Deep Reinforcement Learning based Intelligent Traffic Control

Abstract

The development of Intelligent Traffic Signal Control (ITSC) systems is crucial for enhancing traffic flow and mitigating congestion, which is a widespread problem in urban areas globally. Presently, RADAR or inductive loop-based intelligent systems are used in metropolises of developed countries, but the large investment and infrastructure requirements rule out their widespread application. This paper explores a nascent Deep Reinforcement Learning (DRL) approach to the Traffic Signal Control (TSC) problem, as opposed to classical optimization or rule-based approaches of the past. To address the challenges that limit past RL approaches, the study leverages the Deep Deterministic Policy Gradient (DDPG) algorithm to optimize traffic light control policies. The proposed DRL approach shows intelligent behavior and reduces the average delay time and congestion when compared to the traditional RL, past DRL, and fixed-time signal approaches. A comparative analysis of the reward functions is also presented, which reveals insights into the variance of performance.