Application and development of reinforcement learning in traffic signal control

Abstract

With the acceleration of urbanization, the traditional timing traffic signal control strategy is difficult to deal with the complex and variable nonlinear traffic flow. Reinforcement learning (RL) shows great potential in alleviating traffic congestion with its strong adaptive perception and optimization ability. However, reinforcement learning still faces many challenges in practical applications. This paper systematically reviews the latest research progress of reinforcement learning based traffic signal control (RL-TSC), focusing on control efficiency optimization, multi-objective constraints and overall architecture adjustment. Research shows that RL-TSC still has some problems, such as insufficient robustness to real-world noise, a lack of characterization of non-motor vehicle behavior, and a computing bottleneck of edge devices. The significance of this paper is to clearly point out the limitations and technical gaps of the current application in the rl-tsc field, and propose that the future research direction should be towards the interpretable RL, LLM hybrid architecture and global active scheduling, so as to provide theoretical guidance and forward-looking reference for the construction of a smart urban transportation system that takes into account safety, efficiency and sustainable development.