Search All of the Math Forum:
Views expressed in these public forums are not endorsed by
Drexel University or The Math Forum.
|
|
|
|
Re: Neural Network performance normalization
Posted:
Feb 8, 2013 9:40 AM
|
|
"EanX" wrote in message <kf2cjg$41g$1@newscl01ah.mathworks.com>...
> Thanks Greg,
> but, as far I know (I'm using matlab 2011b), Matlab automatically takes care of standardizing inputs
>
> net.inputs{1}.processFcns = {'removeconstantrows','mapminmax'};
Yes. However, its unfortunate they use mapminmax. For statistical reasons (e.g., outliers, mse normalization, ...) mapstd is superior. However, to make things less complicated just use zscore or mapstd BEFORE calling any NN functions.
> so, I suppose, correct me if I'm wrong, it's not necessary to implement what you suggest.
No it is not necessary.
HOWEVER, you have a problem. I gave you my preferred solution to your problem.
Why did you think I capitalized ALWAYS ???
> Below, an extract from my matlab code:
>
> inputs = input';
> targets = target';
[ I N ] = size(input) % = ?
[ O N ] = size(target) % = ?
minmax(input) % = ?
minmax(target) % =?
mean(var(target')) % = ?
Why did you complicate your code by specifying properties that were already defaults??
trainParam.goal is the only property you need to specify.
Please see the recent post on ANN_Error Goal.
Also, rather than use feedforwardnet, use the more specialized fitnet for regression and
patternnet for classification.
Greg
|
|
|
|
