Scholarlink

This paper introduces a Multi-Agent Deep Recurrent Q-Network (MADRQN) designed for real-time traffic light control across multiple intersections. The approach aims to improve joint control efficiency while minimizing communication overhead among agents. By modeling traffic light management as a Markov Decision Process and treating each intersection’s controller as an agent, the MADRQN method dynamically clusters agents based on their locations and real-time observations using the Growing Neural Gas algorithm. Within each cluster, information sharing and centralized training are applied to enhance coordination. Simulation results on the Simulation of Urban MObility (SUMO) platform demonstrate that this approach reduces communication load, facilitating more effective and efficient information sharing and training. Consequently, MADRQN achieves shorter average vehicle waiting times compared to leading multi-agent deep reinforcement learning methods, significantly alleviating traffic congestion.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Agent Clustering and Information Sharing Underlying MADRQN for Traffic Light Cooperative Control

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Haoran Cheng,
  • Bo Wang,
  • Jie Liu,
  • Tongchun Du

摘要

This paper introduces a Multi-Agent Deep Recurrent Q-Network (MADRQN) designed for real-time traffic light control across multiple intersections. The approach aims to improve joint control efficiency while minimizing communication overhead among agents. By modeling traffic light management as a Markov Decision Process and treating each intersection’s controller as an agent, the MADRQN method dynamically clusters agents based on their locations and real-time observations using the Growing Neural Gas algorithm. Within each cluster, information sharing and centralized training are applied to enhance coordination. Simulation results on the Simulation of Urban MObility (SUMO) platform demonstrate that this approach reduces communication load, facilitating more effective and efficient information sharing and training. Consequently, MADRQN achieves shorter average vehicle waiting times compared to leading multi-agent deep reinforcement learning methods, significantly alleviating traffic congestion.