This project introduces LLMLight, a framework that leverages Large Language Models (LLMs) as agents for Traffic Signal Control (TSC). It addresses limitations in traditional TSC methods, such as poor generalization across diverse traffic scenarios and lack of interpretability. LLMLight offers enhanced decision-making capabilities, akin to human intuition, and introduces LightGPT, a specialized LLM backbone optimized for TSC tasks, promising cost-effective and more adaptable traffic management solutions. The project is targeted at researchers and engineers in urban planning and intelligent transportation systems.

How It Works

LLMLight operates by feeding real-time traffic conditions to an LLM via carefully crafted prompts. The LLM then utilizes its advanced generalization and reasoning abilities to make traffic control decisions. A novel aspect is the development of LightGPT, a domain-specific LLM fine-tuned on nuanced traffic patterns and control strategies, which enhances the efficiency and cost-effectiveness of the LLMLight framework. This approach aims to provide more intuitive and adaptable traffic signal control compared to traditional engineering or reinforcement learning methods.

Quick Start & Requirements

• Primary Install/Run: Execute Python scripts directly (e.g., run_advanced_mplight.py, run_chatgpt.py, run_open_LLM.py).
• Prerequisites: Python >= 3.9, TensorFlow-CPU == 2.8.0, CityFlow, Pandas == 1.5.0, NumPy == 1.26.2, Wandb, Transformers == 4.45.0, PEFT == 0.7.1, Accelerate == 0.27.2, Datasets == 2.16.1, Fire, VLLM == 0.6.2. CityFlow requires a Linux environment (tested on Ubuntu). Running GPT-3.5/GPT-4 necessitates setting an API key.
• Links:
  • Website: https://sqlai2099.github.io/LLMLight-Page/
  • Demo Video: https://github.com/usail-hkust/LLMTSCS/assets/62106026/90567f61-5d58-4dac-8c8e-836a421f3ff9
  • LightGPT Models: https://huggingface.co/collections/usail-hkust/llmlight-lightgpt-673ac5a619cbbe309165b56d

Highlighted Details

• The project has been accepted by KDD'2025 and received a Special Award and Golden Medal at the International Exhibition Inventions of Geneva 2025.
• LLMLight demonstrates remarkable effectiveness, generalization, and interpretability across nine real-world and synthetic datasets, outperforming nine transportation-based and RL-based baselines.
• New LightGPT models, fine-tuned on Qwen2 and Llama3 backbones, are available on HuggingFace.
• The code structure is based on the Efficient_XLight framework.

Maintenance & Community

Information regarding specific maintainers, community channels (like Discord/Slack), or a public roadmap is not detailed in the provided README.

Licensing & Compatibility

The project is licensed under the MIT License, which is generally permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

Documentation and PyPI packaging are marked as "in progress," indicating the project is under active development. The core CityFlow simulator requires a Linux environment, specifically Ubuntu. Running proprietary LLMs like GPT-3.5/GPT-4 requires managing API keys. Utilizing VLLM for faster inference incurs higher GPU memory costs.