Asia Pacific Management Review Review (2006) 11(3), Iraj Mahdavi et al./Asia Pacific Management (2006)155-162 11(3), 155-162 Conceptual Framework for Quantitative Modeling of Semi-structured MADM Iraj Mahdavia∗, Babak Shirazia, Namjae Chob and Nezam Mahdavi-Amiric a Department of Industrial Engineering, College of Technology, Mazandaran University of Science & Technology, PO Box734, Babol, Iran b School of Business, Hanyang University, 17 Haegdang-dong, Seongdong-gu, Seoul, Korea c Faculty of Mathematical Sciences, Sharif University of Technology, Tehran, Iran Accepted in June 2006 Available online Abstract In some situations, MADM matrix is not distinguished completely at the first stage of decision making, because of the complexity of environment. These complexities lead to incomplete cognition and non-optimal decision making. In such “semi-structured” environment, due to its high degree of complexity, the whole environment is not identifiable for Decision Maker (DM). We design an autonomous agent for semi-structured MADM that solves problems when alternatives have incomplete structure and DM is not able to recognize the whole alternatives of the environment for optimal decision making. The proposed model is a systematic approach for semi-structured MADM with multi-layer mathematical model. The Agent’s Stepwise Response Generator (ASRG) moves in semi-structured environment over decision surface step by step to generate hidden alternatives. The new alternatives are designed to go through Feasibility Analyzer and Dynamic Filter Module. The procedure is continued with a closed loop feedback which results in the construction of the Meta-Decision phase. Keywords: Autonomous agent; Decision surface; Meta-decision; Semi-structured MADM represent different dimensions of an alternative, they may conflict with each other. For instance, cost may conflict with profit, etc. Most of the MADM methods require that attributes be associated with weights of importance. Usually, these weights are normalized to add up to one. Several methods have been proposed for solving multi-attribute decision making problems. A major criticism to MADM is that different techniques yield different results when applied to the same problem. A simulation comparison of selected methods was performed by Zanakis et al. (1998). 1. Introduction Multi Attribute Decision Making (MADM) is the most well known branch of decision making. It is a branch of a general class of Operations Research (OR) models which deal with decision problems under the presence of a number of decision criteria (Triantaphyllou et al., 1998). This class of models is very often called Multi Criteria Decision Making (MCDM). According to many authors (Zimmermann, 1996), MCDM is divided into Multi Objective Decision Making (MODM) and Multi Attribute Decision Making (MADM). MODM studies problems in which the decision space is continuous. A typical example is mathematical programming problems with multiple objective functions. On the other hand, MADM concentrates on problems with discrete decision spaces. In these problems the set of decision alternatives tends to be predetermined. Although MADM methods vary widely, many of them have certain aspects in common (Chen and Hwang, 1992). Each MADM problem is associated with multiple attributes. Attributes are also referred to as “goals” or “decision criteria”. Attributes represent different dimensions from which the alternatives can be viewed and measured. In cases where the number of attributes is large, attributes can be arranged in a hierarchical manner. That is, some attributes are defined as major attributes. Each major attribute is associated with several sub-attributes. Similarly, each sub-attribute may be further associated with several sub-sub-attributes and so on. Since different attributes ∗ Email: 2. Multi Attribute Decision Making During the recent decades, the classical decision making of optimization with one criterion or one objective function has evolved into Multiple Criteria Decision Making (MCDM) models for complex decision making problems. These models can be linear, nonlinear, or hybrid. Two categories of decision making with multiple criteria are identified. Multiple Objective Decision Making (MODM) model is used to support planning and MADM is designed to select the best alternative (Hwang and Yoon, 1971; Triantaphyllou et al., 1998). Full-structured MADM model is formulated in the form of decision making matrix as shown in Table 1. [email protected] 155 Iraj Mahdavi et al./Asia Pacific Management Review (2006) 11(3), 155-162 Table 1. Full-structured Static MADM Matrix then we could use the decision making techniques as SAW, ISAW, TOPSIS, LINMAP, ELECTRE, etc. (Hwang and Yoon, 1971; Triantaphyllou et al., 1998). Classical MADM techniques using deterministic and specific mathematical models are applicable when all system variables are determined and there are no uncontrollable variables. In contrast, the uncertainty and complexity in problems and alternatives cause incomplete understanding and thus, incomplete decision making (Gu and Zhu, 2006; Kaebling et al., 1998; Leung et al., 2006). These complexities can be the cause for a decision bottleneck of a system. Complexities arise from the relationship between DM and Environment (Kendall, 2000). Specifically the complexity is related to the following aspects: 1. DM intelligence: Inability in recognizing problems, and in defining problems, prioritization, and organizing information. A1 , A2 , …, Am in matrix D are predefined alternatives and x1 , x2 , …, xn represent utility attributes applicable to the alternatives. Each component aij stands for the coefficient value of j-th attribute for i-th alternative. If a problem does not have a clear structure due to environmental complexity, we can not construct matrix D completely. The problem domain can also take static or stochastic dynamic nature (Howard, 1971; Kaebling et al., 1998; Leong, 1994; Monahan et al., 2000). So we can apply MADM to different types of problems as shown in Figure 1. 2. Problem planning: Inability in generating alternatives, and in explaining and evaluating alternatives. 3. Selecting: Inability in selecting a right solution method and selecting alternatives. 4. Counting: Large number of alternatives. 5. Environment: Environment with uncontrollable variables, environmental disturbance. Here we focus on the static structure. In static MADM system, decision making does not depend on time and the matrix has non-dynamic structure. But it can be incomplete when the problem is semi-structured MADM. If the decision making system were full-structured static MADM, These factors lead to an “environment with high degree of complexity”. In these cases a part of MADM matrix is invisible. This structure is called “semi-structured” decision model as shown in Table 2. Alternatives Am+j , for j ≥ 1, are invisible. Table 2. Semi-structured MADM Matrix Figure 1. MADM Classification 156 Iraj Mahdavi et al./Asia Pacific Management Review (2006) 11(3), 155-162 Figure 2. High-level Descriptive Functions System the system) that may naturally possess nonlinear elements (Boutilier et al., 1999). We can use regression methods to fit a function in an n-dimensional Euclidean space. Some high level functions which describe the system behavior (i.e. sub-goals in our model) are stability, productivity, performance and flexibility: Because of the invisibility in this incomplete structure, the analyst’s task is to help DM to make better decision through incremental exploration of alternatives (Leung et al., 2006). This proposed model should be in a way that help decrease “decision bottleneck”. For this purpose we suggest an autonomous decision maker agent. This agent recognizes the incomplete structure of design environment, and begins to interpret the environment step by step and generate unknown alternatives and provides feedback to redesign the alternatives to complete the matrix structure. The agent, thus, determines the unspecified portion of the decision for a DM and helps him to proceed in decision making. We design and suggest a multi-layer architecture of agent and the basic rules of each layer. Each layer core uses OR rules for designing sub-modules. Subgoal1 = Stability =F1 (x1 , x2 , …, xn); Subgoal2 = Productivity= F2 (x1 , x2 , …, xn) ; Subgoal3 = Performance = F3 (x1 , x2 , …, xn) ; Subgoal4 = Flexibility= F4 (x1 , x2 , …, xn). 