2025 Theses Doctoral

A Novel Game-Theoretic Framework for the Autonomous Mobility Ecosystem

Chen, Xu

This dissertation aims to answer the research question of how to optimally design decision-making processes for autonomous vehicles (AVs), focusing on their dynamic velocity control and route choice within transportation networks. Unlike traditional traffic models that abstract away individual-level control, this work explicitly models multi-agent interactions among AVs, treating each as an intelligent, decentralized agent in a shared environment.

We formulate AV coordination as a large-population differential game, which converges in the many-agent limit to a class of mean field games (MFGs). These MFGs are designed to capture decentralized decision-making under mutual interactions and network constraints. Two core formulations are developed: (i) Spatiotemporal MFGs (ST-MFGs), which model continuous-time velocity control and agent distribution across space and time, and (ii) Graph-based Dual MFGs (G-dMFGs), which address simultaneous driving and routing decisions across transportation networks. Solving these MFGs involves challenging coupled forward-backward PDEs that are often computationally intractable in large-scale, high-dimensional systems. To overcome this, we propose a suite of AI-enhanced, learning-efficient solution methods.

These include: a hybrid Reinforcement Learning and Physics-Informed Deep Learning (RL-PIDL) framework, a Pure-PIDL method, and a Physics-Informed Graph Neural Operator (PIGNO) tailored for solving networked MFGs. These algorithms integrate physical structure and domain knowledge into deep learning architectures, leading to faster convergence, improved scalability, better generalization, and greater data efficiency compared to conventional solvers. This dissertation contributes a unified, interpretable, and scalable framework that merges mean field game theory with physics-informed machine learning to support decentralized, efficient coordination among AVs. It offers new algorithmic tools and theoretical insights for applying AI to solve large-scale multi-agent control problems in intelligent transportation systems.