Non-Parametric Regression

Non-parametric regression (NPR) provides improved response quality and is initialized with one of the available DOE types. NPR algorithm is implemented in DesignXplorer as a metamodeling technique prescribed for predictably high nonlinear behavior of the outputs with respect to the inputs.

NPR belongs to a general class of Support Vector Method (SVM) type techniques. These are data classification methods that use hyperplanes to separate data groups. The regression method works similarly. The main difference is that the hyperplane is used to categorize a subset of the input sample vectors that are deemed sufficient to represent the output in question. This subset is called the "support vector" set.

The internal parameters of the response surface are fixed to constant values and are not optimized. The values are determined from a series of benchmark tests and strike a compromise between the response surface accuracy and computational speed. For a large family of problems, the current settings provide good results. However, for some problem types (like ones dominated by flat surfaces or lower order polynomials), some oscillations might be noticed between the DOE points.

To circumvent this, you can use a larger number of DOE points or, depending on the fitness landscape of the problem, use one of the several optimal space filling DOEs provided. In general, it is suggested that the problems first be fitted with a quadratic response surface, and the NPR fitting adopted only when the Goodness of Fit metrics from the quadratic response surface model is unsatisfactory. This ensures that the NPR is only used for problems where low order polynomials do not dominate.


Release 18.2 - © ANSYS, Inc. All rights reserved.