Abstract

We introduce a new experimental design criterion for contexts where model selection and checking are the primary goals. Our work adapts the surface and posterior distribution exploration criterion proposed by (Joseph et al. 2015, 2018) to the model discrimination context. Our new criterion leads to a sequential design approach that trades off two tasks: distinguishing competing models of interest and maintaining space-filling properties, the latter being essential for model checking and prediction. We demonstrate through simulations that designs generated by our method perform well at distinguishing competing linear regression models and Gaussian process kernels, while also providing data to check the possibility that none of the models initially considered are adequate.

Keywords Model discrimination, experimental design, surface exploration.