3. Autonomous Decision Maker Agent Decision making in semi-structured environments needs a support for searching unknown portion of the decision environment. This support can be either an automated decision making process or a mechanism for generating unknown alternatives (Alami et al., 1998; Baroni et al., 1995; Ferber, 1999; Leger, 1999). Here we focus on the autonomous decision maker agents. An agent acts as an interface to reduce complexity of the semi-structured decision environment. Agent helps DM in making decision so that the relationship between DM and Agent is depicted as shown in Figure 3. Under a semi-structured environment without a clear structure we need to analyze the system step by step following the hierarchical structure. The hierarchical diagram describes the behavior of the system as shown in Figure 2 (Alpert, 1971; Armacost and Hosseini, 1994; Youngpil et al., 2003). This structure has 3 layers of Goal, Sub-Goals (F1 , F2 , …, Fk) and attributes (x1 , x2 , …, xn). Each sub-goal is a function of attributes describing system behavior. These functions reflect past experiences of the system (DM’s and analyst’s beliefs, past activities of 157 Iraj Mahdavi et al./Asia Pacific Management Review (2006) 11(3), 155-162 Figure 3. «DM – Agent » Relationship Autonomous agent generates new alternatives that DM could not recognize independently from other agents in the multi-agent system and as a partial substitute for DM. Due to the semi-structured nature of problems, autonomous agent decision making is sequential (stepwise) to reach the final goal of a system (Littman et al., 1996). As an agent receives additional information from the environment, he helps or even replaces DM in an evolutionary manner (Cardon et al., 2000). Leger (1999) shows comparison between human DM and autonomous DM agent. As to be seen in the next section, a DM agent has a multilayered structured environment that searches through a vague space to generate alternatives (Littman et al., 1996). The agent search locally on the decision surface and varies the search incrementally using different paths as shown in Figure 4 (Barraquand et al., 1992; Menczer et al., 2001). Figure 5. Autonomous Decision Maker Agent Architecture The concept of the modules in the architecture can be summarized as follows: Module 1- Start from an initial point: Provides initial starting point on the decision surface for beginning to move. Module 2- Autonomous Stepwise (Sequential) Response Generator (ASRG): Autonomously generates new responses step by step away from the initial starting point. Module 3- Feasibility Analyzer (FA): Tests the feasibility of the generated alternatives in the solution space of a given problem. A0 is one of m predefined alternatives (A1 , A2 , …, Am) and A11, A12 are new generated alternatives. An autonomous agent for decision making in a semi-structured environment with high degree of complexity should be able to deal with the vague nature of the structure. This agent should be structured as given in Figure 5. Module 4- Dynamic Filtering System (DFS): Presents the new alternative to DM, eliminates irrelevant alternatives based on the interaction with DM and suggests the selection of an alternative that satisfies a specific condition. Module 5- Meta-Decision Synthesizer: Extracts effective alternatives and adds them to the initial matrix. Re-Decision: Revises MADM matrix based on newly generated alternatives and applies classic MADM techniques to the incrementally expanded matrix. 4. Decision Surface and Autonomous Stepwise (Sequential) Response Generator The method we use for generating the alternatives in a vague semi-structured environment follows the sequential movement logic. Since the process of alternative generation is a step by step movement on the decision surface, a stepwise response generator is needed. The agent uses a movement method on the decision surfaces as shown in Figure 6. Figure 4. Type of Agent Search Via Path ω1 or ω The environmental complexity is interrelated to the movement on the surface. As for the modeling of the movement on this surface, we can use the regression method to fit the function in an n-dimensional Euclidean space. 