In this fourth installment of the TMIP AI Webinar Series, we will explore AI-DCM models in more depth, using TB-ResNets as an example. AI-enhanced discrete choice models or AI-DCM combine traditional logit discrete choice models with various types of artificial neural networks. The core motivation is straightforward: pure logit models are interpretable and theoretically grounded in the Random Utility Maximization (RUM) framework, but they miss complex patterns and interactions that more flexible methods can find. Pure neural networks can discover those patterns, but they are opaque, behaviorally unconstrained, and difficult to audit or defend to stakeholders. Most AI-DCMs address this by combining the two: the logit provides a behavioral backbone; the neural network captures what the logit misses. The result is a family of models that are more accurate than either alone, while remaining substantially more interpretable than a pure data-driven approach.
TB-ResNets, introduced by Dr. Wang, are a simple ensemble of a logit model and deep neural network, structured so that the whole model can still be interpreted as a discrete choice model. He will introduce the framework, its motivation, application, and interpretation using mode choice as an example. Dr. Mishra will share his recent work with TB-ResNets and other AI-DCM frameworks in the context of activity scheduling.
Shenhao Wang, PhD, is an assistant professor and the director of the Urban AI Laboratory at the University of Florida. As an urban and computer scientist, he develops novel AI approaches to focus on three research themes: (1) resilient urban systems, (2) travel behavior and networks, and recently (3) generative AI for design and planning. Dr. Wang completed his interdisciplinary Ph.D. in Computer and Urban Science at Massachusetts Institute of Technology in 2020.
Sabya Mishra, PhD, is a Professor in the Department of Civil Engineering at the University of Memphis. His research has included various national and state transportation projects funded by organizations such as the Federal Highway Administration, and the Maryland, Michigan, and Tennessee Departments of Transportation. In his research he has used several techniques such as machine learning, artificial intelligence, and deep learning concepts to model user behavior.