Probabilistic Models for Data-efficient Reinforcement Learning