AdaptLight: Toward Cross-Space-Time Collaboration for Adaptive Traffic Signal Control

Recent multi-agent deep reinforcement learning (MADRL) approaches have shown notable benefits in traffic signal control. However, the spatial-temporal coupling, hysteresis, and heterogeneity of collaborative agents are usually ignored. States and actions among multiple intersections induce complex coupling and hysteresis in both space and time dimensions, while the actions also present spatial-temporal heterogeneity due to fluctuated traffic. These characteristics impose a critical impact on the efficiency and flexibility of coordinated control. In this paper, we propose AdaptLight, an MADRL-based model to achieve cross-space-time collaboration. It captures the interactions among spatial-temporal traffic components and exploits action repetition to adaptively adjust decision granularity for heterogeneous traffic. For the spatial-temporal coupling and hysteresis issue, AdaptLight first establishes a feature extraction network based on spatial-temporal graph Transformer. To tackle the spatial-temporal action heterogeneity problem, an action-repetition-enabled MADRL module is designed, which can decide asynchronous-cooperative actions spanning multiple timesteps. Experiments present that AdaptLight shows competitive performance on different datasets.