A Spatiotemporal Multi-Agent Framework Based on CoevoMARL for Multi-Intersection Traffic Signal Control

Abstract

Managing traffic signals across multiple adjacent intersections remains difficult because vehicle flow changes rapidly and congestion spreads across both space and time. In this paper, we adapt key ideas from CoevoMARL to build a lightweight, spatiotemporal multi-agent reinforcement learning setup for adaptive signal control. We treat each intersection as an independent agent that tracks local queue pressure, while coordinating decisions across the network using graph attention layers for spatial neighbors and recurrent units for temporal trends. Testing our framework on a 16-intersection (4 × 4) CityFlow network over a one-hour period showed clear performance gains. While standard fixed-time control resulted in an average travel time of 1137.80 s, our approach dropped that to 675.44 s, a 40.64% reduction with top-performing episodes reaching 466.36 s. These findings highlight that targeted spatial and temporal coordination can boost urban traffic flow without incurring heavy computational overhead.