Evolving Rewards to Automate Reinforcement Learning