1 158 Iraj Mahdavi et al./Asia Pacific Management Review (2006) 11(3), 155-162 Ω(A) ω1 A1 A11 A2 Ω (A) = Ω ( x1 , x2 , …, xn) x2 x1 Figure 6. Agent Movement and Generation of New Alternatives Figure 7. Agent Movements Over Decision Surface Ω To solve a semi-structured problem, we begin with the decision surface which contains the initial point of the system with m predefined alternatives, and we can assign more complex functions to this surface to deal with a more complex environment. Agent begins movement on the decision surface Ω from specified points A1 , A2 , …, Am through the path ω constructed by algorithms φh-1 (to be described in the next section). Figure 7 shows how the agent starts its movement from A1 and generates A11 through path ω1 over decision surface Ω (A). 4.1 Definition: Decision Surface 4.2 Sequential (Stepwise) Algorithms for Moving on a Decision Surface F is a general form of the descriptive function and the set of points A= {x1 , x2 , …, xn } is considered to satisfy the followings: Rc = {A | A ∈ R n, Fi (A) ∈ R} Suppose that Ω(A) is the decision surface. An agent starts from a selected and applicable point Ak on the decision surface. We define ξ=ξk as the amount of step size. If it is too small, the speed of agent’s movement will be too slow. On the other hand, if it is too large the system will face overshooting. For this reason, the selection of ξk is of special importance. Algorithm φh -1 is a suggestion for generating alternatives as given below: and Ω is the decision surface, 4.2.1 Algorithm φh -1 F: R n → R ( x1 , x2 , …, xn ) ∈ R n | xi ∈ R Î A ∈ Rn Î Fi(A) ∈ R , i=1, 2 , …, k ( k descriptive functions ) 1- Begin decision surface = Ω (A) = {A | A satisfies the above conditions} Ω(A) = Ω( x1 , x2 , …, xn ) is one of the descriptive functions Fi (A); A = ( x1 , x2 , …, xn ). 2- Initialize A, σ 3- While || ∂Ω (A) || > σ do 4- Determine ξ 5- AÅ A – ξ ∂Ω (A) Decision surface is an n-dimensional subspace of the n+1 dimensional space which is divided into Rc and R n-Rc. Depending on the level of complexity we can associate several forms of functions with the decision surface: Ω(A) = b0 + ∑ bi xi 6- Return A 7- End. (A: new generated alternative) 1≤ i≤ n To design a method for setting ξ , suppose that the function can be approximated by a second-order expansion around the value A = Ak : ( linear decision surface for low complexity) 1≤ i≤ n , 1≤ j≤ n Ω(A) = b0 + ∑ bi xi +∑ ∑ bij xi xj ( quadratic decision surface for medium complexity) Ω(A) ≈ Ω(Ak)+ ∂ Ω t(A-Ak) + ½ ( A-Ak) t H (A-Ak) Ω(A) = b0 + ∑ bi xi +∑ ∑ bij xi xj +∑ ∑ ∑ bijk xi xj xk 1≤ i ≤ n , 1≤ j ≤ n , 1≤ k ≤ n (highly complex environment). where H is the Hessian matrix (second partial derivatives ∂2 Ω / ∂Ai∂Aj ) evaluated at Ak . 159 Iraj Mahdavi et al./Asia Pacific Management Review (2006) 11(3), 155-162 Using the above approximation, we will have: Ω(Ak+1) ≈ Ω(Ak) - ξ k || ∂Ω || 2 + ½ ξ k 2 ∂Ω t H ∂Ω . Feasibility Analysis of A11 : 1) x1L ≤ x111≤ x1R 2) x2L ≤ x211≤ x2R Minimizing with respect to ξ k we will obtain: ξ = ξ k = || ∂Ω || 2 / ∂Ω t H ∂Ω. Of course, one may use a line search strategy to set ξ in each iteration to guarantee reduction in every step (see Nocedal and Wright, 1999). Ω(A) ω 11 A As a second-order approach for minimization, we can minimize the quadratic approximation in each iteration obtaining the following algorithm (the so-called modified Newton’s method): A0 x111 4.2.2 Algorithm φh -2 x211 1- Begin x2 2- Initialize A, σ x1 3- Repeat 4- p= -H -1∂Ω(A) ( p solution of Hp= - ∂Ω(A)) Figure 8. Extracting New Alternative Attributes 5- A Å A+ α p (α obtained from a line search) 6- Until || p || ≤ σ - F(A ) = min F(A), F(A+ ) = max F(A) . 7-Return A We can design three general forms of filters as shown in Figure 9: 8-End. In the above algorithm, the step size in each iteration can be obtained using a line search strategy. Alternatively, one may use robust minimization techniques such as secant (quasi-Newton) methods or, to avoid line search, trust region methods (Nocedal and Wright, 1999). 5. a) Low–pass filter: Eliminates all alternatives which satisfy F(A) ≥ F(A+) or passes all alternatives which satisfy F(A) <F(A+ ). b) Medium–pass filter: Eliminates all alternatives which do not satisfy F(A ) ≤ F(A)≤ F(A+) or passes all alternatives which satisfy F(A ) < F(A) < F(A+). Feasibility Analyzer For some newly generated alternatives, some attributes xi may not satisfy xiL ≤ xi ≤ xiR . For this reason we use the Feasibility Analyzer to find feasible solutions among the newly generated alternatives. A new alternative A is feasible if all of its attributes are in the feasible region (xiL ≤ xi ≤ xiR , for all i ,1 ≤ i ≤ n ). Over the decision surface