A Technique for Use of Gaussian Processes in Advanced Meta-Modeling

Current robust design methods rely heavily on meta-modeling techniques to reduce the total computational effort of probabilistic explorations to a combinatorially manageable size. Historically most of these meta-models were in the form of Response Surface Equations (RSE). Recently there has been interest in supplementing the RSE with techniques that better handle non-linear phenomena. One technique that has been identified is the Gaussian Process (GP). The GP has fewer initial assumptions when compared to the linear methods used by RSEs and, therefore, fewer limitations. The initial implementation and employment techniques proposed in current literature for use with the GP are barely modified versions of those used for RSEs. A better, more tailored technique needs to be developed to properly make use of the nature of the GP, and minimize the effect of some of its limitations. Such a technique would allow for rapid development of a reusable, computationally efficient and accurate GP. A new technique is presented here that includes potential revisions to the design of initial experiments, modification of training and validation techniques, and the addition of a “self healing” algorithm to improve the performance of the process during employment.