부산대학교 도서관

Adaptive traffic signal control (ATSC), a networked traffic signal control system with real-time coordination of traffic control signals across intersections, aims to address the aggravated traffic congestion in urban areas. Recent years have seen a growing body of literature that employs multiagent reinforcement learning (MARL) for ATSC, where each traffic controller located at one road intersection is an agent aiming to optimize an objective like minimizing traffic delay. The centralized RL control paradigm, however, incurs high communication costs and demands substantial data, making it challenging to implement in large-scale, real-world road networks. Accordingly, decentralized RL control is more desirable. On the other hand, data sharing and exchange could be challenging in real-time, which demands training a coordinated ATSC using individual datasets collected from each road intersection. In this paper, we apply federated reinforcement learning (FedRL) to ATSC for its benefits in reducing communication cost while maintaining collaborative control. By comparing FedRL to centralized RL and distributed RL through experiments conducted on real-world road networks, we demonstrate the efficiency and superior performance of the FedRL approach in addressing traffic congestion.