Neural Network Test Data

Hi everyone; I have the this neural Network made by

the training Function is a Bayesian regularization backpropagation. I don't understand why the test value are not good even if the training data are almost perfect

Answer 1

[Greg Heath]

It looks like a case of overtraining an overfit net. If you have O-dimensional outputs and use the default Ntrn ~ 0.7*N then you have

Ntrneq = 0.7*N*O = 0.63*N training equations

Whereas the number of unknown weights is

 Nw = (I+1)*H1+(H1+1)*H2+(H2+1)*H3 +(H3+1)*O
    = 3*25++26*15+16*15+16*9 ~849

N >> 849/.63 ~ 1348 So how much data do you have?

Using Bayesian regularization should help.

However insufficient details.

Greg

[Greg Heath]

I'm old fashioned and guided by the following principle

1. Make life as pleasant as possible:

 a. Don't worry if you can't model more than 99% 
    of the average target variance

 b. Use as few hidden layers as possible

 c. Use as few hidden nodes as possible.

2. Search both NEWSREADER and ANSWERS with

greg Hmin Hmax

Hope this helps.

Greg