Jing Zhou

Jing Zhou

Auto-generated profile, based on publicly available information

Jing Zhou

Dalian Navy Academy, Dalian, Liaoning, China

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
10 April 2026
Verified
Conference Paper
Open access
HLHao Long
Hao Long
JZJing Zhou
Jing Zhou
WJWang junsen
Wang junsen
To address the challenges of low exploration efficiency, slow convergence speed, and poor decision interpretability in traditional Deep Q-Network (DQN) for maze navigation tasks, this paper proposes a maze navigation decision method based on Large Language Model(LLM) collaborative reinforcement learning. This method leverages LLM to perform dual evaluation of RL agent's single-step decisions(critical step quality) and global trajectories(path quality), and incorporates the evaluation results as priority weights into the experience replay mechanism to achieve priority learning of high-quality experiences. By constructing an LLM-RL collaborative decision framework, utilizing LLM for dual semantic evaluation of RL agent's single-step decisions and global trajectories, integrating evaluation results as priority weights into the experience replay mechanism, and designing a heuristic+LLM verification two-stage analysis mode to reduce computational overhead, priority learning of high-quality experiences is achieved. Through comparative experiments in a 6×6 grid maze environment, results show that compared to standard RL, Prioritized Experience Replay(PER), and Temporal Difference(TD) sampling methods, the proposed method improves navigation success rate by over 25% in the early training stage, reduces the number of training episodes required to reach 80% of final performance by nearly 40%, while maintaining the same optimal asymptotic performance and significantly enhancing decision interpretability. This validates the effectiveness of LLM-enhanced experience replay in improving the sample efficiency of DQN, and provides a practical reference for the fusion of large language models and value-based reinforcement learning.