Ω, the new alternative A11 is generated following path ω. To test the feasibility of A11 as appeared in Figure 8 we consider the following constraints: c) High–pass filter: Eliminates all alternatives which satisfy F(A) ≤ F(A ) or passes all alternatives which satisfy F (A) > F (A ). Filtering newly generated alternatives is a dynamic activity because the alternatives are generated autonomously by the agent without the DM’s control. Filters may have different structures when compared with the above general conditions, but they should eliminate some unnecessary alternatives. So we can consider different types of filters according to DM’s perception and judgment. Constraints x iL ≤ x i ≤ x iR in the n+1 dimensional space (due to n attributes) form a subspace intersecting the decision surface. 6. 7. Meta-Decision Synthesizer and Re-Decision Dynamic Filtering System Definition: Meta-decision Designing a filter is a multi-stage, repetitive activity that requires an interaction with DM. A filter has a dynamic nature due to the changes in priorities and attitude of the DM. The filter eliminates some alternatives so that the remaining alternatives satisfy DM’s needs. Let A and + A be arguments of minimum and maximum limits of the descriptive function describing the attitude of a DM: The movements on the decision surface Ω from certain points A1 , A2 , …, Am through the path ω with the algorithm φh-1 generate new alternatives which DM did not have at the beginning. These additional alternatives form a newly updated matrix that can lead to a new decision. Besides the responses generated by traditional approaches, 160 Iraj Mahdavi et al./Asia Pacific Management Review (2006) 11(3), 155-162 Step 4. Use sequential algorithms to move on the decision surface from the best existing alternative as described in Section 4.2 and obtain new alternatives. Step 5. Apply feasibility analyzer and dynamic filtering to the new alternatives as illustrated in sections 5 and 6 respectively. Step 6. Obtain an acceptable new alternative and create a new MADM matrix. If there is no new alternative, then stop or jump to the next movement with subsequent permutation. Figure 9. Low-Pass-Filter, Medium-Pass-Filter and High-Pass-Filter (left to right) Step 7. Let I=I+1. If I is less than NA then go to Step 1, otherwise Stop. new alternatives are added by the agent to the decision matrix as shown in Figure 10. We call this process to improve the accuracy of a decision through a repeated re-creation of the decision matrix as the “meta-decision”. 8. Conclusions This study suggests the use of a Decision Maker Agent, which autonomously generates unspecified alternatives independent of DM. The agent adds newly generated alternatives to the initial matrix. MADM matrix extends to contain more alternatives for the decision maker. For generating alternatives not recognized by the DM from the beginning, the autonomous agent moves step by step on the decision surface and uses the path that has been calculated by the corresponding algorithm. All newly generated alternatives are not added to the initial matrix, because the agent generates them autonomously without the control of DM. For this reason, the agent has a layer to filter out alternatives which do not satisfy the perceptual requirements of the decision maker (Feasibility Analyzer and Dynamic Filter). Newly generated alternatives created by the stepwise generator that pass through Feasibility Analyzer and Dynamic Filtering System can be added to the initial decision system. For this newly generated matrix, we design a new decision surface on the basis of previously obtained stage alternatives. Sequential algorithms are then applied for movement on the decision surface to obtain next new alternatives. MADM technique is used here to make a decision again (Re-Decision). Improved outcomes are produced with the repetition of the Meta-Decision process. The solution can lead to an effective and flexible selection through an active interaction with the DM. Finally the steps of the proposed conceptual framework for quantitative modeling of semi-structured MADM follow (NA is the number of new alternative to be created). The autonomous Decision Maker Agent interacts with DM to increment the decision matrix to be reconsidered on the basis of the classical MADM to achieve the decision