R help - Apr 2008 - Variables selection in Neural Networks

Hi folks,

I want to apply a neural network to a data set to classify the observations
in the different classes from a concrete response variable. The idea is to
prove different models from network modifying the number of neurons of the
hidden layer to control overfitting. But, to select the best model how I can
choose the relevant variables? How I can eliminate those that are not
significant for the model of neural networks? How I can do this in R?

dataset.nn=nnet(response.variable~., dataset, subset = training, size=1,
decay=0.001, linout=F, skip=T, maxit=200, Hess=T)

What I am doing is to vary size between 0 and 1 since with a single layer it
can learn any type of function or continuous relation between a group of
input and output variables. But this only would give two different models.
The ideal would be to be reducing nonsignifictive variables. How I can prove
other different models?

Regards.

_Fede_
-- 
View this message in context:
http://www.nabble.com/Variables-selection-in-Neural-Networks-tp16911299p16911299.html
Sent from the R help mailing list archive at Nabble.com.

Hi Fede,

You would have to eliminate the variables less correlated with the response
variable. And for the explanatory variables to choose those that are very
correlated to each other. I don't know if exists some function of R that
does this by you.

Mathew Barber


_Fede_ wrote:
> 
> Hi folks,
> 
> I want to apply a neural network to a data set to classify the
> observations in the different classes from a concrete response variable.
> The idea is to prove different models from network modifying the number of
> neurons of the hidden layer to control overfitting. But, to select the
> best model how I can choose the relevant variables? How I can eliminate
> those that are not significant for the model of neural networks? How I can
> do this in R? I do this:
> 
> dataset.nn=nnet(response.variable~., dataset, subset = training, size=1,
> decay=0.001, linout=F, skip=T, maxit=200, Hess=T)
> 
> What I am doing is to vary size between 0 and 1 since with a single layer
> it can learn any type of function or continuous relation between a group
> of input and output variables. But this only would give me two different
> models. The ideal would be to reduce the model eliminating nonsignifictive
> variables. How I can prove other different models?
> 
> Regards.
> 
> _Fede_
> 
-- 
View this message in context:
http://www.nabble.com/Variables-selection-in-Neural-Networks-tp16911299p16921302.html
Sent from the R help mailing list archive at Nabble.com.