Journal of Experimental Psychology: General 1992. Vol. 121, No. 2.222-236 Copyright 1992 by the American Psychological Association, Inc. 0096-3445/92/S3.00 Predictive and Diagnostic Learning Within Causal Models: Asymmetries in Cue Competition Michael R. Waldmann Universitat Frankfurt, Frankfurt/Main, Federal Republic of Germany Keith J. Holyoak University of California, Los Angeles Several researchers have recently claimed that higher order types of learning, such as categoriza- tion and causal induction, can be reduced to lower order associative learning. These claims are based in part on reports of cue competition in higher order learning, apparently analogous to blocking in classical conditioning. Three experiments are reported in which subjects had to learn to respond on the basis of cues that were defined either as possible causes of a common effect (predictive learning) or as possible effects of a common cause (diagnostic learning). The results indicate that diagnostic and predictive reasoning, far from being identical as predicted by associationistic models, are not even symmetrical. Although cue competition occurs among multiple possible causes during predictive learning, multiple possible effects need not compete during diagnostic learning. The results favor a causal-model theory. Tasks as different as classical conditioning, category learn- ing, and causal induction can be viewed as examples of multiple-cue contingency learning. In each of these tasks, a number of cues, which might represent conditional stimuli, features, or causes, are combined to elicit a response. Because of this apparent formal similarity between different types of multiple-cue learning situations, it is tempting to postulate a common learning mechanism. Indeed, a number of research- ers have recently claimed that higher order types of learning, such as categorization and causal induction, can be explained by principles that govern lower order learning in animals, such as classical conditioning. Gluck and Bower (1988), for example, suggested that adaptive associative networks can provide powerful models of human categorization. These connectionist networks consist of an input layer that repre- sents potential cues, such as symptoms of possible diseases observed in a patient, and an output layer that might represent classification responses, such as diagnoses of alternative dis- eases. The responses of the network are computed by a linear function of the weighted cues. The weights are learned using the least mean squares (LMS) learning rule (Widrow & Hoff, 1960), in which the weights are incrementally updated in proportion to the response error they produce. Gluck and Bower showed that a simple model of this sort compares favorably with other models of human categorization (see also Estes, Campbell, Hatsopoulos, & Hurwitz, 1989; for a critique see Shanks, 1990a, 1990b). Because the LMS rule is formally equivalent to Rescorla and Wagner's (1972) theory of classical This research was supported by a National Institutes of Health Biomedical Research Support Grant to Keith J. Holyoak. A prelim- inary report of these experiments was presented at the 1990 meeting of the Psychonomic Society in New Orleans, Louisiana. We thank Patricia Cheng, Eric Melz, Frederic Vallee-Tourangeau, and two anonymous reviewers for helpful discussions. Correspondence concerning this article should be addressed to Michael R. Waldmann, who is now at Psychologisches Institut, Universitat Tubingen, Friedrichstrafle 21, 7400 Tubingen, Federal Republic of Germany. conditioning (Sutton & Barto, 1981), these findings suggest that categorization can be viewed as a special case of associa- tive learning. Similarly, Shanks and Dickinson (1987) argued that learning of causal relationships can be reduced to asso- ciative learning (see also Wasserman, 1990). In the associative framework, cues typically correspond to potential causes, and responses correspond to predictions of potential effects. Weights representing the strengths of the relationships be- tween causes and effects are learned in an incremental fashion. Cue Competition in Associative Models of Multiple- Cue Contingency Learning All modern associative learning theories emphasize the competitiveness of cues; indeed, cue competition can be viewed as the single most important feature of current asso- ciative learning theories (see Gallistel, 1990). The classic evidence for cue competition involves the phenomenon of blocking, first observed by Kamin (1969) in experiments on aversive conditioning in rats. Such blocking experiments typ- ically consist of two learning phases. In Phase 1, a rat learns to associate an initial conditioned stimulus (CS,), for example, a tone, with an unconditioned stimulus (US), for example, shock. In Phase 2, the previously conditioned CS, (tone) is presented together with a new, redundant CS 2 (e.g., light), and the compound is reinforced by the US. In the critical test phase, the rat sees each CS by itself. As expected, when presented with CS, (tone) alone, the rat still shows fear reac- tions. However, CS 2 (light) typically does not elicit fear reac- tions, even though the light was constantly paired with the US in Phase 2, and during this period, the shock never occurred in the absence of the light. Learning about the CS, seems to have blocked acquisition of associative strength for the CS 2 . Rescorla and Wagner (1972) developed their theory of associative learning to account for blocking and other findings involving cue interactions. Within the Rescorla-Wagner the- ory, blocking is viewed as the result of the failure of the 222