Multi-Objective Deep Reinforcement Learning for Variable Speed Limit Control

This study proposes a novel approach to address the multi-objective challenge of traffic control on congested freeways using Deep Reinforcement Learning (DRL). The approach involves developing a Learning-based Variable Speed Limit (VSL) agent specifically designed to optimize traffic efficiency while enhancing road safety. This agent employs a multi-objective reward function, striking a delicate balance between safety and mobility by optimizing speed limits to minimize collision risks and maximize traffic flow simultaneously. To illustrate the architecture of the Framework, a Deep Q-Networks (DQN) topology is presented. Through a comprehensive simulation study on a congested section of the A1 freeway in Morocco, our multi-objective DRL-VSL Framework showcased significant improvements, with mean speed increasing by 2.55% and Time to Collision values demonstrating a 57.33% enhancement. These results highlight the Framework's ability to simultaneously improve traffic flow and safety in real-world scenarios, contributing to the advancement of intelligent traffic control systems through DRL.