Learning search heuristics from examples: a study in computer chess
-
-
[1] University of Ulster
University of Ulster
Reino Unido
-
- Localización: CAEPIA'97: actas / coord. por Asociación Española de Inteligencia Artificial, Vicente J. Botti Navarro, 1997, ISBN 84-8498-765-5, págs. 695-704
- Idioma: inglés
- Texto completo no disponible (Saber más ...)
-
Resumen
Search over large spaces is facilitated by heuristics which reduce the amount of search without commensurate loss in the quality of results. Machine learning techniques now offer methods for automatic discovery of such heuristics from examples. In this paper we report an attempt to discover search heuristics in computer chess. Our approach is based on the hypothesis that in any position, there is a definite relationship between control of space and its reinforcement and exploitation by good moves. These concepts are quantified and a quantitative relationship between them is realised by a multi-layer, feed-forward neural network trained on a large number of examples taken from master and grandmaster games. The output of the trained network is used to order legal moves according to their potential strength. A search heuristic incorporating the move ordering algorithm directs game-tree search away from less promising lines of play and towards more promising ones.
