Variable Selection For Computer Calibration Model

With the continuous updating and creation of science and technology,more and more computer models are applied to deal with real problems.In order to simulate the real system,the inaccurate computer model calibration is also deeply concerned by researchers.The variable selection method could select the variables that have significant influence on the output value of the computer,and use the variable selection method to calibrate the computer model,which can further improve the accuracy and interpretability of the computer model.In this thesis,we study the general computer calibration model firstly,and then propose the variable selection method under penalty likelihood to calibrate the inaccurate computer model.Three methods,namely Lasso,Adaptive Lasso and Elastic Net,are selected.It is also proved that the estimation of calibration parameters obtained by this method is consistent and satisfies asymptotic normality under certain conditions.To solve the problem of parameter identification of computer calibration model,an orthogonal computer calibration model is proposed in this paper.Finally,Lasso,Adaptive Lasso and Elastic Net were used to calibrate the general computer calibration model and the orthogonal computer calibration model respectively,and the simulation performance of the three methods was compared.Numerical simulation results and case analysis show that the orthogonal computer calibration model using Lasso,Adaptive Lasso and Elastic Net methods is superior to the general computer calibration model using the above three methods in accuracy and stability of simulation,thus solving the problem of parameter identification.It improves the interpretability of the model.