Survey on Large Language Model-Enhanced Reinforcement Learning: Concept, Taxonomy, and Methods

Survey on Large Language Model-Enhanced Reinforcement Learning: Concept, Taxonomy, and Methods

VOL. X, NO. X, XXX 2024 | Yuji Cao, Huan Zhao, Member, IEEE, Yuheng Cheng, Student Member, IEEE, Ting Shu, Guolong Liu, Member, IEEE, Gaoqi Liang, Member, IEEE, Junhua Zhao, Senior Member, IEEE, Yun Li, Fellow, IEEE
This survey provides a comprehensive review of the emerging field of *LLM-enhanced Reinforcement Learning (RL)*, aiming to clarify the research scope and directions for future studies. The authors define *LLM-enhanced RL* as methods that utilize the multi-modal information processing, generating, reasoning, and other capabilities of pre-trained, knowledge-inherent AI models to assist the RL paradigm. They propose a structured taxonomy to categorize LLM functionalities in RL, including four roles: information processor, reward designer, decision-maker, and generator. Each role is detailed with methodologies, specific RL challenges addressed, and future directions. The survey highlights the potential applications, opportunities, and challenges of LLM-enhanced RL, emphasizing the need for a unified framework to integrate LLMs into the RL paradigm. The contributions of the survey include defining the *LLM-enhanced RL* paradigm, proposing a unified taxonomy, and reviewing algorithmic advancements in each LLM role. The survey also discusses the characteristics and characteristics of LLM-enhanced RL, such as multi-modal information understanding, multi-task learning, improved sample efficiency, long-horizon handling, and reward signal generation. Finally, it outlines potential future research directions, focusing on improving the generalization and adaptability of LLM-generated rewards and enhancing the effectiveness of explicit reward code generation.This survey provides a comprehensive review of the emerging field of *LLM-enhanced Reinforcement Learning (RL)*, aiming to clarify the research scope and directions for future studies. The authors define *LLM-enhanced RL* as methods that utilize the multi-modal information processing, generating, reasoning, and other capabilities of pre-trained, knowledge-inherent AI models to assist the RL paradigm. They propose a structured taxonomy to categorize LLM functionalities in RL, including four roles: information processor, reward designer, decision-maker, and generator. Each role is detailed with methodologies, specific RL challenges addressed, and future directions. The survey highlights the potential applications, opportunities, and challenges of LLM-enhanced RL, emphasizing the need for a unified framework to integrate LLMs into the RL paradigm. The contributions of the survey include defining the *LLM-enhanced RL* paradigm, proposing a unified taxonomy, and reviewing algorithmic advancements in each LLM role. The survey also discusses the characteristics and characteristics of LLM-enhanced RL, such as multi-modal information understanding, multi-task learning, improved sample efficiency, long-horizon handling, and reward signal generation. Finally, it outlines potential future research directions, focusing on improving the generalization and adaptability of LLM-generated rewards and enhancing the effectiveness of explicit reward code generation.
Reach us at info@study.space
[slides and audio] Survey on Large Language Model-Enhanced Reinforcement Learning%3A Concept%2C Taxonomy%2C and Methods