New Zealand Statistical Association 2024 Conference

Martin Hazelton

University of Otago

Model-based priors for network tomography

Network tomography is a challenging type of statistical linear inverse problem. A common example arises in transport engineering, where the goal is to estimate volumes of origin-destination traffic flow based on traffic counts observed at various sites over the network. In that context, the difficulties for statistical inference are exacerbated by the existence of multiple plausible routes connecting most origins and destinations of travel. The observed data provide limited information about the route choice probabilities, and so the availability of an informative prior is critical. In this talk I describe how such a prior can be constructed using classical route choice models founded on game theory. Such models are computationally expensive, and so I also discuss the use of cheap emulators.