Structure learning with Temporal Gaussian Mixture for model-based Reinforcement Learning