Self-Imitation Learning

This paper proposes Self-Imitation Learning (SIL), a method that learns to imitate the agent’s past good trajectories and when combined with an actor-critic architecture can achieve better exploration and better performance. Specifically, our empirical results show that SIL improves the advantage actor-critic (A2C) on several hard exploration Atari games and is competitive to the state-of-the-art count-based exploration methods. We also show that SIL improves proximal policy optimization (PPO) on continuous control tasks.

• Machine Intelligence