ANN cannot extrapolate any data, which method can do?

i tried to model an ANN model to predict the behavior of concrete due to sulfate attack but the model cannot predict in out of inputs range.

for instance, i trained the network to predict the concrete behavior at different ages like 10, 100, 500, 1000, and 2000 days. the model can predict the propertie at any time between 10 and 2000 days, but it has a large error at predicting the properties at 20000 days.

i knew that the ANN cannot extrapolate any results from training data, how can i extrapolate some results for the previous mentioned example. which method can i use?

Answer 1

[Shashank Prasanna]

ANN is non-parametric and model free, if your training sample small and doesn't really capture the empirical relationship really well then there is only so much you can expect from it. Maybe the relationship changed erratically between 2000 and 20000, there is no way ANN or any other method you use will know that.

That said, you can always use splines or fit a custom model, if you know a parametric equation, in order to extrapolate.

There are some examples for splines. http://www.mathworks.com/help/matlab/ref/spline.html

But unless you see the data (or show us) there would be no universal method we can suggest that would work perfectly.