Large Language Models Achieve Smarter Traffic Control during Unforeseen Incidents

Adaptive traffic signal control struggles to cope with sudden incidents like accidents and roadworks, often necessitating slow and inefficient manual adjustments. Shiqi Wei, Qiqing Wang, and Kaidi Yang from the National University of Singapore, along with their colleagues, present a novel approach leveraging the power of Large Language Models (LLMs) to create a ‘virtual traffic police’ system. Unlike previous attempts to replace existing infrastructure, this research augments current traffic signal control with an LLM agent that intelligently fine-tunes signal parameters in response to real-time events. By combining a self-refined traffic language system with an LLM-based verifier, the team demonstrates a significantly more reliable and efficient method for managing unforeseen disruptions, paving the way for smarter and more responsive urban traffic networks.

This breakthrough avoids the costly and unreliable practice of replacing established systems with entirely LLM-based solutions, instead leveraging the strengths of both approaches. The team achieved this by designing a self-refined traffic language retrieval system (TLRS), employing retrieval-augmented generation to draw upon a tailored database encompassing traffic conditions and controller operation principles.

The study unveils a system where the LLM agent operates at a higher level, responding to real-time incident reports and adjusting lower-level signal controllers, a process that significantly improves operational efficiency and reliability. To enhance the trustworthiness of the LLM, the researchers devised an LLM-based verifier that continuously updates the TLRS during the reasoning process, mitigating the risk of ‘hallucinations’ and ensuring domain-specific accuracy. Experiments show that this framework allows LLMs to function as dependable virtual traffic officers, adapting conventional TSC methods to unexpected events with a level of performance previously unattainable. This innovative approach addresses the limitations of current adaptive TSC methods, which often struggle with incidents beyond their pre-programmed parameters and rely on slow, labor-intensive manual interventions by traffic police.
Furthermore, the research establishes a method for continuous improvement of the traffic language database, ensuring the LLM remains informed about evolving traffic patterns and control strategies. The TLRS, combined with the LLM verifier, creates a feedback loop that refines the system’s knowledge base over time, enhancing its ability to handle a wider range of unforeseen incidents. This work opens new avenues for intelligent transportation systems, moving beyond reactive responses to proactive adaptation, and potentially reducing congestion, improving safety, and lowering operational costs for traffic management authorities. The team’s findings demonstrate that LLMs can be effectively harnessed to create a more resilient and efficient urban transportation network.

LLMs augment traffic control via retrieval-augmented generation

Scientists developed a hierarchical framework augmenting existing adaptive traffic signal control (TSC) systems with large language models (LLMs) to address inefficiencies caused by unforeseen traffic incidents like accidents and road maintenance. The research team engineered a virtual traffic police agent operating at a higher level, dynamically fine-tuning parameters of lower-level signal controllers in real-time response to disruptions. This innovative approach moves beyond complete system replacement, mitigating the risks associated with LLM hallucinations and reducing implementation costs. Experiments employ a retrieval-augmented generation technique, drawing knowledge from a tailored traffic language database encompassing both traffic conditions and controller operation principles.

To enhance the reliability of the LLM in domain-specific scenarios, the study pioneered a self-refined traffic language retrieval system (TLRS). This system utilises retrieval-augmented generation to access a curated database containing detailed information on traffic states and signal controller functionalities. Moreover, researchers devised an LLM-based verifier that continuously updates the TLRS throughout the reasoning process, ensuring the agent’s knowledge base remains current and accurate. The team harnessed this verifier to refine the retrieval process, improving the quality and relevance of information provided to the LLM agent.

The core of the methodology centres on the LLM’s ability to interpret incident reports and translate them into actionable parameter adjustments for the traffic signal controllers. The system delivers a nuanced response, avoiding blanket changes and instead focusing on targeted modifications to optimise traffic flow around the incident. This process involves the LLM agent analysing real-time traffic data, identifying the nature and location of the incident, and then generating a set of recommended parameter changes, such as green time adjustments or cycle length modifications, for the affected intersection. The TLRS provides the LLM with the necessary contextual information to make informed decisions, while the verifier ensures the recommendations align with established traffic control principles.

Results demonstrate that LLMs can function as trustworthy virtual traffic police officers, adapting conventional TSC methods to unforeseen incidents with significantly improved operational efficiency and reliability. The study’s findings indicate a substantial reduction in vehicle delay and improved safety compared to traditional manual intervention methods. This innovative framework offers a promising solution for enhancing urban traffic management, enabling a more proactive and responsive approach to handling unexpected disruptions and optimising traffic flow under dynamic conditions. Experiments revealed that this approach significantly improves operational efficiency and reliability compared to conventional manual interventions. The core innovation lies in the Self-Refined Traffic Language Retrieval System (TLRS), which employs retrieval-augmented generation to draw knowledge from a tailored database encompassing traffic conditions and controller operation principles.

The team devised an LLM-based verifier that continuously updates the TLRS during the reasoning process, enhancing domain-specific reliability. Results demonstrate that the LLM agent successfully adapts conventional TSC methods to unforeseen incidents, avoiding the need for costly and unreliable system replacements. Measurements confirm the system’s ability to address limitations of existing adaptive TSC methods, which often struggle with diverse and unexpected incidents. The work addresses the critical need for rapid response to traffic disruptions, currently reliant on labor-intensive manual adjustments by traffic police officers, a process often slow and prone to delays.

Scientists recorded substantial improvements in traffic flow management by leveraging the LLM’s reasoning capabilities to adjust signal timings based on incident specifics. The TLRS, a key component of the system, accesses a database of traffic conditions and controller operation principles, enabling informed decision-making. Tests prove that the LLM-based verifier effectively refines the TLRS, ensuring the accuracy and consistency of the information used for signal control. This continuous refinement process mitigates the risk of LLM hallucinations, a crucial factor for safety-critical applications like traffic management.

Furthermore, the breakthrough delivers a solution that avoids the high costs associated with replacing existing traffic control infrastructure. The team’s framework seamlessly integrates with current systems, allowing for a more practical and scalable implementation. Measurements confirm the system’s responsiveness, enabling sub-second-level adjustments to phase transitions, a critical requirement for effective traffic signal control.