Toward a Theory of Well-Guided Search

Susan L. Epstein

A game learner’s experience is no more than the nodes it encoumers in the search space. For a challenging game, only a small fraction of that space can ever be visited. People’s ability to learn to play difficult games well is a clear indication that not all nodes are equally relevant to learning. There are, it is argued here, key nodes particularly important to the development of expertise, and those key nodes appear in clusters in the game tree. A game learning program might arrive at a key node by chance, be drawn there by a choice it makes, or be driven there by the moves of its opposition. Trainer guidance has some similarity to varying the sequence of training examples in ordinary induction, but here is delegated to file program itself. This paper offers empirical evidence of substantial improvement in the quality of play when a program is steered to clusters of key nodes, and considers several ways to do so.


This page is copyrighted by AAAI. All rights reserved. Your use of this site constitutes acceptance of all of AAAI's terms and conditions and privacy policy.