Developing Deep Learning Networks for Dynamic Traffic Light Control

Overview

Cary James John Cangelosi Theodore Taylor The average Los Angeles commuter spends upwards of 20 work days per year in traffic. Though the rise of electric vehicles has reduced greenhouse gas emissions, “a green traffic jam is still a traffic jam”. Congestion is caused by a number of theories such as overcrowding, spontaneous generation, interaction with pedestrians, and road work. However, another large factor is inadequate traffic light management. Adaptive traffic light control in which light timings are biased based on the number of cars at each node at an intersection have been tested and shown to improve commute speeds by 10% in a 9x16 block section of midtown NYC. Many of these centralized systems use traditional algorithmic control. Recent studies using deep learning improved simulated wait times up to 25%. In this research, a real intersection was chosen to model in Python with observed traffic data. This simulation models vehicles on a grid as an environment to evolve neural network agents that control traffic light timings, whose goal is to minimize wait times. Neuroevolution was used to produce a neural network which reduced wait times 12% and increased throughput 2% on average as compared to the traditional algorithms. The implementation of this model has the possibility to reduce costs as it is open source and can run on an SBC, which can accept inputs from detectors such as cameras and can control the lights itself. Combining these controllers across a city can allow the software to find solutions to minimize waiting and increase efficiency.