Guaranteed Policy Performance In Reinforcement Learning