Q* Approximation for Batch Reinforcement Learning: A Theoretical Comparison