Harnessing Discrete Representations for Continual Reinforcement Learning.