LLM-enabled universal traffic signal control across different intersections and traffic flows

With the development of urbanization, the issue of traffic congestion is becoming increasingly severe, and traffic signal control (TSC) remains a challenging and significant task. While reinforcement learning (RL)-based TSC methods have shown good performance in dynamic and complex traffic flows, how to construct a universal TSC framework to adapt to intersections with different structures, phases, and traffic volumes remains a pressing matter for exploration. Given the powerful generalization, zero-shot learning, and knowledge transfer capabilities of large language models (LLMs), this paper directly employs LLMs as the TSC agent, proposing a pretrained LLMs-based generalized TSC framework. The framework transforms different intersection configurations into a unified state and action representation, leveraging existing RL methods to generate expert experience covering various intersection configurations. Guided by this, a resource-efficient fine-tuning strategy is deployed to tailor the LLMs to the intricacies of TSC tasks. Furthermore, bisimulation learning is harnessed to eradicate spurious correlations between states and actions within offline data, thereby fortifying the resilience of TSC across intersections of varying structures. To demonstrate the foundational performance and generalization capabilities of the proposed framework, we established training and testing datasets. The training dataset comprised eight different intersection configurations with a medium traffic flow input mode. The testing dataset expanded the traffic flow input modes to low, medium, and high levels, additionally testing other unseen intersection configurations. The experiments indicated that across intersections with varying traffic flow inputs, phases, and structures, whether trained or untrained, our framework consistently exhibited exceptional performance, particularly in high-traffic environments.