Neural network's simulation result OUTRANGE the target data.

I've trained neural network using LM backpropagation on matlab several times. When i test (simulate) the network using new input data, the test result OUTRANGE the target data. My target data, range between 0.1 to 0.3 but the simulation result sometimes show value greater than 0.3. Is it common? I only use 18 data pair since I couldn't find more data

Answer 1

[Greg Heath]

The range of your output data depends on the output activation function. The default NN for regression and curve-fitting is FITNET. By default, it uses the linear function PURELIN which does not constrain the output range.

The important thing is how much of the average biased (divide by N instead of N-1) target variance

    vart1 = mean(var(target',1))

can you model? This is quantified by the normalized mean square error

   error = target - output;

    NMSE = mse(error)/vart1

and the corresponding COEFFICIENT OF DETERMINATION ( AKA known as the RSQUARE) where

   Rsq = 1-NMSE

is a measure of just how much of vart1 is modelled.

 See Google and Wikipedia regarding RSQUARE and COEFFICIENT OF DETERMINATION for more 
insight on how to measure modelling performance.

Hope this helps,

Thank you for formally accepting my answer

Greg

P.S. In many of my earlier NEWSGROUP and ANSWERS postings, I have used the notation MSE00 instead of vart1.