It IS IMPORTANT to choose a correct set of values for the
hyperparameters C and sigma in order to get good generalization.
The best way to do this is to use k-fold cross validation.
There are other measures too, but they don't always work well.
You may want to look at a comparison we did of various measures:
http://guppy.mpe.nus.edu.sg/~mpessk/comparison.shtml
CJ Lin's LIBSVM now has in it hyperparameter tuning. You may want to
try it.
Keerthi
On Fri, 17 Jan 2003, Zhang Ruijie wrote:
> hi, everyone :
>
> Now when i train my SVM, i select sigma & c randomly or
> just by experience.
>
> Do you know any paper that can tell me how to choose these
> two parameter automatically
>
>
>
> Thanks
>
|