Deep Reinforcement Learning with Guaranteed Performance