Kriging Model Variable Selection Under Computer Experiments

With the development of technology,the problems to be solved in the real world become more and more complex.Computer experiment is becoming more and more popular as the substitute and assistant of physical experiment.Kriging model is an important model in computer experiment and could be applied to various fields.Variable selection is one of the common issues in statistical modeling.Whether it is to simplify the model or improve the prediction accuracy of the model,variable selection plays an important role in those aspects.The variable selection of Kriging model is concerned by researchers.In this dissertation,Lasso,Adaptive Lasso,Elastic Net are chosen to select variables of universal Kriging model.However,the issue of parameter identification in Kriging model can not be used to select significant variables more accurately.In this dissertation,the identification issue of Kriging model can be effectively solved by orthogonalization of universal Kriging model.Furthermore,a variable selection method based on Fiducial inference is also proposed.Finally,Lasso,Adaptive Lasso,Elastic Net and Fiducial are used to select variables for universal Kriging model and orthogonal Kriging model respectively.The performance of several variable selection methods in variable selection is compared.The simulation results show that the orthogonal Kriging model is superior to the universal Kriging model in variable identification rate and prediction accuracy.It is proved that the orthogonal Kriging model can effectively solve the problem of parameter identification.In addition,we find that Fiducial method has a great advantage in prediction accuracy and stability by comparing several variable selection methods.