Metamodel Validation with Deterministic Computer Experiments
Yao Lin, Janet K. Allen, Farrokh Mistree · 9th AIAA/ISSMO Symposium on Multidisciplinary Analysis and Optimization · 2002
Problem: Computer-based simulation programs are widely used in design to help analyze the system performance given a set of inputs. Metamodels are usually developed to replace complex, time-consuming simulation programs and used in the early stages of design to facilitate fast and accurate analysis. Methods for testing the validity of design and analysis of physical experiments are not ideal for deterministic computer experiments since there is no random error. In this paper, we intend to answer the following questions: Is cross validation an appropriate method for model validation in deterministic computer experiments? And if the answer is no, what techniques should we use to validate the metamodel and illustrate the prediction uncertainty?