Machine Learning-Based High Fidelity Mesoscopic Modeling Tool for Traffic Network Optimization

Abstract

Traffic signal control of the future will adapt in real time to traffic, driving, and environmental conditions. The adaptations will require real-time optimization based on traffic monitoring. However, for optimization to work, it must be able to predict what traffic will do in response to changes in signal timing. Furthermore, it must do this prediction very efficiently, in a short time horizon, so that the optimization process can evaluate many alternative signal timing parameters. Instead of using a “one size fits all” legacy function for prediction of traffic flow, this project took steps to pioneer the use of machine learning techniques to learn how traffic behaves on particular segments of streets in response to traffic and driving conditions and used that learning to build a high-speed simulator to predict the arrival times of cars at intersections.