Sequential projection maximin distance sampling method

Abstract

Design optimization for engineering problems often requires severe computer simulations. Thus, to perform a design optimization efficiently, surrogate models replacing the time-consuming simulator by using the adequate number of computer experiments, i.e., design and analysis of computer experiments (DACE) have been developed. Our goal in this paper is to propose a sequential design of experiments to construct a global surrogate model. The proposed method employs the priority of variables defined from non-linearity, contribution ratio or global sensitivity. The priority gives a chance to have better projective property to more important variable, because relatively more important variable significantly influences on the accuracy of surrogate model. Consequently this causes a decrease in the error of surrogate model and a reduction of the total number of sample points. The proposed method is compared with sequential maximin distance design and optimal Latin hypercube design with two examples.