Transforming Equilibrium Analysis Paradigm for Modeling Transportation Networks with Intelligent Traveling Agents

点击翻译按钮获取中文摘要

This project will transform transportation network equilibrium modeling paradigm that metropolitan planning organizations use to plan emerging connected and automated mobility systems for our nation. Considering that future traveling agents (connected drivers or automated vehicles) will possess strong learning and computation capability, and their travel decisions can be algorithmic, strategic and adaptive, this project will first investigate the day-to-day evolution of network traffic dynamics with these intelligent traveling agents, and examine whether the notion of equilibrium remains relevant for modeling and planning future mobility systems. Leveraging massive empirical data made available by connectivity, this project will then develop an end-to-end learning framework that directly learns relevant modeling components and the equilibrium state from empirical data. The planned modeling paradigm, if successful, has a great potential for widespread market adoption, and will save time and resources for metropolitan planning organizations to build and maintain their planning models. It can potentially help them better plan and manage their transportation networks to reduce traffic congestion and vehicle emissions, without requiring much new investment on expanding the existing infrastructure. This project will consist of two thrusts. The first thrust develops dynamical systems by explicitly modeling day-to-day travel choices of intelligent traveling agents and then examines the convergence and stability properties of the dynamical systems. By demonstrating that Wardropian user equilibrium can still emerge in the day-to-day evolution of network traffic dynamics, this project intends to establish a behavioral basis for the equilibrium modeling paradigm. The second thrust aims to integrate implicit deep learning with network equilibrium analysis to develop an end-to-end framework that directly learns the behaviors of traveling agents, the equilibrium state, and other modeling components, if needed, from empirical data. This research, if successful, makes fundamental contributions to advance transportation network science. First, it innovates the methodology of modeling day-to-day traffic dynamics by explicitly capturing the decision-making process of day-to-day choices of travelling agents. Second, the end-to-end learning and optimization framework represents a paradigm shift for modeling and planning transportation networks. The framework melds the data-decisions pipeline by integrating learning and decision/optimization into a single end-to-end system. Lastly, the research enriches the literature of game theory by offering novel application and further development of mean-field game theory, and presenting a new way of integrating machine learning with game theory.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.