Traffic Signal Control with Extended Perception for Enhanced Robustness andResilience Using Graph Attention and Temporal Encoders

Authors:

Abstract
Deep Reinforcement Learning has demonstrated significant potential in optimizing Traffic Signal Control; however, existing Multi-Agent approaches often struggle with the non-stationarity inherent to abnormal traffic conditions. Most models, trained in nominal simulation environments, lack the resilience to withstand unforeseen disturbances such as sensor failures, road blockages, or extreme weather. To address this gap, this paper proposes a robust Spatio-Temporal Graph Reinforcement Learning framework explicitly designed for structural stability. Our architecture facilitates intelligent inter-agent collaboration by selectively weighting relevant neighbors via a Graph Attention Network (GATv2), while a Transformer-based Temporal Interaction Encoder captures long-term historical trends. This dual mechanism enables agents to distinguish between transient noise and persistent failure patterns, facilitating proactive rather than reactive control. Furthermore, we introduce a rigorous stress-testing benchmark comprising stochastic perturbation scenarios; ranging from high-frequency random blockages to critical arterial collapses, to quantify system resilience. Extensive experiments on a stochastic grid network demonstrate that our approach not only achieves superior throughput under high-demand nominal conditions but also exhibits exceptional resilience, recovering stable traffic flow significantly faster than state-of-the-art baselines following severe structural disruptions.

This publication has 14 references indexed in Scilit: