Traffic Signal Control
Project Overview
Traffic signal control has long been managed by algorithms and intelligent systems. This project aims to validate a machine learning method for traffic signal control, enhancing technologies used in traffic and transportation management.
Challenges
The primary goal of this project was to build and train a reinforcement learning model for traffic signal control. The main challenges included simulating real-world roads and infrastructure in software, and converting traffic conditions into a numerical reward function for training.
Methodology
- Simulation: Build a simulated environment to replicate real-world roads and traffic signals.
- Algorithm Design: Develop a prototype of the reinforcement learning algorithm.
- Training and Iteration: Train the model using the designed reward function and make necessary adjustments.
- Optimization: Select the best-performing model by fine-tuning the hyperparameters.
- Validation: Report the model performance within the relevant simulated environment.
- Documentation: Record all steps, detailing the algorithm, hyperparameters and parameters used in the algorithm.
Tools
The project primarily utilized traffic simulation software, Vissim and SUMO, along with OpenAI Gym and PyTorch in Python for model building and training. Git and GitHub were employed for version control and collaboration. Testing and deployment were supported by continuous integration/continuous deployment (CI/CD) tools and practices. Amazon Web Services (AWS) were deployed to train the reinforcement learning model in a highly parallelized simulation environment.
Results
The final outcome was a reinforcement learning model trained to control traffic signals. The results clearly indicated improved traffic conditions compared to traditional signal control programs. Traffic jams were reduced in the simulated environments, and vehicle flow was cleared more quickly than before.