1 Introduction

Official code for the article "Expression might be enough: representing pressure and demand for reinforcement learning based traffic signal control".

This article has been received by ICML 2022

If you use our method, please cite our article.

@inproceedings{advanced_xlight,
  title={Expression might be enough: representing pressure and demand for reinforcement learning based traffic signal control},
  author={Zhang, Liang and Wu, Qiang and Shen, Jun and L{\"u}, Linyuan and Du, Bo and Wu, Jianqing},
  booktitle={International Conference on Machine Learning},
  pages={26645--26654},
  year={2022},
  organization={PMLR}
}

The code structure is based on Efficient_XLight.

2 Requirements

python3.6,tensorflow=2.4, cityflow, pandas, numpy

cityflow needs a linux environment, and we run the code on Manjaro Linux.

3 Usage

Parameters are well-prepared, and you can run the code directly.

Our proposed method:

• For Advanced-MPLight, run:

python run_advanced_mplight.py

• For Advanced-CoLight, run:

python run_advanced_colight.py

• For Advanced-MP, run:

python run_advanced_maxpressure.py

For the baseline methods:

• Efficient-PressLight

python run_efficient_presslight.py

• Efficient-CoLight

python run_efficient_colight.py

• Efficient-MPLight`

python run_efficient_mplight.py

• Fixed-Time

python run_fixedtime.py

• Max-Pressure

python run_maxpressure.py

• PressLight

python run_presslight.py

• MPLight

python run_mplight.py

• Colight

python run_colight.py

3.1 Newly added baseline

We newly added two methods(AttendLight and PRGLight) as the basline.

PRGLight

The official code is avaliable at https://github.com/guang0123/PRGLight.git.

AttendLight

• For AttendLight, the officical code is not avaliable and we realized it according to their article. But, our-realized is not exactly the same as article described. We would like to run AttendLight in a DQN approach, but it is designed for Actor-Critic framework and cannot coverge under DQN framework.
• Through analysis, we find that the action-attention block cannot work well in DQN framework. Finally, we adopt the idea of AttentionLight(https://github.com/LiangZhang1996/AttentionLight.git) and use self-attention to replace the action-attention block.
• Our-realized AttenLight can converge and work well under all the datasets. It still retains the property as a universal model for any intersections.
• For fair comparison, AttendLight uses the same hyper-parameters with other methods.

4、Code details

4.1、structure

• models: contains all the models used in our article.
• utils: contains all the methods to simulate and train the models.

4.2、Reference

The code is modified from Efficient_XLight. The Max-Pressure is created by ourselves, based on MaxPressure .

• PressLight: Bsed on LIT model, which comes from Colight.
• Colight : Based on Colight.
• Fixed-Time: From MPLight.
• MPLight: From MPLight.

License

This project is licensed under the GNU General Public License version 3 (GPLv3) - see the LICENSE file for details.