End-to-End Learning of User Equilibrium: Expressivity, Generalization, and Optimization
该研究提出一个端到端学习框架，通过神经网络参数化交通网络中的供需组件并嵌入隐层来强制用户均衡条件，直接从多日交通状态观测中学习均衡状态，并设计了基于自动微分的梯度下降算法进行高效训练。
This paper establishes an end-to-end learning framework for constructing transportation network equilibrium models. The proposed framework directly learns supply and demand components as well as equilibrium states from multiday traffic state observations. Specifically, it parameterizes unknown model components with neural networks and embeds them in an implicit layer to enforce user equilibrium conditions. By minimizing the differences between the predicted and observed traffic states, parameters for supply and demand components are simultaneously estimated. We demonstrate that the end-to-end framework is expressive: when parameterized with sufficiently large neural networks, it can replicate any unique, differentiable equilibrium state that solves a well-posed variational inequality. Moreover, it can generalize to new, unseen data when trained with sufficient observations. For efficient training, we design an autodifferentiation-based gradient descent algorithm that handles link- and path-based user equilibrium constraints and ensures local convergence. The proposed framework is demonstrated using three synthesized data sets. History: This paper has been accepted for the Transportation Science Special Issue on Machine Learning Methods for Urban Passenger Mobility. Funding: This work described in this paper was partly supported by research grants from the USDOT Center for Connected and Automated Transportation and the National Science Foundation, Division of Civil, Mechanical and Manufacturing Innovation [Grant CMMI-2233057].