Learning Patterns for Playing Strategies 1

Abstract

We describe a first-order inductive framework, PAL, capable of learning chess patterns from a combination of two sources, viz. general-purpose knowledge about chess and simple example descriptions supplied to it. It is believed to be the first time useful chess patterns have been so learned. In order to establish that patterns so learned are applicable in play, a simple playing strategy for the King and Rook against King (KRK) endgame has constructed with patterns learned by PAL. A sketch of PAL’s limitations and of first-order inductive frameworks’ limitations in general is given; in chess-like domains, restrictions are essential for limiting the number of clauses required for induction and for guiding the search for them. Conclusions are given about the state of pattern learning achieved and achievable with presently available means.