Applies to: NeuralTools 5.x and Up

Neural Net Predictions Outside Training Range

If a neural net receives input values outside the range of the training data, it is effectively being asked to extrapolate. How well it performs depends on the net type:

• GRN/PN (Probabilistic) nets:

  • Designed primarily for interpolation within the training range.

  • Excellent at predicting within familiar data patterns.

  • Extrapolation is generally unreliable, because these nets focus on statistical interpolation rather than generalizing beyond the training data.

• MLF (Multiple-Layer Feedforward) nets:

  • Can detect general patterns in the data.

  • Better capability to extrapolate outside the training range, though accuracy cannot be guaranteed.

Practical advice:

• Out-of-range predictions may require custom checks to warn the user.

• You can stress-test extrapolation by splitting the dataset so the testing set contains out-of-range values.

• Using the NeuralTools XDK, you can automate multiple tests, similar to the Testing Sensitivity analysis in v6.

 

Attachments:
Extrapolation_NeuralTools.xlsx