The performance of hidden neurons

Hi all,

after several runs of the neural network to select the optimum number of hidden neurons, I found that the performance changes when the HN are within a loop rather than when it was taken one by one. why ?

thank you for your help

Answer 1

[Greg Heath]

I do not understand your problem. However if you want to see examples of searching for the optimal (e.g., smallest) value of hidden nodes that will yield a specified performance, see some of my zillion examples.

Search both the NEWSGROUP and ANSWERS using

 greg Hmin dH Hmax

Hope this helps.

THank you for formally accepting my answer

Greg

[Lilya]

Thank Dr. Heath for your answer. the problem is the performance changed when the number of HN changed manually. In spite of used within a loop! it doesn't gave the same result.

[Greg Heath]

I don't understand. Please post code.

Greg

[Greg Heath]

Oh! … If you mean that you are getting different weight values from different random initializations, that is to be expected. There are an uncountable number of weight combinations and permutations that will yield acceptably small answers.

Greg

Answer 2

[Greg Heath]

I think I misinterpreted the question. Now I think you mean when I increase the number of hidden nodes from 4 to 5 why do I start the H=5 from scratch rather than just add a new node to either all or just the better H=4 solutions?

ANSWER: It is easier and faster.

That was good enough for me.

If you want to make a convincing TIME AND ERROR comparison of the two techniques, please use MATLAB datasets and alert me when you post it.

Greg