For each filtering output parameter, DesignXplorer builds an internal meta-model (polynomial response surface) based on the sample points generated during the correlation update.
A first response surface is built based on major input parameter returned by the filtering by using the correlation values.
Once a first response surface has been built, DesignXplorer tests the contribution of each input parameter in the prediction of the output parameter, and refines the response surface by removing the insignificant input parameters and by adding new significant input parameters.
The following notation is used in equations:
the number of sample points
the R-squared of the response surface
based on current major input parameters
the R-squared of the response surface
based on current major input parameters without the
major input parameter
the R-squared of the response surface
based on current major input parameters and the
minor input parameter
the number of terms used in the regression of the response surface
the number of terms used in the regression of the response surface
the number of terms used in the regression of the response surface
DesignXplorer computes the Fisher Test to test the significance of major input parameter and the gain in prediction:
Under the null hypothesis that response surface does not provide a significantly better R-squared than the response
surface
,
has an
distribution, with
degrees of freedom.
The null hypothesis is rejected if the calculated from the data is greater than the critical value of the
distribution for the false-rejection probability equals to 0.05. This
means that if the p-value of the Fisher Test is greater than 0.05, the major input parameter
becomes a minor input parameter.
DesignXplorer unselects all major input parameters when .
DesignXplorer computes the Fisher Test to test the significance of major input parameter and the gain in prediction:
Under the null hypothesis that response surface does not provide a significantly better R-squared than the response
surface
,
has an
distribution, with
degrees of freedom.
DesignXplorer promotes a minor input parameter as a major input parameter when (1%) and when the p-value of the Fisher Test is less than 0.05
(5%).
For each output parameter, DesignXplorer computes the contribution of each input parameter (
or
).
For a given output, to compute the relevance of contribution, DesignXplorer scales
contribution by the best known
contribution for this output, and transforms this value with the following
function:
