Edge-Assisted Multi-Agent Deep Reinforcement Learning for Emergency and Emission-Aware Traffic Signal Control

Abstract

Urban traffic signal control must reduce congestion, prioritize emergency vehicles, and limit environmental impact under dynamic traffic conditions. Fixed-time control cannot adapt effectively to such variations. Existing learning-based methods have shown promising results, but most of them do not jointly address emergency awareness, emission reduction, and decentralized edge-assisted operation. This paper proposes an edge-assisted multi-agent Double Dueling Deep Q-Network (D3QN) framework for intelligent traffic signal control. Each signalized intersection is controlled by a local edge-based agent using local traffic observations, which enables decentralized realtime operation. The proposed reward function jointly captures queue length, waiting time, emergency response, and local $\mathbf{C O}_{2}$ emission, and guides the agent toward balanced signal control. The proposed method is implemented in a SUMO-TraCI environment and compared with fixed-phase, alternating, and existing benchmark policies. Results show that, relative to the alternating baseline, the proposed method reduces the average queue length by 21.2%, the average waiting time by 19.7%, and the local $\mathbf{C O}_{2}$ emission by 5.9%, while maintaining comparable emergency-clearance performance.