Advanced optimization and deep learning for enhanced urban traffic management and safety

This PhD thesis addresses the critical need for advanced traffic management and safety solutions in urban environments through the development and application of optimization algorithms and deep learning models. It presents a series of research findings that collectively aim to enhance the efficiency, reliability, and adaptability of traffic systems. By focusing on specific challenges within the realm of traffic management and safety, this work introduces novel computational techniques and methodologies that offer significant improvements over traditional models. The thesis begins with the development of the ”Trust Contraction” algorithm, a new method for solving the Stochastic User Equilibrium (SUE) problem more efficiently than the widely-used Method of Successive Averages (MSA). This algorithm accelerates conver- gence and demonstrates robust performance across different traffic scenarios, establishing a new standard for traffic assignment optimization. Building on this foundation, the thesis then explores a groundbreaking approach to Origin- Destination (OD) matrix estimation. It proposes a single-level formulation that simplifies the traditionally complex bi-level optimization challenge, employing Automatic Differen- tiation and the Levenberg-Marquardt algorithm for enhanced computational efficiency. This method proves to be effective across various network configurations, offering a scal- able solution for integrating traffic congestion and route choice models into traffic planning and simulation. In the realm of deep learning, the research introduces an innovative application of Graph Neural Networks (GNNs) to predict traffic flow patterns. This model moves beyond static equilibrium assumptions, providing a dynamic framework for traffic assignment that adapts to real-time data. The GNN approach showcases the potential to revolution- ize traffic management systems by enabling more responsive and accurate adjustments to traffic conditions. The thesis also addresses the critical issue of road traffic accident severity prediction using a machine learning-enhanced model. By analyzing a comprehensive dataset with factors ranging from weather conditions to road characteristics, the study applies conformal pre- diction to assess accident severity with high accuracy. This method, supported by SHapley Additive exPlanations (SHAP) for model interpretability, identifies key determinants of accident severity, guiding the development of targeted safety measures. Overall, the thesis contributes a suite of advanced computational tools for traffic man- agement and safety, marking a significant step forward in the quest for more efficient, safe, and adaptive urban traffic systems. These innovations not only provide practical solutions to current challenges but also open new avenues for future research in traffic system optimization and safety enhancement.