Transferable Deep Reinforce Learning With Limited Target Data

While there has been significant progress in the application of transfer learning within reinforcement learning, most existing research primarily focuses on online reinforcement learning, with few studies addressing the optimization of offline reinforcement learning. This paper proposes a novel transferable deep reinforcement learning method designed to enhance limited offline samples in the target domain by utilizing online interactive samples from a similar source domain. The method is versatile and can be integrated with any offline reinforcement learning algorithm. To validate its effectiveness, extensive evaluations were conducted across multiple Mujoco environments and D4RL datasets. The results indicate that by employing the proposed method, the algorithm's performance can approach that of using 100% of the target domain data, even when only 10% of the data is available.