Convergent Nonparametric Variable Screening and Metamodeling Methods
Jian Tu · 2004
For a CAE model whose true underlying function form is totally unknown, the nonparametric metamodeling approach is more flexible and is likely to produce more accurate nonlinear approximations of the CAE model’s multivariate relationship. Nevertheless, it still faces the so-called curse-of-dimensionality, according to which the required sample size may increase exponentially with the number of input variables. This paper presents a convergent variable screening strategy, in which a nonparametric univariate-based additive modeling approach is used with a uniform sample partitioning technique to ensure robust variable screening. After reducing the large input variable set to a moderate size, the resulting nonparametric multivariate metamodel can thus capture the key relationships in the CAE model.