Re: How many hidden neurons???

[EMAIL PROTECTED] (Jose M. Castro) wrote in message 
news:<[EMAIL PROTECTED]>...
> I built a recursive training process that calculates the total error
> for all the points to predict (simple sum of square of errors) and
> store the weights in a text file, then go for a new training (allways
> using the same training pairs), if this give less error then replace
> the weights file, if not then go for a new training. 

1. I hope you are using an independent validation set to 
   calculate the error.
2. If not, do you add a weight penalty term to SSE?
3. Are initial weights and biases the only difference
   between these runs?

> At training 654
> the program stops for insuficient memory, (I think I should clean
> memory for each training but don't know how to do this in visual
> basic). I've got better results than those obtained with the non
> linear regresion of the Excel's solver (half the error of this).

I hope Excel's statistical routines are better than they used to be.
Again, beware of comparing design data error rates.

> Now it is time to add a new recursive loop to change the number of
> hidden nodes and run the above process for every network arquitecture,
> graph the results and see what we got.
> Thanks for the feedback

What benchmark data sets are you using?

Good Luck.

Greg