Improving Efficiency by Learning Intermediate Concepts

One goal of explanation-based learning is to transform knowledge into an operational form for efficient use. Typically, this involves rewriting concept descriptions in terms of the predicates used to describe examples. In this paper we present RINCON, a system that extends domain theories from examples with the goal of maximizing classification efficiency. RINCON'S basic learning operator involves the introduction of new intermediate concepts into a domain theory, which can be viewed as the inverse of the operationalization process. We discuss the system's learning algorithm and its relation to work on explanation-based learning, incremental concept formation, representation change, and pattern matching. We also present experimental evidence from two natural domains that indicates the addition of intermediate concepts can improve classification efficiency.