Traffic Control RL

Results

Overview

This project implements deep reinforcement learning to optimize traffic signal timing. Using SUMO with RL algorithms, the system learns traffic light patterns to minimize congestion and improve traffic flow. To exploit learning gains from single-intersection traffic tasks, the DuelDQN model is modified for easier weight transferring. This is done since the observation can be made that internal logic for n x n traffic control scenarios use similar internal logic, only distinguished by different observation spaces and action spaces. Using base_policy_weights.pth, derived from 2way-single-intersection-vhvh.sh experiment, traffic control agent(s) are able to converge quickly and perform well under previously unseen traffic scenarios.

For traffic scenarios with more than one intersection, multiple TrafficAgent models are instantiated to handle each intersection in a n x n grid.

Installation

git clone https://github.com/yourusername/Traffic-Control-RL.git
cd Traffic-Control-RL
sudo bash setup.sh

Usage

Training

For Training, you can trigger the training process for a given net and rou file using the --train flag:

python src/main.py --config src/config.yaml --net path/to/net.xml --route path/to/rou.xml --path models/single-intersection --train

Testing

For Testing, you can trigger the render process for a given net and rou file by providing weight args:

python src/main.py --config src/config.yaml --net path/to/net.xml --route path/to/rou.xml --path models/single-intersection --weights models/single-intersection

Alternatively, if you're running the rendering job on a cloud instance, you can simluate via xvfb with the following bash command:

xvfb-run -a -s "-screen 0 1400x900x24"     python main.py --config config.yaml --net single-intersection.net.xml --route single-intersection.rou.xml --path Outputs/