Urban traffic congestion presents a complex challenge driven by intricate spatial dependencies and non-stationary temporal dynamics. While Multi-Agent Deep Reinforcement Learning has shown promise for Traffic Signal Control, existing approaches often struggle with partial observability and fail to coordinate effectively across large-scale, heterogeneous road networks. In this paper, we propose HydraLight (HYbrid Deep Reinforcement Learning Architecture for Traffic Lights), a novel spatio-temporal framework that integrates Graph Attention Networks and Temporal Transformers. To overcome the localized myopia of standard graph methods, HydraLight introduces a Global Pooling Context module that broadcasts macroscopic, citywide traffic summaries, enabling agents to proactively mitigate systemic gridlock. Furthermore, to facilitate robust multi-scenario training, we introduce a Unified Prioritized Experience Replay (Unified PER) module that normalizes Temporal-Difference errors, preventing task dominance across diverse topologies. Extensive experiments on the RESCO benchmark across five synthetic and real-world networks demonstrate that HydraLight consistently outperforms state-of-the-art baselines (including X-Light and CoSLight).Byreducing traffic congestion, travel delays, and idle waiting times, the proposed framework also contributes to more sustainable urban mobility through improved traffic flow efficiency, lower fuel consumption, and reduced vehicular carbon emissions. Notably, the proposed architecture excels in structurally irregular environments, achieving up to 13.07% reduction in average travel time on complex arterial networks and consistently improving queue stability and waiting-time minimization across both synthetic and real-world RESCO benchmarks compared to state-of-the-art baselines.
Dabbagh et al. (Fri,) studied this question.