I need a starting point for choosing "spread" when using newrb()

My data sets consist of 62 inputs and one output and I want to do function approximation. I understand that the optimum "spread" value is usually determined by trial and error. However, I was wondering if there is any way of approximating this value ( just to get a sense of its greatness )? My second question is regarding the minimum number of training samples required when using newrb. Is it just like the feedforward neural networks, the more the better?

Thank you for your support

Answer 1

[Greg Heath]

If you standardize inputs (zscore or mapstd) the unity default is a good starting place.

The best generalization performance comes from using as few hidden neurons as possible.

Search the neural net literature (e.g., comp.ai.neural-nets FAQ) using the terms

 overfitting

 overtraining

[Shadan]

thanks!