A typical objective of training is, given data ( DESIGN + NONDESIGN ) that can be considered to come from the same probability distribution, use
1. ONLY design (training + validation) data
2. As few hidden neurons as possible
to reduce the MSE of NONDESIGN (test) data below a specified threshold. For example
3. mse(ttst-ytst) < 0.01*mean(var(ttst',1))
This is usually attempted by trying to achieve a specified goal of MSEtrn proportional to mean(var(ttrn',1)) before MSEval reaches a minimum.
If you desire better fitting at extrema, use the errorWeight option of mse
help mse
doc mse
Hope this helps.
Thank you for formally accepting my answer
Greg
