Reinforcement learning solve Atari game