Federated Reinforcement Learning for Autonomous Vehicle Coordination in Urban Traffic Networks

Urban traffic management represents a high-stakes multi-agent coordination problem where autonomous vehicles must collaborate to optimize network-wide throughput while respecting privacy constraints that prohibit sharing raw sensor data. We propose FedRL-Traffic, a federated reinforcement learning framework that enables AVs to collectively optimize traffic flow through decentralized policy gradient methods and periodic model aggregation. Our approach introduces a communication-efficient gradient compression scheme and handles non-stationary reward distributions arising from dynamic traffic patterns. Experiments in SUMO simulation across 3 urban network topologies demonstrate 34% reduction in average travel time and 28% decrease in fuel consumption compared to non-cooperative baselines, with convergence achieved in 40% fewer communication rounds than vanilla federated RL.