Learning long-term chess strategies from databases

摘要

We propose an approach to the learning of long-term plans for playing chess endgames. We assume that a computer-generated database for an endgame is available, such as the king and rook vs. king, or king and queen vs. king and rook endgame. For each position in the endgame, the database gives the “value” of the position in terms of the minimum number of moves needed by the stronger side to win given that both sides play optimally. We propose a method for automatically dividing the endgame into stages characterised by different objectives of play. For each stage of such a game plan, a stage-specific evaluation function is induced, to be used by minimax search when playing the endgame. We aim at learning playing strategies that give good insight into the principles of playing specific endgames. Games played by these strategies should resemble human expert’s play in achieving goals and subgoals reliably, but not necessarily as quickly as possible.