Toward Emphatic Reinforcement Learning