Surprising Predictions

This paper describes a computational framework for theoretical models of classical conditioning. It is based on the engineering technique of Kalman filtering, which describes how systems can infer regularities in noisy and changing environments based on the results of partial experiments. The resulting class of conditioning theories has similarities with three existing (and apparently incompatible) ones; those due to Rescorla and Wagner (1972), Mackintosh (1975) and Pearce and Hall (1980). As in the Rescorla-Wagner rule, learning is dependent on errors in prediction of the unconditioned stimulus, but, as in the other rules, the amount of learning that accrues to each conditioned stimulus is a function of its own conditioning history.