Virtual Traffic Police: Large Language Model-Augmented Traffic Signal Control for Unforeseen Incidents

TL;DR

The paper tackles the challenge of unforeseen traffic incidents in adaptive signal control by introducing a two-level, LLM-augmented framework where a virtual traffic police agent at the top tunes the lower-level controller parameters $oldsymbol{ heta}_s$ in response to incident descriptions $m{E}_s$. Reliability is enhanced through a self-refined traffic language retrieval system (TLRS) that ground knowledge via retrieval-augmented generation and an LLM-based verifier that updates the TLRS with new experiences. The approach is evaluated in SUMO on four incident types with two lower-level controllers (max-pressure and MPC), showing that LLM-augmented TSC with CoT prompting and TLRS substantially outperforms baselines, especially under unseen incidents, by reducing average delays $AD$, queue lengths $AQL$, and pedestrian crossing measures like CCR. These results indicate that integrating LLM reasoning with conventional, high-frequency controllers can deliver reliable, scalable, and cost-efficient incident-responsive traffic management in practice, with potential extensions to richer knowledge graphs and safety guarantees.

Abstract

Adaptive traffic signal control (TSC) has demonstrated strong effectiveness in managing dynamic traffic flows. However, conventional methods often struggle when unforeseen traffic incidents occur (e.g., accidents and road maintenance), which typically require labor-intensive and inefficient manual interventions by traffic police officers. Large Language Models (LLMs) appear to be a promising solution thanks to their remarkable reasoning and generalization capabilities. Nevertheless, existing works often propose to replace existing TSC systems with LLM-based systems, which can be (i) unreliable due to the inherent hallucinations of LLMs and (ii) costly due to the need for system replacement. To address the issues of existing works, we propose a hierarchical framework that augments existing TSC systems with LLMs, whereby a virtual traffic police agent at the upper level dynamically fine-tunes selected parameters of signal controllers at the lower level in response to real-time traffic incidents. To enhance domain-specific reliability in response to unforeseen traffic incidents, we devise a self-refined traffic language retrieval system (TLRS), whereby retrieval-augmented generation is employed to draw knowledge from a tailored traffic language database that encompasses traffic conditions and controller operation principles. Moreover, we devise an LLM-based verifier to update the TLRS continuously over the reasoning process. Our results show that LLMs can serve as trustworthy virtual traffic police officers that can adapt conventional TSC methods to unforeseen traffic incidents with significantly improved operational efficiency and reliability.

Figures (10)

• Figure 1: Comparison of conventional management of unforeseen traffic incidents and our proposed LLM-augmented framework
• Figure 2: Illustration of our proposed incident-aware LLM-augmented TSC framework.
• Figure 3: Illustration of LLM-based generator and verifier prompts.
• Figure 4: Examples of incident reports (left) and their corresponding Q-A pair representations (right) stored in the Traffic Language Retrieval System (TLRS).
• Figure 5: Overview of the Self-refinement Traffic Language Retrieval System (SRTLRS). The system retrieves top-$k$ relevant historical Q-A pairs based on the embedding of the current traffic event $E_s$. The LLM Agent generates a control strategy, which is then evaluated by the LLM Verifier through judgment and feedback. Verified outputs are used to assist the controller and are also added as new Q-A pairs to continually refine the retrieval system.

Theorems & Definitions (5)

• Definition 1: Contextual description $\bm{E}_s$ of the real-time traffic incident
• Remark 1: Access to contextual description $\bm{E}_s$
• Definition 2: Zero-shot CoT prompting kojima2022large
• Remark 2: Extensibility of the Agent Design
• Definition 3: Traffic language database