Flexible Learning of Problem Solving Heuristics Through Adaptive Search

Abstract

Noting that the methods employed by existing learning systems are often bound to the intended task domain and have little applicability outside that domain, this paper considers an alternative learning system design that offers greater flexibility without sacrificing performance. An operational prototype, constructed around a powerful adaptive search technique, is presented and applied to the problem of acquiring problem solving heuristics through experience. Some performance results obtained with the system in a poker betting domain are reported and compared with those of a previously investigated learning system in the same domain. It is seen that comparable levels of performance are achieved by the two systems, despite the latter's dependence on a considerable amount of domain specific knowledge for effective operation.