Uncertainty Quantification for Unobserved Variables in Dynamical Systems and Optimal Experimental Design
Abstract
We consider a problem of optimal experimental design in dynamical systems, where we aim to find measurements that reduce uncertainty in model predictions. For a dynamical system, the set of parameters that fit the data well can be very broad or very narrow, but the size of this set does not necessarily translate into uncertainty over predictions. Thus, this needs to be determined another way. We solve an optimization problem that searches for a pair of models that each provide a good fit for the observed data, yet have maximally different predictions. This is called "prediction deviation." Prediction deviation can determine the possible impact that an additional measurement would take to reduce uncertainty. Thus, we can use prediction deviation to estimate which experiment to perform (which measurement to take) in order to maximally reduce uncertainty. We use this technique study optimal experimental design in a model of interferon-alpha inhibition of HIV infection. We will also prove a theoretical result showing that prediction deviation provides bounds on the trajectories of the underlying true model. The collaborators on this project will be Ben Letham, Portia A. Letham, and Edward P. Browne.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 12, 2017
- Source ID
- W911NF1510155
Entities
People
- Cynthia Rudin
Organizations
- Army Contracting Command
- Massachusetts Institute of Technology
- United States Army