I am curious - how does the SANN - ANS select which networks to retain. If I train 200 models and retain 20, there doesn't seem to a obvious characteristic that identifies the 'best performing networks'. Is 'best' selected on the correlation coefficient of the train, test or validate subset. Or is it on the SOS error of the train, test or validate subset. Or is it based on behind the scenes, learning outcomes from training the entire 200 models. I have included the summary of the 20 retained models, I can't see which characteristic informs the selection of 'best'.
Networks with the lowest error for regression are retained, and highest classification rate for classification will be retained. This is done on Test and/or Validation set depending on the options you have chosen. You can also reference the help in Statistica as well. It is extensive, with detail on all the options for the algorithms, not just Neural Networks but all algorithms. It's a great resource.