Titolo della tesi: Advanced Optimization and Deep Learning for Enhanced Urban Traffic Management and Safety
This PhD thesis presents a comprehensive examination of fixed-point problems, traffic assignment models, and road traffic accident severity prediction, leveraging advanced computational and algorithmic approaches to address critical challenges in transportation and traffic safety. The work introduces the Trust Contraction algorithm as a novel solution to the fixed-point problem, demonstrating significant improvements over the Method of Successive Averages (MSA) in solving the Stochastic User Equilibrium (SUE) model for traffic assignment. Through theoretical justification, convergence analysis, and numerical experiments, the Trust Contraction algorithm is shown to offer a faster and more robust solution for traffic assignment under various conditions, including congestion levels, perception variance, and network sizes.
The thesis further explores a single-level joint formulation for Origin-Destination (OD) matrix estimation under SUE, incorporating traffic congestion and route choice with traffic measurements. Employing the Levenberg-Marquardt algorithm enhanced by Automatic Differentiation and the Generalized Minimal Residual method, this approach demonstrates efficiency and resilience across different network scenarios, emphasizing the importance of balancing information quantity with variable estimation in large-scale transportation networks.
Addressing the application of deep learning in traffic assignment, the thesis introduces a data-driven method using Graph Neural Networks, specifically the Message-Passing Neural Network model. This novel approach surpasses traditional and conventional deep learning techniques in predicting traffic distribution, showing adaptability to dynamic changes in traffic demand and supply.
In tackling road traffic accident severity prediction, the thesis presents an innovative analysis using a comprehensive dataset from Rome, employing machine learning models with a focus on the Extreme Gradient Boost (XGBoost) algorithm. The application of one-hot encoding, the Synthetic Minority Over-sampling Technique (SMOTE), and conformal prediction enhances model performance and interpretability, with SHapley Additive exPlanations (SHAP) identifying key factors influencing accident severity.
Collectively, this thesis contributes significant advancements in the fields of traffic assignment and safety, offering novel computational solutions and insights into traffic management, demand estimation, and accident severity prediction. Through a blend of theoretical development, algorithmic innovation, and practical application, the research provides a foundation for future advancements in transportation analysis and safety strategies.