Abstract

Traffic Signal Control (TSC) is a crucial component in urban trafficmanagement, aiming to optimize road network efficiency and reduce congestion.Traditional TSC methods, primarily based on transportation engineering andreinforcement learning (RL), often struggle with generalization abilitiesacross varied traffic scenarios and lack interpretability. This paper presentsLLMLight, a novel framework employing Large Language Models (LLMs) asdecision-making agents for TSC. Specifically, the framework begins byinstructing the LLM with a knowledgeable prompt detailing real-time trafficconditions. Leveraging the advanced generalization capabilities of LLMs,LLMLight engages a reasoning and decision-making process akin to humanintuition for effective traffic control. Moreover, we build LightGPT, aspecialized backbone LLM tailored for TSC tasks. By learning nuanced trafficpatterns and control strategies, LightGPT enhances the LLMLight frameworkcost-effectively. Extensive experiments conducted on ten real-world andsynthetic datasets, along with evaluations by fifteen human experts,demonstrate the exceptional effectiveness, generalization ability, andinterpretability of LLMLight with LightGPT, outperforming nine baseline methodsand ten advanced LLMs.