LLM-Enhanced MARL for Smarter Traffic Control

The rapid advancement of artificial intelligence has driven the application of Deep Reinforcement Learning in multi-agent systems, offering new opportunities for complex collaborative scenarios in transportation, energy, and communications. As urban traffic congestion intensifies, adaptive traffic signal control based on Multi-Agent Reinforcement Learning (MARL) has emerged as an effective approach to optimize traffic flow and reduce intersection delays. However, existing research largely focuses on single-intersection control, with limited exploration of coordinated multi-intersection scheduling. This study proposes a MARL-based method for coordinated traffic signal optimization, utilizing Multi-Agent Proximal Policy Optimization (MAPPO), which demonstrates superior stability and convergence over other algorithms. To address MARL's dimensionality challenges, we introduce two key innovations: (1) optimizing the MAPPO network architecture by integrating a Transformer module to enhance model expressiveness and accelerate convergence; and (2) incorporating Large Language Models (LLMs) to leverage their reasoning capabilities for improved multi-agent collaboration. Experimental results show significant improvements in traffic control efficiency, laying a solid foundation for intelligent transportation systems.