Toward Virtuous Reinforcement Learning