The Optimal Reward Baseline for Gradient-Based Reinforcement Learning