Understanding Similarity in Choice Behavior: A Connectionist Model

Abstract

Understanding Similarity in Choice Behavior: A Connectionist Model Frank Y. Guo (fyguo@ucla.edu) UCLA, Department of Psychology, 405 Hilgard Ave. Los Angeles, CA 90095-1563, USA Keith J. Holyoak (holyoak@psych.ucla.edu) UCLA, Department of Psychology, 405 Hilgard Ave. Los Angeles, CA 90095-1563, USA Abstract Classical choice theories assume choice behavior is based on value maximization computed over the entire choice set. However, empirical evidence has revealed violations of axioms of rational choice that cannot be explained by value maximization. We argue that choice behavior can be reconceptualized as value maximization constrained by categorization processes, and describe a neural network model developed to account for key empirical findings. The model simulates two important phenomena that have been construed as irrational choice behavior, namely, the similarity effect and the attraction effect. We argue that there are important commonalities among choice behavior, categorization and perception. Introduction Many axiomatic theories of choice behavior are based on the assumption that decision making is based on a process of value maximization performed over all attributes (c. f., Tversky & Simonson, 1993). However, empirical evidence has demonstrated that axioms of rational decision making are often violated in choice behavior, and value maximization alone is unable to explain these violations. Recently, an alternative perspective that is concerned with the relations between similarity processes and decision processes has been proposed to conceptualize choice behavior and to understand violations of rational decision making (Medin, Goldstone, & Markman, 1995). That view has been embodied in a comprehensive computational model of choice behavior (Roe, Busemeyer, & Townsend, 2001). In the spirit of this alternative perspective, we have developed a connectionist model to account for two key violations of rational choice, namely, the similarity effect and the attraction effect. Both of these phenomena involve adding a third alternative (decoy) to a choice set of two options, thereby leading to inconsistency of choice. If the decoy is similar and competitive (two alternatives are competitive when their additive utilities are almost identical to each other) to one of the original options, then the addition of the decoy decreases the choice probability of that option. This phenomenon is called the similarity effect (Tversky, 1972). If the decoy is similar to and dominated by one of the two original alternatives but not the other, then the addition of the decoy increases the choice probability of the dominant option more than the other alternative. This phenomenon is referred to as the attraction effect (Huber, Payne, & Puto, 1982). Both phenomena can potentially lead to violations of rational choice. Few theories were able to provide an integrated explanation of both phenomena prior to the model proposed by Roe et al. (2001), which is a neural network instantiation of the decision field theory (Busemeyer & Townsend, 1993). That model explains the two effects (in addition to several other important choice phenomena) by taking into consideration similarity relations among options and the dynamic nature of decision processes. The model described here is similar to that of Roe et al. in that it also takes into account similarity among alternatives; however, the manner in which similarity is represented and processed differs between the two models. We will briefly discuss the relationship between the two models after we present our proposal. Neural network models have been one of the major modeling tools in cognitive science (Rumelhart, McClelland, & PDP Research Group, 1986). However, such models have had only limited applications to decision behavior (Holyoak & Simon, 1999; Roe et al., 2001; Thagard & Millgram, 1995). The model we describe here, like that of Roe et al. (2001), uses a neural network approach to provide an account of the similarity and attraction effects. Operation of the Model Decision Scenario and Model Architecture The decision scenario used here is adapted from that used by Roe et al. (2001). The decision maker has to choose one car from a set of two or three alternatives by evaluating their ratings on two attributes: gas mileage and performance (see Figure 2). A simple neural network is constructed for this scenario. Figure 2 shows the architecture of the model, adapted from ECHO (Thagard, 1989), a neural network