Generative Adversarial Self-Imitation Learning