Potential Applications of Opponent-Model Search1

Abstract

An opponent is modelled by assumed knowledge of his evaluation of positions in a game. Exploiting this knowledge and assuming the opponent to be fallible, the opponent may be outwitted by anticipating his errors. Though the moves so generated need not be optimal in some minimax sense, the model may confer an advantage to the modelling player. Conditions are derived for what is, in essence, a minimum distance between the two player’s strategies; notably, an impetuous opponent is seen to labour under the same disadvantage as one with shallower search depth.