Reasoning about MDPs Abstractly: Bayesian Policy Search with Uncertain Prior Knowledge