Research work · 1994

Learning Optimal Chess Strategies

M Bain, S Muggleton

Machine Intelligence 13

Abstract

Abstract Move-perfect databases for chess endgames can be constructed by full-width backup from won positions. Complete tabulations are available for certain endgames, such as King and Rook against King (KRK), which contain optimal depth-to-win information. Previous work has applied decision-tree learning to construct rules for the classification of positions from such databases as won or not won, but with limited success. The current work takes an Inductive Logic Programming approach, using methods of generalization in first-order logic and specialization by predicate invention. Results are given on learning rules from example black-to-move KRK positions which are won-for-white in a fixed number of moves.

DOI: 10.1093/oso/9780198538509.003.0012