Data-Efficient Reinforcement Learning with Probabilistic Models