Proposition of an enhanced deep reinforcement learning algorithm for managing traffic.

Abstract

Due to the rapid growth of the population as well as the quality of life, reliance and the use of roads is increasing drastically which leads to an inevitable increase in traffic. As a result, many traffic control systems have been in operation to regulate traffic but still need improvements to be more suitable especially in our country. Indeed, these systems are designed to work on predefined patterns which do not always correspond to the real and particular cases of our road traffic. In order to tackle this challenge we propose in this work an upgrade approach of an already existing traffic monitoring model based on Deep QLearning to traffic lights control with an agent simulator using the framework SUMO. Our main objective is to minimize the average waiting time of the vehicles at the intersection. To this end we propose many improvements to the algorithm, by adding some important layers to the convolutional neural network and adapting the values of some crucial hyper parameters. The obtained simulation results have shown that our approach performs better than the existing algorithm in terms of mean waiting time.