Contrastive Learning Methods for Deep Reinforcement Learning

摘要

Deep reinforcement learning (DRL) has shown promising performance in various application areas (e.g., games and autonomous vehicles). Experience replay buffer strategy and parallel learning strategy are widely used to boost the performances of offline and online deep reinforcement learning algorithms. However, state-action distributions shift leads to bootstrap errors. Experience replay buffer learns policies with elder experience trajectories, limiting its application to off-policy algorithms. Balancing the new and the old experience is challenging. Parallel learning strategies can train policies with online experiences. However, parallel environmental instances organize the agent pool inefficiently with higher simulation or physical costs. To overcome these shortcomings, we develop four lightweight and effective DRL algorithms, instance-actor, parallel-actor, instance-critic, and parallel-critic methods, to contrast different-age trajectory experiences. We train the contrast DRL according to the received rewards and proposed contrast loss, which is calculated by designed positive/negative keys. Our benchmark experiments using PyBullet robotics environments show that our proposed algorithm matches or is better than the state-of-the-art DRL algorithms.