A Novel Decentralized Optimization Technique with Explainable AI for Responsive Traffic and Pedestrian Management

Actuated and pre-timed traffic signal controllers have been beneficial to the improvement of traffic in cities and urban environments across the world. While these methods have been effective, recent developments in the field of machine learning demonstrate that these traffic signal controllers can be further improved through the proper implementation of this growing field of study. However, it is equally important for practical application that when multiple traffic signal controllers in a local environment exist, that they are capable of working together in a way that is beneficial to everyone. Not only is it important for the framework of each intersection to be capable of cooperating, it is also ideal that the algorithm itself be capable of accounting for parameters beyond vehicles, such as pedestrians who may rely on the traffic signal controllers to safely cross intersections.

This research includes the application of a lightweight queue-prediction algorithm that can work in tandem with a machine learning algorithm to better build predictive models, which showed minimal error that slowly increases as queue sizes increase. A modular model-based reinforcement learning algorithm was created that was tested with an accurate recreation of historical data provided through a series of traffic cameras placed at intersections along Shallowford Road in Chattanooga, TN. The algorithm was proven to demonstrate a stable performance of reducing the cumulative delay of vehicles waiting while balancing this objective with avoiding any neglect of idle vehicles on less congested roadways. This work finally concludes with the creation and testing of a decentralized traffic signal controller algorithm that leverages model-based predictions to improve learning rates and allow for intersections to collectively work towards optimal solutions while minimizing communication between one another and spreading the computational burden among all intersections via a decentralized framework.