Compare trainlm with early stopping and trainbr generalization

Hi, I'm goingo to develop my neural network with both the standard trainlm algorithm (with early stopping at 6 validation fails) and with trainbr: I wish to compare the two trials, but since in trainbr the validation set is not present, how to compare them? Thank you

Answer 1

[Greg Heath]

There is no reason for TRAINBR to exclude validation subset stopping! There is published proof that using both is superior for many difficult problems. I think the references are given in COMP.AI.NEURAL-NETS. If not, I will find and post.

 1. CONTACT MATLAB and COMPLAIN that the validation subset option is not allowed in TRAINBR.
 2. Try to see if MATLAB will allow the use of TRAIN with both the msereg and valstop options. 
 3. Also remember that you can use any of the train options directly. For example

    [net tr y e ] = trainlm(net,x,t);

 4. Finally, do not forget that a net cannot be overtrained if it is not overfit. Therefore,
    another approach is to use the minimum number of hidden nodes that yields the desired 
    training goal. This is best determined by trial and error in a for loop. For examples use  
    the search words 

    greg Ntrials

Hope this helps.

Thank you for formally accepting my answer

Greg

[Greg Heath]

Sorry, cannot find the reference. However, it is recent and very believable.

I don't think I made myself clear above. In the reference the comparison

is between the 3 choices: 1. valstop 2. regularization 3. both

Both is better for small training sets..

Greg