Greedy Algorithms for Sparse Reinforcement Learning