Optimizing Traffic Flow With Reinforcement Learning: A Study on Traffic Light Management
Amal Merbah, Jalel Ben‐Othman
Abstract
The non-adaptive management of traffic lights has proven inefficient for a number of drawbacks. They mainly impinge on CO2 emissions, fuel consumption, traffic waiting time, and heavy traffic. In this study, we propose a traffic signal control system that combines the accuracy of mathematical modeling with the real-time and adaptation features of deep learning (DL) by basing the DL configuration on a mathematical model of the interaction between the environment and the intersection as a Markov decision process (MDP) while taking structural and safety issues into consideration. As a resolution method, we suggest in this study a policy iteration (PI) method, which gives the best policy to follow so as to choose the action that determines the phase duration. These phases minimize the reward, which is the average waiting time (AWT) for all vehicles crossing the intersection. The PI has demonstrated greater efficiency compared to management systems based on fixed durations in various traffic situations. Instead of triggering the PI system for each new situation encountered and minimizing the processing time, the PI will act as a learning method for the DL program. We build a learning database by storing several situations represented by the variables: input flow, latest switching dates, output flows, traffic light states, and queue lengths, with their respective solutions returned by PI as the policy for selecting next switching dates. Due to this configuration, DL has been able to respond optimally and in real-time to different levels of throughput: low, medium, and high.