Learning with Incrutable Theories

Abstract

This paper addresses the problem of learning problem solving knowledge when the domain theory of the learning system is inscrutable, i.e., not represented declaratively. We draw the following conclusions: First, although most of the algorithms and methods investigated in Explanation-Based Learning (EBL) literature assume declarative theories, at least in some domains, similar speedups can be achieved by simpler techniques that do not make this assumption. Second, learning with inscrutable theories blurs the distinction between “empirical― and “explanation-based― methods, both of which can be viewed as implementing some form of bias in their algorithms. Third, the utility problem is best addressed by implementing a bias that exploits the structure of the domain. We support our conclusions with experiments in the Eight Puzzle domain.