Using Generative Adversarial Nets on Atari Games for Feature Extraction in Deep Reinforcement Learning