Sample Trajectory Selection Method Based on Large Language Model in Reinforcement Learning

This paper introduces a method for trajectory selection using large-scale pre-trained language models, aiming to improve sample and training efficiency in reinforcement learning. By using a carefully designed prompt, the large language model can fully utilize its prior knowledge, effectively understanding and assessing the quality of trajectories produced through agent-environment interactions in reinforcement learning. This approach allows selecting more informative trajectories for the current agent’s learning. Unlike other works that indirectly improve reinforcement learning training efficiency by generating actions or decisions through large language models, our method employs these models to choose high-quality trajectories, thereby enhancing sample efficiency in reinforcement learning more directly. The approach was evaluated across multiple benchmark tasks in OpenAI’s Gym and RLcard. The results indicate a significant reduction in the number of environment interactions, and a 37% increase in the average reward compared to the original method.