Systematic structural discrepancy assessment for computer models.

Abstract

Model or structural discrepancy is an essential component in the analysis of computer simulators, representing the differences between the outputs of the simulator and the real-world system that the simulator seeks to represent. This discrepancy can arise from various sources such as simplifications of the model science in the simulator, choices made in our particular implementation of that science, and epistemic uncertainties such as the absence of features or science that we did not know to include or have yet to discover. In this paper, we define and distinguish two types of discrepancy: internal discrepancy that can be assessed by experiments on the simulator itself; and external discrepancy which lies outside the scope of such experiments. We present a tractable methodology and workflow for the assessment of structural discrepancy on the basis of collections of experiments applied to the computer model and illustrate our approach in the context of a simple biological model.This article is part of the theme issue 'Uncertainty quantification for healthcare and biological systems (Part 2)' .