Heuristics for Inductive Learning

Abstract

A number of heuristics have been developed which greatly reduce the search space a learning program must consider in its attempt to construct hypotheses about why a failure occurred. These heuristics have been implemented in the HANDICAPPER system [Salzberg 1983, Atkinson & Salzberg 1984], in which they significantly improved predictive ability while demonstrating a remarkable learning curve The rationalization process has been developed as a verification system for the hypotheses suggested by the heuristics. Rationalization uses the causal knowledge of the system to ascertain whether or not a hypothesis is reasonable. If the hypothesis is not supported by causal knowledge, it is discarded and another hypothesis must be generated by the heuristics. The resulting learning system, by integrating causal knowledge w i th heuristic search, has quickly gone from essentially random predictive accuracy to a system which consistently outperforms the experts at predicting events in its problem domain.