Improving predictive quality of Kriging metamodel by variogram adaptation

Application of interpolation/approximation techniques (metamodels, for brevity) is commonly adopted in numerical optimization, typically to reduce the overall execution time of the optimization process. A limited number of trial solution are computed, cov- ering the design variable space: those trial points are then used for the determination of an estimate of the objective function in any desired location of the design space. The behaviour of the prediction of the objective function in between two trial points depends on the structure of the adopted metamodel, and there is no possibility, in principle, to determine a priori if one method is preferable to another. Nevertheless, some metamodels require the adjustment of a set of tuning parameters, and this operation is critical for the prevision qualities of the metamodel. In this paper, some base parameters of the kernel of the kriging metamodel are tuned in order to improve the overall quality of the prediction.