A Bayesian approach to relevance in game playing

1. The point of game tree search is to insulate oneself from errors in the evaluation function. The standard approach is to grow a full width tree as deep as time allows, and then value the tree as if the leaf evaluations were exact. This has been eeective in many games because of the computational eeciency of the alpha-beta algorithm. Our approach is to form a Bayesian model of our uncertainty. We adopt an evaluation function that returns a probability distribution estimating the probability of various errors in valuing each position. These estimates are obtained by training from data. We thus use additional information at each leaf not available to the standard approach. We utilize this information in three ways: to evaluate which move is best after we are done expanding, to allocate additional thinking time to moves where additional time is most relevant to game outcome, and, perhaps most importantly, to expand the tree along the most relevant lines. Our measure of the relevance of expanding a given leaf provably approximates a measure of the impact of expanding the leaf on expected payoo, including the impact of the outcome of the leaf expansion on later expansion decisions. Our algorithms run (under reasonable assumptions) in time linear in the size of the nal tree and hence except for a small constant factor, are as time eecient as alpha-beta. In a given amount of time our algorithm can thus in principle grow a tree several times as deep as alpha-beta along the relevant lines. We have tested our approach on a variety of games, including Othello, Kalah, Warri, and others. Our probability independence approximations are seen to be signiicantly violated, but nonetheless our tree valuation scheme was found to play signiicantly better than minimax or the Probability Product rule when both competitors search the same tree. Our full search algorithm was found to outplay a highly ranked, directly comparable alpha-beta Othello program even when the alpha-beta program was given sizeable time odds, and also performed well against the three top Othello programs on the Internet Othello Server.