Statistically weighted maximin distance design

Abstract

In a computational experiment, a metamodel, which is an approximation model, is widely used to perform optimization efficiently. The accuracy of a metamodel significantly depends on the way of choosing sample points. This process is known as the design of experiment (DOE). An important property of DOE is space filling that is developed to obtain information evenly on the overall design domain. However, space filling may be ineffective in optimization because this property does not consider output information. The proposed novel sequential DOE places more sample points in the neighborhood of the interested region in terms of optimization. The proposed method employs the weighted distance concept that considers output information. The weighted distance is evaluated through proposed parameters that are obtained from the basic statistical distribution of output information, e.g., probability density or cumulative distribution function, while satisfying space filling.