A high-performance explanation-based learning algorithm

Abstract

The main contribution of this paper is a new domain-independent explanation-based learning (EBL) algorithm. The new EBL∗DI algorithm significantly outperforms traditional EBL algorithms both by learning in situations where traditional algorithms cannot learn as well as by providing greater problem-solving performance improvement in general. The superiority of the EBL∗DI algorithm is demonstrated with experiments in three different application domains. The EBL∗DI algorithm is developed using a novel formal framework in which traditional EBL techniques are reconstructed as the structured application of three explanation-transformation operators. We extend this basic framework by introducing two additional operators that, when combined with the first three operators, allow us to prove a completeness result: in the formal framework, every EBL algorithm is equivalent to the application of the five transformation operators according to some control strategy. The EBL∗DI algorithm employs all five proof-transformation operators guided by five domain-independent control heuristics.