It is often assumed that agents in multiagent systems with state uncertainty have full knowledge of the model of dy- namics and sensors, but in many cases this is not feasible. A more realistic assumption is that agents must learn about the environment and other agents while acting. Bayesian methods for reinforcement learning are promising for this type of learning because they allow model uncertainty to be considered explicitly and offer a principled way of dealing with the exploration/exploitation tradeoff. In this paper, we propose a Bayesian RL framework for best response learn- ing in which an agent has uncertainty over the environment and the policies of the other agents. This is a very general model that can incorporate different assumptions about the form of other policies. We seek to maximize performance and learn the appropriate models while acting in an online fashion by using sample-based planning built from power- ful Monte-Carlo tree search methods. We discuss the theo- retical properties of this approach and experimental results show that the learning approaches can significantly increase value when compared to initial models and policies.
展开▼
