Learning Diverse Policies with Soft Self-Generated Guidance

Abstract

Reinforcement learning (RL) with sparse and deceptive rewards is a significant challenge because nonzero rewards are rarely obtained, and hence, the gradient calculated by the agent can be stochastic and without valid information. Recent work demonstrates that using memory buffers of previous experiences can lead to a more efficient learning process. However, existing methods usually require these experiences to be successful and may overly exploit them, which can cause the agent to adopt suboptimal behaviors. This study develops an approach that exploits diverse past trajectories for faster and more efficient online RL, even if these trajectories are suboptimal or not highly rewarded. The proposed algorithm merges a policy improvement step with an additional policy exploration step by using offline demonstration data. The main contribution of this study is that by regarding diverse past trajectories as guidance, instead of imitating them, our method directs its policy to follow and expand past trajectories, while still being able to learn without rewards and gradually approach optimality. Furthermore, a novel diversity measurement is introduced to maintain the diversity of the team and regulate exploration. The proposed algorithm is evaluated on a series of discrete and continuous control tasks with sparse and deceptive rewards. In comparison with the existing RL methods, the experimental results indicate that our proposed algorithm is significantly better than the baseline methods in terms of diverse exploration and avoiding local optima.