Bayesian Inference for Data-Efficient Reinforcement Learning