A Survey on Transfer Reinforcement Learning

Abstract

The reinforcement learning(RL) paradigm enables machines to autonomously complete a series of tasks through continuous trial and error. With the development of deep learning technology, reinforcement learning has not only made breakthroughs in the gaming field, but also been widely applied in fields such as autonomous driving, and robot control. However, how to improve its sample efficiency is still a challenge for deep reinforcement learning(DRL). Transfer learning, as a technology that utilizes external professional knowledge to promote the learning process in the target domain, is a promising machine learning method to solve the above problems. In this survey, we systematically summarize the technology of transfer reinforcement learning(TRL) on single-agent and multi-agent systems. Moreover, we summarize the emerging trends in TRL, including meta-RL and self-supervised learning. Finally, we explore the latest applications of TRL algorithms. Through systematic comparisons between TRL and DRL paradigms, we quantify TRL's superiority in cross-task generalization and rapid adaptation.