eMARLIN+: Addressing Partial Observability to Promote Traffic Signal Coordination by Leveraging Historical Information

In Adaptive Traffic Signal Control (ATSC) systems, real-time responsiveness relies on sensor data for signal timing adjustments. However, limitations in sensor capabilities result in an incomplete representation of the true system state. Hence, practical controllers can only access restricted dynamical features within specific detection areas that lead to partial observability, where identical observations may correspond to different system dynamics, hindering optimal decision-making. To address these challenges, we explore the existence and sources of partial observability in ATSC, formulating it within the framework of Markov decision processes. The global ATSC problem is factorized and decoupled to reveal structural properties in underlying system dynamics. This enhanced understanding reveals the dominant information that should be considered by decentralized controllers and guides the derivation of eMARLIN+. Experimental validation on synthetic and real-world scenarios demonstrates eMARLIN+'s effectiveness in enhancing agent-level coordination and surpassing strong baselines in minimizing travel delay. Additional diagnostic analysis of our learned controller further validates the effectiveness of our information-sharing scheme.