Generating Hypotheses to Explain Prediction Failures

Steven Salzberg

Learning from prediction failures is one of the most important types of human learning from experience. In particular, prediction failures provide a constant source of learning. When people expect some event to take place in a certain way and it does not, they generate an explanation of why the unexpected event occurred [Sussman 1975) [Schank 1982]. This explanation requires hypotheses based on the features of the objects and on causal relations between the events in the domain. In some domains, causal knowledge plays a large role; in some, experience determines behavior almost, entirely. This research describes learning in intermediate domains, where causal knowledge is used in conjunction with experience to build new hypotheses and guide behavior. In many cases, causal knowledge of the domain is essential in order to create a correct explanation of a failure. The HANDICAPPER program uses domain knowledge to aid it in building hypotheses about why thoroughbred horses win races. As the program processes more races, it builds and modifies its rules, winning horses. and steadily improves in its ability to pick winning horses.

This page is copyrighted by AAAI. All rights reserved. Your use of this site constitutes acceptance of all of AAAI's terms and conditions and privacy policy.