Deep Reinforcement Learning based Intelligent Traffic Control System

Abstract

Metropolitan cities are witnessing enormous surge in everyday traffic resulting in delays, congestions and longer commute times. This rise is the consequence of the ever-increasing population and the corresponding demand of vehicles. However, present traffic control systems are not reciprocating this increasing demand. Upgrades in the form of real-time intelligent systems are necessary. Our solution to this problem is the formulation of a deep reinforcement learning traffic control system that actively monitors the traffic environment and makes precise predictions. These predictions in the form of traffic light phases provide for the most efficient flow of traffic through an intersection. Using deep Q-learning networks, we present an enhanced solution to the problem that reduces the average waiting time of vehicles in the environment by about 60%. We compare and evaluate multiple novel neural network architectures and demonstrate their results in our custom environment. We also account for variety of vehicles on road and their nuances, to accurately depict a realistic environment.