A comparison of learning models (Q1899548)

Behavior changes with experience. An individual takes an action, observes some of the consequences, and perhaps later takes a different action in similar circumstances. The purpose of the present paper is to investigate such behavioral changes, which we regard as the manifestation of learning. We are interested in several questions, both normative and descriptive. Does behavior converge to an optimum? If so, is convergence as rapid as possible given available information? If not, can behavioral changes be described reasonably well by any simple model? These and similar questions can be asked in quite general contexts. We confine our attention to a particular laboratory version of a ``medical diagnosis'' task. After a short review of research in probabilistic learning situations, we describe the medical diagnosis task, the analytic framework we use, the question of optimality, and the formalization of the models for this task. We base our analysis on five models, the normative (optimal) Bayesian model, the Fuzzy Logical Model of Perception (FLMP), the two-layer Connectionist model (CMP), an Exemplar model, and a recent exemplar-based network model (ALCOVE). The asymptotic predictions and learning predictions of these models are all presented in the context of the medical diagnosis task in order to keep the exposition self-contained and to fix ideas and establish notation. Next we turn to the data. In addition to existing data, we analyze new data from simplified medical diagnosis tasks that include not only asymmetric but also symmetric base rates. We find evidence that subjects do learn in the sense that response frequencies to the symptom configurations approach objective probabilities more closely on later trials than on earlier trials. All of the descriptive models give a more accurate account of performance than the normative Bayesian model. Relative to a benchmark measure, however, none of these models does an especially good job of characterizing asymptotic performance or the learning process. In the last section we summarize our findings. Our results raise more questions than they answer, so we conclude by suggesting directions for more definite work.