Opponent modeling in a PGM framework

We consider the situation where two agents try to solve their own task in a common environment. In particular, we study a type of sequential Bayesian games with unlimited time horizon where two players share a visible scene, but the tasks (termed assignments) of the players are private information. We present a probabilistic graphical model (PGM), together with recursive modeling techniques, for representing this type of games. We introduce the type tree which can be used to calculate policies and to efficiently approximate the opponent’s state of belief. Secondly, we propose a method for reasoning with a mixture of models when a true model of opponent agents is unknown in the game due to their private information. We demonstrate its performance in a Grid game.