Reasoning About Mdps Abstractly: Bayesian Policy Search With Uncertain Prior Knowledge