"Number of hidden neurons ANN"

chaosbringer

Hi,
i have a question... of course
The doc for the neural netowrk operators says: the (hidden) layer size will be set to (number of attributes + number of classes) / 2 + 1.

2 question:

1. Why not using 2*num_attrs+1 hidden neurons? With that number is can be guaranteed to approximate any function, as you might know
2. What is the number of classes, if i am making a regression? 1?

Thank you very much

land

Hi,

well actually beside from asking you could answer some of the questions you already have been solved

1. If you can approximate any function you might really simple overfit. I suggest reading the nice paper "Overfitting S&P 500" at http://www.shookrun.com/documents/stupidmining.pdf

2. Yes, is 1 for regression.

Actually you might add this to the operator documentation, if you would be so kind?

Greetings,
Sebastian