A Comparison of Learning Models

Abstract

We investigate learning in a probabilistic task, called "medical diagnosis." On each trial, a subject is presented with a stimulus configuration indicating the value of four medical symptoms. The subject responds by guessing which of two diseases is present and is then given feedback about which disease was actually present. The feedback is determined according to fixed conditional probabilities unknown to the subject. We test a normative Bayesian model as well as simple variants of well-known psychological models including the Fuzzy Logical Model of Perception, an Exemplar model, a two-layer Connectionist model and an ALCOVE model. Both the asymptotic predictions of these models (i.e., predictions regarding behavior after it has stabilized and learning is complete) and predictions of trial-by-trial changes in behavior are tested. The models are tested against existing data from Estes et al. (1989, Journal of Experimental Psychology: Learning, Memory, & Cognition,15, 556-571) and new data from medical diagnosis tasks that include not only asymmetric but also symmetric base rates. Learning was observed in all cases in that subjects tended to match the objective probabilities of the symptom configurations more closely in later 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. We suggest that future experiments should address individual performance, rather than group learning curves.