Traffic signal control based on weighted mean field multi-agent reinforcement learning

Abstract

Cooperative optimization of intelligent traffic light control is an effective approach to alleviating traffic congestion and improving the efficiency of transportation networks. We propose WMFLight (Weighted Mean Field multi-agent reinforcement learning-based traffic Light control method), a method that combines dynamic clustering and multi-agent mean field reinforcement learning. This method uses K-Means++ to implement dynamic clustering, and employs a weighted mean field mechanism to accurately characterize the attenuation characteristics of urban traffic flow in spatial topology and global coordination requirements, thereby adapting to the dynamic changes of traffic flow. Extensive experimental results demonstrate that this method has significant advantages over baseline methods in terms of convergence performance, reducing average travel time, and adapting to the dynamic changes of traffic flow.