Reinforcement learning studies the problem of solving sequential decision making problems. Model-based methods learn an effective policy in few actions by learning a model of the domain and simulating experience in their models. Typical model-based methods must visit each state at least once, which can be infeasible in large domains. To overcome this problem, the model learning algorithm needs to generalize knowledge to unseen states and provide information about the states in which it needs more experience. In this paper, we use existing supervised learning techniques to learn the model of the domain. We empirically compare their effectiveness at generalizing knowledge across states on three different domains. Our results indicate that tree-based models perform the best after training on a small number of transitions, while support vector machines perform the best after a large number of transitions.
In Proceedings of the ICML/UAI/COLT Workshop on Abstraction in Reinforcement Learning, June 2009.

Todd Hester Postdoctoral Alumni todd [at] cs utexas edu
Peter Stone Faculty pstone [at] cs utexas edu