Game theory-based opponent modeling in large imperfect-information games

We develop an algorithm for opponent modeling in large extensive-form games of imperfect information. It works by observing the opponent’s action frequencies and building an opponent model by combining information from a precomputed equilibrium strategy with the observations. It then computes and plays a best response to this opponent model; the opponent model and best response are both updated continually in real time. The approach combines gametheoretic reasoning and pure opponent modeling, yielding a hybrid that can effectively exploit opponents after only a small number of interactions. Unlike prior opponent modeling approaches, ours is fundamentally game theoretic and takes advantage of recent algorithms for automated abstraction and equilibrium computation rather than relying on domain-specific prior distributions, historical data, or a handcrafted set of features. Experiments show that our algorithm leads to significantly higher win rates (than an approximateequilibrium strategy) against several opponents in limit Texas Hold’em — the most studied imperfect-information game in computer science — including competitors from recent AAAI computer poker competitions.