Dr. Fan visits RPI

Dr. He invites Dr. Yueyue (YoYo) Fan from UC Davis to give a talk on “Physics-Informed Data Analytics Approaches Using Constrained Optimization - Exploiting Domain Knowledge with Hard Information in a Transportation Network.” Thank you for sharing your research with us!

Her talk’s abstract:

Civil infrastructure as a system often faces challenges and complexity brought by interactions between spatially- and functionally- distributed components. Recognizing and incorporating these physical interactions in data driven approaches present challenges but also unique research opportunities for domain experts. In this talk, I will use transportation networks as examples to discuss how constrained optimization, by providing a flexible modeling framework for integrating domain knowledge, statistics, and data-driven approaches, could help address some fundamental data challenges that frequently arise in transportation applications. The first example shows how stochastic programming (SP) can be used to provide a statistically consistent and efficient estimate of global variables (network-level travel demand) that are not directly measurable based on partial local measurements (link-level traffic flows). In this example, domain knowledge reflecting network physics is modeled explicitly as constraints, and data samples are treated in some sense as uncertain scenarios in a SP framework. The second example shows how domain knowledge regarding the usage of data may be directly incorporated in data representation learning. In this example, objectives of the downstream application may be included in the design of the loss function in the learning process. The results demonstrate the importance of usage-aware representation learning for networked data.