Interpretable Dynamics Models for Data-Efficient Reinforcement Learning