|Date: Friday, March 29, 2019
Location: 1084 East Hall (3:00 PM to 4:00 PM)
Title: Finding the most useful data via simulation-based Bayesian experimental design
Abstract: Experimental data play a crucial role in developing and refining models of physical systems. However, some experiments produce more useful data than others, and well-chosen experiments can provide substantial resource savings. Optimal experimental design (OED) thus systematically quantifies and maximizes the value of experiments. We describe general mathematical frameworks and algorithmic approaches of OED, targeting to handle nonlinear and computationally-intensive models. The formalism employs Bayesian statistics and an information-theoretic objective, which rigorously defines the conditions under which batch experiments (experiments planned simultaneously) and sequential experiments (forward-looking designs with data feedback) are truly optimal. Finding these optimal designs quickly becomes intractable, and we develop practical numerical methods for OED by employing and advancing computational techniques on several fronts, including stochastic optimization, polynomial chaos surrogate modeling, approximate dynamic programming, and deep policy gradient. The overall procedure is demonstrated on the design of combustion experiments, ice sheet borehole selection, and sensor placement for contaminant source inversion.
Speaker: Xun Huan
Institution: University of Michigan
Event Organizer: AIM seminar organizers