How can I improve the performance prediction of a net in extrapolation

Training data are in [minTrain maxTrain], test data are in [minTest maxTest] such as 30% (for instance) of the data are outside the range [minTrain maxTrain]. It is known that neural networks do not work well in extrapolation. How can I improve the network predictions in this case just to have acceptable results. Do you have any idea?

Answer 1

[Greg Heath]

Use whatever info you have to fabricate additional data.

[Platon]

Can you explain more your proposal?

[Greg Heath]

Not much more to say. If theory and/or non-training data gives you some idea of data characteristics beyond those limits, use tat info to fabricate training data with those characteristics.