In the figure below the trained data (shown in green ) is not covering fully to the target data by Artificial neural network technique. I used feed forward ANN technique. Is it the limitation of the ANN or anything else

Answer 1

[Greg Heath]

You have to decide how much of the target variance you want to model. For a regression net (e.g, FITNET) I try to model at least 99%.

Plots are very useful. However, the actual calculation of the normalized mean-square-error is the proof.

I have posted many, many FITNET examples in the NEWSGROUP and ANSWERS. You may want to check the most recent posts first.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Greg Heath]

What are

   1. [ I N ] = size(input) 
   2. [ O N ] = size(output)
   3. H       = No. of hidden nodes

All I can do is estimate from eyeballing the plot to obtain

   1. O = I = 1
   2. N ~ 1620, Ntrn ~ 0.7*N ~1143
   3. Nlocmax ~ 16
   4. H >= 2*Nlocmax ~ 32
   5. Nweights =(I+1)*H+(H+1)*O = O+(I+O+1)*H >= 97

I suspect that increasing H could lead to a better fit. However, is it really worthwhile? What is

 NMSE = mse(target-output)/var(target',1)  ?