Colloquium

Title:            Robust Bayesian Optimal Experimental Design under Model Misspecification

Abstract:      Bayesian optimal experimental design (BOED) is increasingly used to improve uncertainty quantification by optimizing the way scientists and engineers gather data. However, as with all methods, understanding the impact of assumptions and model discrepancy is critical. In fact, this is particularly important for BOED because it guides the way data is gathered potentially causing BOED not to sample data that could falsify assumptions or sampling data whose errors are poorly quantified due to low-fidelity models.

In this talk we discuss a new information criterion, Expected Generalized Information Gain (EGIG), for BOED problems that incorporate model discrepancy. This criterion is intended to augment traditional BOED based on Expected Information Gain by defining a trade off in robustness vs performance of the experimental design. We will discuss both the theoretical aspects of this new approach to OED and scalable algorithms incorporating it into BOED.