Nonparametric Bayesian Approaches for Reinforcement Learning in Partially Observable Domains