Near-Optimal Reinforcement Learning with Self-Play