The Crossroads of LLM and Traffic Control: A Study on Large Language Models in Adaptive Traffic Signal Control

Recent advancements in Large Language Models (LLMs) have ushered in opportunities to craft agents that exhibit human-like cognitive abilities, notably reasoning and planning. Leveraging the vast knowledge bases and sophisticated understanding of interrelations inherent in LLMs, this study explores their application in the dynamic and complex realm of traffic management. Specifically, we focus on Adaptive Traffic Control Systems (ATCS) as a critical use case, where designing effective traffic controllers parallels the cognitive tasks of reasoning and planning. This paper introduces a novel LLM-based agent development framework to create two types of traffic controllers: a Zero-Shot Chain of Thought approach and a Generally Capable Agent (GCA) approach. The latter integrates new knowledge from environmental interactions to enhance reasoning and planning capabilities. We implement and compare these controllers within a simulated traffic flow scenario at a single intersection using the Simulation of Urban Mobility (SUMO) against conventional traffic control methods, including fixed-time, gap-based, and delay-based controllers. Our results demonstrate that GCA-based controllers notably outperform traditional systems, reducing halted vehicle numbers by 48.03% and increasing average vehicle speed by 25.29%. These controllers also exhibit superior flexibility, generating diverse phase patterns that adapt dynamically to changing traffic conditions. This study underscores the transformative potential of LLMs in traffic management, promising significant enhancements in efficiency, responsiveness, and versatility in urban traffic control, ultimately improving urban life by mitigating traffic congestion’s economic and environmental impacts.