Multi-agent Reinforcement Learning for Real-time Traffic Control in 6G-Enabled VANET Environments
Unknown authors
Abstract
The growth of urban mobility and networked vehicles has expedited the need for smart traffic management systems that can operate optimally in the next-generation wireless environment.This paper introduces an innovative Multi-Agent Reinforcement Learning (MARL) system for real-time traffic signal control in 6G-enabled Vehicular Ad Hoc Networks (VANETs).In contrast to conventional approaches that optimize solely traffic-related metrics like delay or queue length, the presented system adopts traffic-oriented and communication-aware metrics such as average waiting time, congestion index, packet delivery ratio, and bandwidth utilization under a common reward framework.Dynamic scalarization allows MARL agents to adaptively weigh these objectives based on real-time traffic and network conditions.The proposed framework was validated using four datasets: the Vehicular Network State Dataset (Kaggle), Secure VANET Vehicle Dataset, VANET Real-Time Route Optimization Dataset, and 5G-VANET MEC Dataset.All traffic signals are designed as independent MARL agents, learning signal phases based on local observations and fast inter-agent communication via 6G links.The innovation is the simultaneous optimization of vehicle flow and network-level performance for end-to-end real-time coordination and scalability in highly dense urban VANET environments.Compared to standard fixed-time control algorithms, MA2C, DQN-based controllers on the same dataset, the suggested framework provides an 32.6% decrease in mean waiting time, a noteworthy reduction in congestion, a 21.4% increase in packet delivery ratio, and a 27.8% increase in bandwidth efficiency.The results confirm the framework's robustness and generalizability across diverse VANET scenarios.