FEEDFORWARD ARTIFICIAL NEURAL NETWORKS FOR SOLVING DISCRETE MULTIPLE CRITERIA DECISION-MAKING PROBLEMS

Decision making involves choosing some course of action among various alternatives. In almost all decision making problems, there are several criteria for judging possible alternatives. The main concern of the Decision Maker (DM) is to fulfill his conflicting goals while satisfying the constraints of the system. In this paper, we present an Adaptive Feedforward Artificial Neural Network (AF-ANN) approach to solve discrete Multiple Criteria Decision Making (MCDM) problems. The AF-ANN is used to capture and represent the DM's preferences and then to select the most desirable alternative. The AF-ANN can adjust and improve its representation as more information from the DM becomes available. We begin with the assumption that an AF-ANN topology is given, i.e., specific numbers of nodes and links are predetermined. To adjust the parameters of the AF-ANN, we present an iterative learning algorithm consisting of two steps: (a) generating a direction, and (b) a one-dimensional search along that direction. We then present a methodology to obtain the most appropriate AF-ANN topology and set its parameters. The procedure starts with a small number of nodes and links and then adaptively increases the number of nodes and links until the proper topology is obtained. Furthermore, when the set of training patterns (alternatives with their associated evaluations by the DM) changes, the AF-ANN model can adapt itself by re-training or expanding the existing model. Some illustrative examples are presented. To solve discrete MCDM problems by an AF-ANN, we show how to incorporate basic properties of efficiency, concavity, and convexity into the AF-ANN. We formulate the MCDM problems and use the AF-ANN to rank the set of discrete alternatives where each alternative is associated with a set of conflicting and noncommensurate criteria. We present a method for solving discrete MCDM problems through AF-ANNs which consists of: (a) formulating and assessing the utility function by eliciting information from the DM and then training the AF-ANN, and (b) ranking and rating alternatives by using the trained AF-ANN model. Some computational experiments are presented to show the effectiveness of the method.