Potential Applications of Opponent-Model Search1

Opponent-Model Search in Games with Incomplete Information

“…Opponent models can come in diverse forms. Iida et al (1993Iida et al ( , 1994 propose opponent models for games with perfect information, where models are given by the evaluation function and the search depth of the opponent. Sturtevant, Zinkevich, and Bowling (2006) propose opponent models given by opponents' preferences over the outcomes of a game.…”

“…In addition, given a strategy represented by a mixed strategy or another linear representation (e.g. sequence form (Koller and Megiddo 1992) or evaluation function (Iida et al 1993)), its equivalent behaviour strategy can be computed in linear time.…”

“…This algorithm can be considered as a generalisation of OM search as proposed by Iida et al (1993), which only considers games with perfect information for which OMs are described by MIN's evaluation functions.…”

“…On the other hand, computing maxmin strategies for the original game tree without using the best-defence model is not ideal either, for these strategies fail to exploit any assumption one may have about their adversary, such as that they have limited computational power or reasoning depth, or that they have a predictable behaviour. Such assumptions make sense in particular when playing against humans (Iida et al 1993;Stahl and Wilson 1995;Dhami 2019).…”

“…Such models have been explicitly incorporated into game tree search algorithms (e.g. minimax, αβ search, MCTS) for games with perfect information (Iida et al 1993(Iida et al , 1994Sturtevant, Zinkevich, and Bowling 2006) to find robust strategies against a given opponent model, i.e. strategies that guarantee a given payoff against any strategy deemed possible by the opponent model.…”