A Doubly Robust Approach to Sparse Reinforcement Learning