Optimal Control with Natural Images: Efficient Reinforcement Learning using Overcomplete Sparse Codes