A theory of conditioning: Inductive learning within rule-based default hierarchies.
Keith J. Holyoak, Kyunghee Koh, Richard E. Nisbett · Psychological Review · 1989
We present a theory of classical conditioning based on a parallel, rule-based performance system integrated with mechanisms for inductive learning.Inferential heuristics are used to add new rules to the system in response to the relation between the system's predictions and environmental input.A major heuristic is based on "unusuamess": Novel cues are favored as candidates to predict events that are important or unexpected.Rules have strength values that are revised on the basis of feedback.The performance system allows rules to operate in parallel, competing to control behavior and obtain reward for successful prediction of important events.Sets of rules can form default hierarchies: Exception rules censor useful but imperfect default rules, protecting them from loss of strength.The theory is implemented as a computer simulation, which is used to model a broad range of conditioning phenomena, including blocking and overshadowing, the impact of statistical predictability on conditioning, and conditioned inhibition.The theory accounts for a variety of phenomena that previous theories have not dealt with successfully.Intelligence manifests itself in the adaptation of goal-directed systems to complex and potentially dangerous environments.In order to acquire nourishment and avoid injury, an animal must process a large and ever-changing array of perceptual information.From this array it must identify cues that signal important events and occasions for goal-appropriate actions.It must learn significant regularities even though valid cues are only imperfectly predictive, often conflict with one another, and are embedded among a plethora of irrelevancies.The kinds of learning that underlie such adaptation fall under the rubric of induction, broadly defined as those inferential processes that expand knowledge in the face of uncertainty (J.H.