| Learning to play in the presence of independent and self-motivated opponents is a difficult task, because the dynamics of the environment appear to be non-stationary. In recent years there has been considerable amount of research in the field of Multi-Agent Learning, and some of this work has been in the context of Reinforcement Learning. One commonly used approach has been to restrict the opponent to a class of computationally bounded players, creating a parametrized model of it, and then search the model that can best explain the observed opponent behavior. In this thesis we study the problem of Reinforcement Learning in Stochastic Games, and propose two models for the opponent and two search algorithms, one based on Tests of Significance and the other on Maximum a Posteriori probabilities. We integrate the modeled opponent into a Markovian environment, and present an algorithm for solving the resulting MDP. Finally, we perform experiments on the effectiveness of the search algorithms. |
