Abstract:

Adaptive traffic signal control (ATSC) is crucial for alleviating regional traffic congestion, yet it faces severe challenges in real-time response to unexpected events and global coordination. The DRL method relies on pure data-driven approaches, suffering from core limitations such as poor generalization, weak interpretability, and a lack of guidance from emergency disposal knowledge, which makes them difficult to meet the demands of complex traffic scenarios. A control system that integrates knowledge-driven and data-optimized approaches was proposed. The GraphRAG was used to construct a dynamic traffic knowledge graph, providing LLMs with real-time updated historical emergency disposal experience and road network topology knowledge. The LoRA fine-tuning technologywas adopted to accurately inject professional knowledge in the field of traffic signal control while avoiding the high cost of full-scale LLM training. A structured Prompt was also designed to concatenate regional road network states and relevant contexts, enabling LLMs to possess regional coordination and event response capabilities. Simulation experiments show that compared with traditional methods and the optimal reinforcement learning baseline method, this approach reduces the average travel time by 9.85%, 4.87%, and 17.4% respectively, shortens the average waiting time by 4.17% (Ji′nan dataset) and 6.76% (Random scenario) respectively, and increases the throughput by 0.95%, 3.71%, and 8.53% respectively. In addition, the performance improvement is particularly significant in the Random scenario.

Key words: traffic signal control, knowledge graph, graph retrieval-augmented generation, fine-tuning, LLMs, RL