A Deep Reinforcement Learning Approach for Finding Non-Exploitable Strategies in Two-Player Atari Games