TLDR: Chat2SPaT is an AI-powered tool that automates the creation and editing of traffic signal control plans. It uses large language models (LLMs) to interpret natural language descriptions of traffic plans and then employs Python scripts to generate precise signal phase and timing (SPaT) results. This significantly reduces manual effort for traffic engineers and achieves high accuracy (over 90%) in converting ambiguous descriptions into executable plans.
Managing traffic signal control plans has long been a complex and time-consuming task for traffic engineers. Traditional pre-timed traffic signal control, widely used for intersections and coordinated arterial roads, demands extensive manual input for creating and updating signal phase and timing (SPaT) parameters. This often involves handling multiple plans for a single intersection based on time-of-day or day-of-week, leading to repetitive and tedious work.
Introducing Chat2SPaT: AI for Traffic Signal Management
A new tool called Chat2SPaT has been proposed to address these challenges by automating the traffic signal control plan management process. Chat2SPaT leverages the power of large language models (LLMs) to convert semi-structured and often ambiguous user descriptions of signal control plans into precise SPaT results. These results can then be transformed into structured stage-based or ring-based plans, ready to interact with intelligent transportation system (ITS) software and traffic signal controllers.
The core idea behind Chat2SPaT is to make the process user-friendly. It utilizes LLMs’ natural language understanding capabilities to interpret how users describe a signal plan. With carefully designed prompts, the LLM reformulates the plan into a combination of phase sequence and phase attribute results, typically in a JSON format. Following this, Python scripts are employed to accurately locate phases within a cycle, handle the intricate nuances of traffic signal control, and finally assemble a complete and executable traffic signal control plan. The system is designed for iterative use, allowing for further plan editing within a continuous chat.
How Chat2SPaT Works
The methodology of Chat2SPaT involves a two-step workflow. First, prompts guide LLMs to understand user descriptions, considering various plan styles like stages, rings, and phase overlapping. To prevent incorrect reasoning, LLMs are instructed to record special phase treatments as key-value phase attributes, such as permissive left-turns or exclusive pedestrian crossings. The LLM then outputs three key results: phase sequence, phase attributes, and cycle length.
Second, Python scripts take these LLM outputs and perform data cleansing, standardize phase names, and replace placeholders. They calculate the timing for major phases based on the plan structure and integrate phase attributes. The scripts also include fault-tolerance functions to correct potential errors or inconsistencies from the LLM outputs, ensuring the generated plan is accurate. For instance, if a phase is described with both start/end times and a split duration, the system prioritizes the start and end times. The tool also handles complex scenarios like merging multiple occurrences of a phase or managing phase overlapping, especially for pedestrian and right-turn movements.
Validation and Performance
Experiments conducted on a test dataset of over 300 plan descriptions, in both English and Chinese, demonstrated Chat2SPaT’s effectiveness. The tool achieved an accuracy of over 94% for both English and Chinese cases when using top-performing LLMs like ChatGPT-4o and Qwen2.5-72B-Instruct. This makes Chat2SPaT the first benchmark for evaluating LLMs’ capability in understanding complex traffic signal control plan descriptions.
The study also highlighted insights into LLM behavior, categorizing errors into formatting errors, overthinking errors (where LLMs attempt unnecessary calculations), and semantic comprehension errors. The design of Chat2SPaT minimizes the need for LLMs to perform complex mathematical reasoning, offloading such tasks to specialized scripts, which enhances robustness and reduces the chance of hallucinations. This design also makes the tool more feasible for practical application, even with smaller LLMs.
Also Read:
- Large Language Models Prove Effective in Predicting Traffic Incident Impact
- AI’s New Frontier: Detecting Road Crashes with Language Models
Future Outlook
Chat2SPaT offers a promising new building block for more accurate and versatile applications of LLMs in the field of Intelligent Transportation Systems. It has the potential to be integrated into existing traffic signal control software or microscopic traffic simulation tools, significantly reducing the manual workload for traffic engineers. Future developments could include managing day-of-week and time-of-day plans, supporting more transportation modes like bikes and buses, and extending its capabilities to dynamic control systems by specifying detectors and adjusting control parameters for actuated controls. For more details, you can refer to the full research paper: Chat2SPaT: A Large Language Model Based Tool for Automating Traffic Signal Control Plan Management.
