[R] Some simple questions about neural networks

[Ken Spriggs]

Have you ever gotten any response from this post?  I have similar questions
regarding the AMORE package.  



Efferz wrote:
> 
> Hi,
> 
>  
> 
> I have some "simple" questions and annotations about neural networks:
> 
>  
> 
> 1) Which R-package (or which software) would you use to train and validate
> a
> multilayer (2 hidden layers) feed forward neural network. I think "AMORE"
> is
> the only one that can do this task in R.
> 
>  
> 
> 2) When using neural networks for time series prediction (macroeconomic &
> financial time series), how would you precede to avoid overfitting? Split
> the sample in two subsamples, train the NN in the first subsample and then
> test it on the validation set? What is a good split ratio 1/2, 2/3, 3/4?
> Are
> there procedureces which endogenize this step.
> 
>  
> 
> 3) How to select the parameters like the global.learnging.rate, the
> momentum.global, the activation functions of the hidden layer neurons, the
> training method, the n.shows and show.step numbers, the probability
> vector,... when setting up a neural net with "AMORE" or
> 
> the initial weights, decay,... when setting up a neural net with "NNET"?
> 
>  
> 
> Are these all "econometrican choice variables" and must be exogenously
> specified. My own experience shows that the results are far from robust
> and
> highly sensitive to an alternative parameter choice.  Even when using the
> same parameter setup re-training and re-validation delivers different
> results (unless you use the set.seed command). How to get results that are
> replicable? I think there is a great danger of getting spurious results
> when
> snooping the parameter space. Or are there any reasons why to use a decay
> of
> 0.1 instead 0.11 or a momentum of 0.4 instead of 0.5, or ...
> 
> Is it a good choice to use if possible the default values?  Therefore I am
> very skeptical if those new and  highly sophisticated non-linear methods
> (neural networks, svm, etc.) perform really better in time series
> prediction
> than the classical linear methods. Besides the problem which variables to
> use as predictors, how to choose the calibration window (rolling
> expanding,
> rolling fixed), one faces the additional choice of model parameters.
> 
>  
> 
> How do you think about it? Any ideas? Any experiences?
> 
>  
> 
> Best
> 
> Martin
> 
> 
> 	[[alternative HTML version deleted]]
> 
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
> http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
> 
> 

-- 
View this message in context: http://www.nabble.com/Some-simple-questions-about-neural-networks-tp12898957p14814193.html
Sent from the R help mailing list archive at Nabble.com.