establishing KSS, he had provided decision support consultancy services to various companies, banks, retailers, and manufacturers in both Europe and the United States. He was a Professor at the University of Manchester Institute of Science and Technology. He edited the 18th Volume of the Encyclopedia of Systems and Control. During his career, he authored, or co-authored, 17 books and 200 scientific articles in his field. Dr. Singh won a number of honors, awards, and distinctions, including Chevalier des Palmes Academiques (France) and was granted unrestricted permission by Her Majesty the Queen to wear insignias of his honors. He won the IEEE Third Millennium Medal, the Norbert Weiner Award, and the Joseph Wohl Award.

Vous aimerez peut-être aussi