UniTSA: A Universal Reinforcement Learning Framework for V2X Traffic Signal Control

Paper-1, UniTSA | Paper-2, ADLight | Blog-1 | Blog-2 | Model Checkpoint Download

This repository contains the code for the paper "UniTSA: A Universal Reinforcement Learning Framework for V2X Traffic Signal Control".

🎉 News

1. Congratulations! Our research has been accepted by IEEE Transactions on Vehicular Technology. Read the paper here.
2. We have transitioned the simulation platform in the project from Aiolos to TransSimHub (TSHub). We extend our gratitude to our colleagues at SenseTime, @KanYuheng (阚宇衡), @MaZian (马子安), and @XuChengcheng (徐承成) (listed alphabetically) for their contributions. The development of TransSimHub (TSHub) is built upon the foundation of Aiolos.
3. The weights of the model have been uploaded. The uploaded model weights can be found in save_models and can be verified using eval_model.py.

🔑 Key Points

• Universal Model for Different Intersections: UniTSA uses a junction matrix to characterize different intersections, enabling the same model to be applied to various intersection designs.

(a) A 3-way intersection with (b) its junction matrix with zero padding.

• Performance Enhancement at Unseen Intersections: UniTSA incorporates traffic state augmentation techniques that emphasize the relative positioning of vehicles, enhancing the model's adaptability to changing traffic conditions and unfamiliar scenarios.

llustration of three traffic state augmentation methods applied to both 4-way and 3-way intersections.

• Improved Results at Key Intersections: UniTSA integrates the Low-Rank Adaptation (LoRA) method, allowing for efficient model customization to specific intersections with minimal additional training.

The overall framework of UniTSA, including: 1. RL Training and 2. Fine-tuning.

📥 Installation

Before using, make sure TSHub is installed.

git clone https://github.com/Traffic-Alpha/TransSimHub.git
cd TransSimHub
pip install -e ".[rl]"

🏃‍♂️ Training

After installation, run train_UniTSA.py to train the model.

python train_UniTSA.py

The trained reward curve is roughly as follows:

Note: There are large fluctuations here because the rewards vary greatly in different environments. You can rollout in advance to get the mean and variance for normalization.

The trained model will be stored in the save_models directory. This model can be applied to intersections of varying structures. For improved results under diverse traffic flow patterns, we recommend generating a wider variety of routes for training. Please note that currently, only a subset of rou files have been uploaded.

🧪 Testing

Testing can be done later. If you want to test directly on the road network in this article, you can run:

python eval_model.py

If you want to replace it with your own road network, you can put the road network in sumo_datasets, and then modify EVAL_CONFIG to modify the corresponding road network.

For example, you can modify the EVAL_CONFIG as follows:

EVAL_SUMO_CONFIG = dict(
    # Four-way intersection, lane count (3,3,3,3)
    test_four_34=dict(
        tls_id = 'J1',
        sumocfg = 'test_four_34.sumocfg',
        nets = ['4phases.net.xml'],
        routes = ['0.rou.xml', '1.rou.xml',],
        start_time = 0,
        edges = ['E0', '-E1', '-E3'],
        connections = {
            'WE-EW':['E0 E1', '-E1 -E0'],
            'NS-SN':['-E3 E2', '-E2 E3']
        }
    ),
)

Here:

• test_four_34: Modify this to the name of your new folder.
• sumocfg: This is the path to the sumocfg file, which needs to be in the env directory.
• nets: This is the name of the SUMO net file. There can be multiple files.
• routes: This is the file name of SUMO routes, which needs to be in the routes directory. There can be multiple files.

📚 Citation

If you find this work useful, please cite our papers:

@article{wang2024unitsa,
  author={Wang, Maonan and Xiong, Xi and Kan, Yuheng and Xu, Chengcheng and Pun, Man-On},
  journal={IEEE Transactions on Vehicular Technology}, 
  title={{UniTSA}: A Universal Reinforcement Learning Framework for V2X Traffic Signal Control}, 
  year={2024},
  volume={73},
  number={10},
  pages={14354-14369},
  doi={10.1109/TVT.2024.3403879}
}

The paper below presents the preliminary concept of UniTSA, specifically introducing the movement shuffle concept:

@inproceedings{wang2022adlight,
  title={{ADLight}: A Universal Approach of Traffic Signal Control with Augmented Data Using Reinforcement Learning},
  author={Wang, Maonan and Xu, Yutong and Xiong, Xi and Kan, Yuheng and Xu, Chengcheng and Pun, Man-On},
  booktitle={2023 Transportation Research Board Annual Meeting (102nd TRB)},
  year={2023},
}

📫 Contact

If you have any questions, please open an issue in this repository. We will respond as soon as possible.