Donald Bren School of Information and Computer Sciences

A system is described which learns to compose sequences of operators into episodes for problem solving. The system incrementally learns when and why operators are applied. Episodes are segmented so that they are generalizable and reusable. The idea of augmenting the instance language with higher level concepts is introduced. The technique of perturbation is described for discovering the essential features for a rule with minimal teacher guidance. The approach is applied to the domain of solving simultaneous linear equations.