Balancing Multiple Sources of Reward in Reinforcement Learning