Learning Models of Opponent's Strategy Game Playing
David Cannel, Shaul Markovitch · 1993
Most of tile aclivity ill tile area of game playing progran,s is concerued wit h efficient ways of searching game troos. There is substantial evidence that game playing involves additiolml types of intelligent processes. One such pl-OCoss porformed by human experts is the acquisil ion and usage of a model of their opponent’s strategy. This work studies the problem of opponent mod-,.lling in game playing. A simplified version of a model is d,’tim’d as a pair of search depth and evaluation funclion. M*. a generalization of the minimax algorithm thai can handle an opponent model, is described. The Iwm’tiI of using opllouent models is demonstrated by comparing the performance of hi * with that of the traditional miuimax algoritlml. An algorithm for learning lho opponent’s strategy using its moves as examples was developed. Experiments demonstrated its ability Io acquire very accurate models. Finally, a full modelh’arniug game-playing system was developed and exp,,rimontally demonstrated to have advantage over nouh,arning player. 1