Using example-based reasoning for selective move generation in two player adversarial games

Abstract

Given the processing speed of the best chess computer is 200 million times faster in terms of positions evaluated per second than a human chess expert, a grandmaster, the question is not how can a computer beat a chess grandmaster but rather how do chess grandmasters beat computers? In computing terms, the human expert's strength lies in the ability to significantly prune the search tree and to correctly evaluate the resulting positions. This paper addresses the first of these strengths and proposes an example-based reasoning mechanism to select candidate moves from a given chess position. The mechanism automatically generates an example-base from a database of grandmaster chess games using Principal Component Analysis to characterise the positions in the games database. Given a new position, its characterisation is compared to those in the example-base and a ranked list of n “similar” moves is returned. This forms the basis of an effective forward pruning mechanism in a two player adversarial game.