Explainable RL Enhances Adaptive Traffic Signal Control

Adaptive traffic signal control, while benefiting from Reinforcement Learning (RL), faces challenges due to the black-box nature of deep RL models, hindering acceptance and trust. This research proposes a novel, explainable entity-centric RL framework designed for transparency and safety in traffic management. Instead of processing traffic states as monolithic vectors, the architecture breaks down real-time intersection observations into distinct lane entities and phase configurations, preserving the intersection's structural topology. A dual-stage attention network dynamically extracts relational dependencies and inter-lane conflicts, generating a real-time affinity matrix. This matrix visually and analytically quantifies the influence of signal phases on traffic volumes and queues, providing full interpretability. To guarantee operational reliability, a deterministic action-masking interface is integrated into the Proximal Policy Optimization pipeline, explicitly preventing invalid phase transitions and ensuring compliance with safety constraints. The framework outperforms state-of-the-art baselines in delay minimization and aligns with established traffic engineering principles.