Dynamic driver behavior modeling with GMM for sustainable traffic signal control in deep reinforcement learning

Abstract

Adaptive traffic signal control is crucial for improving urban mobility and supporting environmental sustainability. Driver behavior heterogeneity has a significant impact on intersection performance, but its effective integration into reinforcement learning models remains challenging. This study proposes HARLight, a style-aware control framework that explicitly incorporates driver behavior heterogeneity to enhance signal efficiency. The framework employs a Gaussian Mixture Model (GMM) trained on real-world trajectory data to identify three representative driving styles: calm, neutral, and aggressive. The resulting driving style composition is incorporated into a Dueling Double Deep Q-Network (D3QN) state representation, enabling the controller to account for behavioral variations in signal phase decisions. Evaluation using the Simulation of Urban MObility (SUMO) platform shows that HARLight outperforms conventional engineering heuristics and standard reinforcement learning approaches. Under typical traffic conditions, HARLight achieves a 21.27% reduction in average intersection delay while also improving traffic progression and environmental performance. Under high traffic demand, the framework maintains effective control performance, while sensitivity analysis further demonstrates its robustness across varying driving style compositions.