Automatic Feature Generation for Problem Solving Systems

Existing methods for constructive induction generally isolate feature generation from problem solving. This paper describes a theory of feature generation that creates features using both a domain theory and feedback performance. An evaluation function based on these features is learned simultaneously and is used to guide a problem solver. The Zenith system, an implementation of this theory, has been applied to two domains. In each domain, Zenith generated useful features, given only a domain theory and the ability to solve problems in the domain.