Real-Time Bottleneck Identification and Graded Variable Speed Limit Control Framework for Mixed Traffic Flow on Highways Based on Deep Reinforcement Learning

Abstract

Effective management is essential for maintaining the smooth operation of highways. The graded variable speed limit (GVSL) method enhances traditional speed limit models by dividing the highway into multiple segments and applying different speed limits to each segment differently. However, there are two main issues: human drivers often choose not to comply with speed limits displayed on gantries, resulting in low compliance rates. Additionally, traditional models lack the flexibility to respond promptly to changing traffic conditions, potentially delaying necessary control measures. This paper proposes a graded speed limit control framework for mixed traffic environments consisting of connected and autonomous vehicles (CAVs) and human-driven vehicles (HDVs). The framework automates the identification of bottlenecks using fundamental diagram analysis and controls the behavior of CAVs through agents. This approach indirectly influences the behavior of HDVs to achieve overall traffic management objectives. The agents are modeled using deep reinforcement learning (DRL), incorporating graded levels and safe control intervals. In experiments conducted on the Taiwan road network, our RL-based GVSL method demonstrates a 3.4% increase in vehicle throughput and a 47.9% reduction in potential collision risks compared to traditional threshold-based control methods. Furthermore, the RL-based GVSL method adapts to various CAV penetration rates, leveraging the controllability of CAVs to achieve better control outcomes as penetration rates increase.