Adaptive Joint Control of Intersection Traffic Signals and Variable Lanes Using Multi-Agent Learning

Abstract

To effectively manage varying traffic flows at urban intersections during peak and off-peak hours, especially under conditions of unbalanced directional demand, we propose a learning-based coordination method for traffic signal control and variable-direction lane control (LCSL) to alleviate traffic congestion. The framework integrates a variable-direction lane control module and a traffic signal control module, leveraging mutual interaction and real-time information sharing to enable dynamic, coordinated decision-making. Additionally, we design an adaptive reward function based on lane balancing and traffic demand to enhance the adaptive coordination between agents. The use of a prioritized experience replay (Pr) mechanism further enhances the efficiency of experience utilization, accelerates algorithm convergence, and ensures the adaptive stability of the agents across varying traffic conditions. The experimental findings indicate that the LCSL method effectively decreases the average delay by 33.5% and the queue length by 48.2%, compared to the current state-of-the-art techniques, exhibiting higher stability and efficiency and improving intersection throughput.