Coordination in Traffic Signal Control: A Mechanism‐Oriented Survey of Multi‐Agent Learning Approaches

ABSTRACT Coordination across multiple intersections remains a fundamental challenge in traffic signal control (TSC), where locally optimized decisions often lead to network‐level congestion, spillback, and unstable traffic dynamics. Learning‐based approaches, particularly multi‐agent learning (MAL), have been increasingly adopted to address this challenge. However, existing surveys mainly categorize methods by algorithms or training paradigms, while the structural nature of coordination and its representation in learning systems are insufficiently examined. This survey adopts a mechanism‐oriented perspective and interprets coordination as a structural coupling problem arising from interdependent traffic flows and delayed spatial interactions in urban networks. We analyze how traffic‐level coordination requirements are translated into learning abstractions and instantiated through specific mechanisms, including reward coupling, information sharing, policy structure design, and temporal planning. To complement the taxonomy, we synthesize clearly sourced public benchmark results from standardized TSC evaluation to illustrate mechanism‐level empirical differences. We further examine structural failure modes by linking learning‐level design limitations to observable traffic phenomena such as non‐stationarity, credit assignment distortion, and spillback propagation. Finally, we discuss open challenges toward interpretable, scalable, empirically reproducible, and deployable coordination strategies and outline directions that extend beyond conventional reinforcement learning formulations. This survey aims to clarify the capabilities and limitations of learning‐based coordination in TSC and to provide a structural foundation for future research on reliable and scalable traffic control systems.