When 2 cues occur together and reliably predict an outcome, Ss often judge the effect of the compound as reducible to the individual effects of the elements. This elemental processing in predictive learning is perhaps the single most important aspect of most theories of human inference. Surprisingly, selectional processing was not observed in either blocking or conditioned inhibition problems. Only when the learner had past experience with another problem encouraging an elemental strategy were the expected selectional processes observed. These proactive effects of prior learning were abolished if the earlier problem required a nonadditive solution. The results suggest that configural cues were guiding predictive inferences in the absence of elemental processes.Many authors have been struck by the formal similarity of the problem of covariation detection in humans and animals (e.g., Gluck & Bower, 1988;Holyoak, Koh, & Nisbett, 1989;Shanks & Dickinson, 1987). Both organisms must learn to detect regularities in their environment and to take appropriate actions. That is, predictive signals must be identified in a stream of irrelevant cues, whether these signals predict the romantic interest of another person or patches of food for a hungry animal. These parallels raise the intriguing possibility that similar learning mechanisms may explain how humans and animals solve formally analogous problems.This possibility is encouraged when variables in the human and animal domains have corresponding effects. For example, in assessing the contingency between an action and an outcome, people are sensitive to the likelihood of the outcome in the presence versus absence of the action (e.g., Chatlosh, Neunaber, & Wasserman, 1985). The outcome is presumed to be contingent only if the action influences the background rate of occurrence of the outcome. By no means are the reported contingency effects equivalent to statistical measures of covariation. People often overestimate zero contingencies, especially if there is a high probability of the outcome (Alloy & Tabachnik, 1984). Animals show a similar sensitivity to contingency (Rescorla, 1968) and also respond inappropriately early in training when there is a nominal zero contingency (e.g., Benedict & Ayres, 1972). Shared imperfections in covariation detection like these can be explained by an associative model (e.g., Wasserman, Elek, Chatlosh, & Baker, 1993).Although there has been considerable progress in our understanding of the cognitive processes underlying predictive

Forms of inhibition were identified in human predictive learning that are qualitatively similar to those identified by P.C. Holland (1984) in rats. When P (positive) signaled the outcome and PN (N = negative) signaled the absence of the outcome, participants learned the discrimination, but the negative cue did not suppress responding to a transfer cue. Post-learning reversal training, in which N was followed by the outcome, did not abolish the original discrimination. These 2 results imply a configural form of inhibition. Negative transfer, which indicated a 2nd, elemental form of inhibition, was observed when neither PN nor N were reinforced during the discrimination stage. Under these conditions, negative transfer and the original discrimination were both abolished by individually pairing N with the outcome. Empirical parallels and differences with the animal conditioning literature are discussed.

Rescorla's (1969) recommendations concerning the logical and empirical operations for inferring Pavlovian conditioned inhibition were examined in light of modern comparison theories of Pavlovian conditioning and new data that question whether excitation and inhibition are opposite ends of a single continuum of associative strength. This reanalysis confirms Rescorla's dictum that the summation and retardation tests are a sufficient basis for inferring inhibition. No theoretical or empirical challenges since 1969 undermine the heuristic or explanatory power of that concept. However, passing both of his two operational tests may not be necessary: A conditioned stimulus may be inhibitory, yet may not pass one or the other test. Recent changes in the conception of inhibition also are reviewed and it is suggested that there may be other forms of associatively based inhibition that are not at all antagonistic to excitation.We thank Bob Tail for many useful discussions about this "accurst" problem and Anastasia Droungas and Ralph Miller for their comments on a draft of this article.

separate groups are illegitimate (e.g., Mahoney, Kwaterski, & Moore, 1975), but only that when differences are found, it is difficult to know what procedural details are contributing to the difference.

Five experiments examined whether past elemental or configural learning affects the coding of novel compound stimuli in human learning. Experiments 1-2B confirmed, in an unspeeded task, that past experience can have a powerful influence on whether a novel compound stimulus is treated as a combination of separate elements or as an undivided whole. Experiment 3 revealed no evidence that humans are malleable in the use of their stimulus coding strategies if the task is speeded rather than unspeeded. Experiment 4 revealed that rule abstraction as measured by a postexperimental questionnaire was not related to the occurrence • of transfer. Implications for elemental and configural theories of associative learning are discussed.