Improving Robustness to Model Misspecification in Bayesian Experimental Design

Alex Forster, Desi R. Ivanova, Tom Rainforth

Keywords: Bayesian experimental design, Model Misspecification, Robustness

Abstract: We propose a method to improve robustness to model misspecification in Bayesian experimental design (BED). Our approach introduces a flexible auxiliary model and jointly optimizes the expected information gain (EIG) in the original model parameters, the predictions of the auxiliary model, and a Bernoulli random variable indicating whether the original model is correct or misspecified. We show this balances learning about the original model, gathering data useful for general prediction, and assessing model fit. By leveraging the domain-specific knowledge embedded in the original model, we guide the design process while maintaining flexibility in the face of model misspecification. This is particularly important in adaptive design settings, where the original model informs early design decisions, but the auxiliary model enables adaptation when new data reveals model inaccuracies.

Submission Number: 34