This is the source code of COMP 767 group project of Dingyi Zhuang (ID: 260894334) and Zhenyuan Ma (ID: 260881867). Our topic is about Traffic Light Control in Special Scenario.

The simulation environment is based on Berkeley Flow project. If you want to implement this code, please follow the installation instruction of Flow.

Our implementation of traffic light control is divided into single- and multi-agent environments. The code structure and containing files are

├─ project/
│  ├─ network/
│  ├─ tl_custom_env.py
│  ├─ tl_env.py
│  ├─ tl_main.py
│  └─ tl_net.py
│
├─ project_ma/
│  ├─ network/
│  ├─ tl_custom_env.py
│  ├─ tl_custom_env_ma.py
│  ├─ tl_env_ma.py
│  ├─ tl_main_ma.py
│  └─ tl_net_ma.py
│
└─ figs/

The suffix _ma is for multi-agent, otherwise single-agent. Here are some explanations of the scripts:

• tl_custom_env.py defines the base class for training RL algorithms in traffic light control scenarios. While tl_custom_env_ma.py defines the multi-agent training environment for multi-agent RL training.
• tl_env*.py defines the flow parameters for simulation, developed by Berkeley Flow project.
• tl_main*.py is the main training script, defines the training algorithms and policy graphs.
• tl_net*.py defines the traffic network parameters to suit our self-defined scenario.
• network/ contains our self-designed highway ramp special scenario with different infrastructure settings.

Our special scenario combines the regular crossroads and highway offramp.

Watch our video in Youtube

Only people in McGill University can access this link!

RLLib training results

Raw simulation video

• Dingyi Zhuang (dingyi.zhuang@mail.mcgill.ca)
• Zhenyuan Ma (zhenyuan.ma@mail.mcgill.ca)