goal. Future research can focus on the development of an extended model to incorporate the stochastic dynamic evolution of the logic by taking into account the possibility of the changes in attribute values and the existence of missing attribute values of known alternatives. Step 1. Let I=0. Step 2. Identify decision surface on the basis of existing alternatives as defined in Section 4.1. Step 3. Obtain the best existing alternative using classic MADM technique. Figure 10. Meta-Decision 161 Iraj Mahdavi et al./Asia Pacific Management Review (2006) 11(3), 155-162 References Kendall, K. E. (2002). Systems Analysis and Design, Fifth Edition, Prentice Hall Inc. Leger, C. (1999). Automated Synthesis and Optimization of Robot Configurations an Evolutionary Approach. Ph.D. Thesis, The Robotics Institute, Carnegie Mellon University, Pittsburgh, Pennsylvania. Leong, T. Y. (1994). An Integrated Approach to Dynamic Decision Making under Uncertainty. Laboratory for Computer Science, MIT/LCS/TR631. Leung, Y., Wu, W. Z. and Zhang,W. X. (2006). Knowledge acquisition in incomplete information systems: A rough set approach. European Journal of Operational Research, 168, 164-180. Littman, M. L. Algorithms for Sequential Decision Making. (1996). Ph.D. Thesis, Department of Computer Science, Brown University. Menczer, F., Street, W. and Degeratu, M. (2001). Evolving Heterogeneous Neural Agents by Local Selection. Advances in the Evolutionary Synthesis of Intelligent Agents, MIT Press 337-366. Monahan, G. E. (2000). Management Decision Making, Cambridge University Press. Nocedal, J. and Wright, S. (1999). Numerical Optimization, NY: Springer. Triantaphyllou, E., Shu, B., Nieto Sanchez, S. and Ray, T. (1998). Multi-Criteria Decision Making: An Operations Research Approach. Encyclopedia of Electrical and Electronics Engineering, 15, 175-186. Youngpil, C., Seongbong , C. and Mooyoung, J. (2003). Satisfaction assessment of multi-objective schedule using neural fuzzy methodology. International Journal of Production Research, 41(8), 1831-1849. Zanakis, S. H., Solomon, A., Wishart, N. and Dublish, S. Multi-attribute decision making: A simulation comparison of select methods. European Journal of Operational Research, 107, 507-529. Zimmermann, H. J. (1996). Fuzzy Set Theory and Its Applications, Second Edition, Kluwer Academic Publishers, MA: Boston. Alami, R., Chatila, R., Fleury, S., Ghallab, M. and Ingrand, F. (1998). An architecture for autonomy. Journal of Robotic Research, 17(4), 315-337. Alpert, M. I. (1971). Identification of determinant attributes: A comparison of methods. Marketing Research, 8(5), 184-191. Armacost, R. L. and Hosseini, J. C. (1994). Identification of determinant attributes using the analytic hierarchy process. Journal of the Academy of Marketing Science, 22(4), 383-392. Barraquand, J., Langlois, B. and Latombe, J. C. (1992). Numerical potential field techniques for robot path planning. IEEE Transactions on Systems, Man, and Cybernetics, SMC, 22(2), 224–241. Baroni, P., Guida, G., Mussi, S. and Vetturi, A. (1995). A distributed architecture for control of autonomous mobile robots. In ICAR’95, 2, 869-877. Boutilier, C., Dean, T. and Hanks, S. (1999). Decision-theoretic planning: Structural assumptions and computational leverage. Journal of Artificial Intelligence Research, 11, 1-94. Cardon, A., Galinho, A. and Vacher, J. P. (2000). Genetic algorithms using multi-objectives in a multi-agent system. Robotics and Autonomous Systems, 33, 179-190. Chen, S.J. and Hwang, C.L. (1992). Fuzzy Multiple Attribute Decision Making: Methods and Applications, Sringer-Verlag, Berlin, Germany. Ferber, J. (1999). Multi-Agent Systems: An Introduction to Distributed Artificial Intelligence, Addison-Wesley. Gu, X. and Zhu, Q. (2006). Fuzzy multi-attribute decision-making method based on eigenvector of fuzzy attribute evaluation space. Decision Support Systems, 4, 400-410. Howard, R. A. (1971). Dynamic Probabilistic Systems: Semi-Markov and Decision Processes, vol. 2 of Series in Decision and Control, NY: John Wiley & Sons. Hwang, C. L. and Yoon, K. (1981). Multiple Attribute Decision Making, NY: Springer-Verlag. Kaebling, L. P., Littman, M. L. and Cassandra, A. R. (1998). Planning and acting in partial observable stochastic domains. Artificial Intelligence, 101(1-2), 99-134. 162
© Copyright 2026 Paperzz