A fast optimal latin hypercube design for Gaussian process regression modeling

Abstract

In engineering applications, Gaussian process (GP) regression method is a new statistical optimization approach, to which more and more attention is paid. It does not need pre-assuming a specified model and just requires a small amount of initial training samples. Based on the design of experiment (DOE), determining a reasonable statistical sample space is an important part for training the GP surrogate model. In this paper, a novel intelligent method of DOE, the translational propagation algorithm, is employed to obtain optimal Latin hypercube designs (TPLHDs). It also proved that TPLHDs' performance is superior to other LHDs' optimization techniques in low to medium dimensions. Using this method, the best settings of the process parameters are determined to train GP surrogate model in the injection process. A automobile door handle is taken as an example, and experimental results show that the proposed TPLHD performs much better than the normal LHD in the quality of fitting GP surrogate model, so taking TPLHDs instead of LHDs' optimization technique for training GP model is practical and promising.