Robustness to model misspecification in Bayesian experimental design

Abstract

Scientific experiments can be costly to conduct due to various reasons. Therefore, researchers need to choose the experimental design which allows them to obtain the highest amount of information possible. As an information-theoretical framework based on Bayesian inference, Bayesian experimental design is utilised to choose the most optimal experimental design. However, there is a possibility of model misspecification which would lead the framework to produce suboptimal results. This thesis conducts a literature review to find the extent of research regarding the model misspecification problem in Bayesian experimental design. The results of the study show that there are two metrics to measure the degree and effect of model misspecification named expected generalized information gain and expected discriminatory information as well as the presence of a method of using hypothesised noise to increase robustness to model misspecification. The thesis also indicated that sequential Bayesian experimental design methods are more vulnerable to the effects of model misspecification.

The thesis concludes that there is a need for studies about the model misspecification problem in the Bayesian experimental design framework since most of the existing literature focuses on identifying and analysing the problem while there is only limited literature proposing a solution to the problem.