Domain Adaptation for Reinforcement Learning on the Atari