Sequentially weighted uniform designs

Abstract

Uniform designs seek to distribute design points uniformly in the experimental domain. Some discrepancies have been developed to measure the uniformity by treating all factors equally. It is reasonable when there exists no prior information about the system or when the potential model is completely unclear. However, in the situation of sequential designs, experimental information, such as the importance of each factor, would be obtained from previous stage experiments. With this fact, the weighted -discrepancy is more suitable than the original discrepancy for choosing follow-up designs. In this paper, the sequentially weighted uniform design is proposed, which is obtained by minimizing the weighted -discrepancy. The weights, indicating the relative importance of each factor, are estimated through a Bayesian hierarchical Gaussian process method based on serial experimental data. Results from several classic computer simulator examples, as well as a real application in circuit design, demonstrate that the performance of our new method surpasses that of its counterparts.