BROWSE
Detail
Conference Paper
LLM-Based Environment Perception and Candidate Action Ranking for Reinforcement Learning
Cited - time in 
Share 
Share
- Authors
- Joonyoung Jung, Dong-oh Kang
- Issue Date
- 2026-10
- Citation
- International Conference on Information and Communication Technology Convergence (ICTC) 2026, pp.1-3
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- Reinforcement learning (RL) is a representative learning approach for solving sequential decision-making problems and has been widely applied in various fields, including autonomous driving and robotic control. However, in real-world environments, the state space and action space are often extremely large, resulting in inefficient exploration during the learning process and slow policy convergence. Recently, large language models (LLMs) have demonstrated the potential to select appropriate actions in complex situations based on semantic understanding of environments and commonsense reasoning. In this paper, we propose an LLM-based RL framework that leverages environment perception and candidate action ranking. The proposed method converts simulator state information into natural language-based environment descriptions and uses an LLM to generate a set of candidate actions that can be performed in the current state. In addition, it estimates the execution probability of each generated candidate action and incorporates these probabilities into proximal policy optimization (PPO)-based policy learning. Through this process, the proposed method can effectively reduce the exploration space and improve the efficiency of policy learning compared with RL methods that explore the entire action space. Experimental results obtained using the VirtualHome simulator demonstrate that the proposed framework effectively understands environmental states, generates meaningful candidate actions, and improves policy convergence in PPO-based RL.
- Keyword
- reinforcement learning, large language model, candidate action, proximal policy optimization, VirtualHome
- KSP Keywords
- Action space, Commonsense reasoning, Complex situations, Current state, Environment perception, Language Models, Learning Process, Learning approach, Natural language, Policy learning, Policy optimization