Probabilistic modeling of heteroscedastic laboratory experiments using Gaussian process regression

Abstract

This paper proposes an extension to Gaussian process regression (GPR) for data sets composed of only a few replicated specimens and displaying a heteroscedastic behavior. Because there are several factors that are out of the control of experimenters, it is often impossible to reproduce identical specimens for a same experiment. Moreover, observations from laboratory experiments typically display a heteroscedastic interspecimen variability. Because experiments and specimen manufacturing are expensive, it is uncommon to have more than three specimens to build a model for the observed responses. The method proposed in this paper uses GPR to predict each tested specimen using a shared prior structure and models the global heteroscedastic behavior by combining observations using conjugate prior distributions. An application of the method to high-performance fiber-reinforced concrete experiments highlights fiber addition benefits for reducing water permeability caused by macrocracks.