Goodness of Fit Criteria

The Table and Chart views display Goodness of Fit metrics for each output parameter. The following criteria are calculated for the points taken into account in the construction of the response surface.


Note:  Criteria marked as advanced options are shown in the GOF table only if you've enabled Show Advanced Options on the Design Exploration page in the Options dialog box. For more information, see Design Exploration Options.


Coefficient of Determination (R2)

The percent of the variation of the output parameter that can be explained by the response surface regression equation. That is, the Coefficient of Determination is the ratio of the explained variation to the total variation. The best value is 1.

The points used to create the response surface are likely to contain variation for each output parameter (unless all the output values are the same, which results in a flat response surface). This variation is illustrated by the response surface that is generated. If the response surface were to pass directly through each point (which is the case for the Kriging response surface), the coefficient of determination would be 1, meaning that all variation is explained.

Mathematically represented as:

where:

Adjusted Coefficient of Determination (Adjusted R2)

Advanced option. The Adjusted Coefficient of Determination takes the sample size into consideration when computing the Coefficient of Determination. The best value is 1.

Usually this is more reliable than the usual coefficient of determination when the number of samples is small (<30). Only available for standard response surfaces.

Mathematically represented as:

where:

Maximum Relative Residual

Advanced option. The maximum distance (relatively speaking) out of all of the generated points from the calculated response surface to each generated point. The best value is 0%. In general, the closer the value is to 0%, the better quality of the response surface.

However, in some situations, you can have a larger value and still have a good response surface. For example, this can be true when the mean of the output values is close to zero. See the following formula.

Mathematically represented as:

where:

Root Mean Square Error

This is the square root of the average square of the residuals at the DOE points for regression methods. The best value is 0. In general, the closer the value is to 0, the better quality of the response surface.

Mathematically represented as:

where:

Relative Root Mean Square Error

Advanced option. This is the square root of the average square of the residuals scaled by the actual output values at the points for regression methods. The best value is 0%. In general, the closer the value is to 0%, the better quality of the response surface.

However, in some situations, you can have a larger value and still have a good response surface. For example, this can be true when part of the output values are close to zero. You can obtain 100% of relative error if the observed value = 1e-10 and the predicted value = 1e-8. However, if the range of output values is 1, this error becomes negligible.

Mathematically represented as:

where:

Relative Maximum Absolute Error

This is the absolute maximum residual value relative to the standard deviation of the actual output data, modified by the number of samples. The best value is 0%. In general, the closer the value is to 0%, the better quality of the response surface.

Note that this value and the Relative Average Absolute Error value correspond to the Maximum Error and Average Absolute Error scaled by the standard deviation. For example, the Relative Root Mean Square Error becomes negligible if both of these values are small.

Mathematically represented as:

where:

Relative Average Absolute Error

This is the average of the residuals relative to the standard deviation of the actual outputs. This is useful when the number of samples is low (<30). The best value is 0%. In general, the closer the value is to 0%, the better quality of the response surface.

Note that this value and the Relative Maximum Absolute Error value correspond to the Maximum Error and Average Absolute Error scaled by the standard deviation. For example, the Relative Root Mean Square Error becomes negligible if both of these values are small.

Mathematically represented as:

where:

= value of the output parameter at the sampling point
= value of the output parameter at the sampling point
= value of the regression model at the sampling point
is the arithmetic mean of the values
is the standard deviation of the values
= number of sampling points
= number of polynomial terms for a quadratic response surface (not counting the constant term)


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