LSSVMs have a regularization parameter

[bhvieira]

Just a minor comment, while in the models list it makes clear SVM have a regularization parameter C || cost to be tuned, the LSSVM also has one: tau. caret doesn't reflect that.

[topepo]

Ah, I overlooked that. Do you want to suggest a range of appropriate values?

[bhvieira]

@topepo Well, I have no theory to back it, 1E-2 is the default. tau 10^(-4:4) brings great variability in iris. See these results, the training error for the iris dataset.

y = c(0.01342282, 0.01342282, 0.006711409,0.01342282,0.02013423,0.02684564,0.03355705,0.04026846,0.1073826,0.3288591,0.3288591)
x = 10^(-5:5)

The region around 10^(-4:-2) is the optimum.

[topepo]

Once I went back and looked at the papers, I think that I will base the
grid off of the same ones for cost in the usual SVM models.

I've sent an email to the package maintainer. I have almost always gotten
an error using the linear kernel with these models and would like to know
if there is a fix.

I'll check in some code as soon as I'm done testing.

On Wed, Jul 27, 2016 at 3:03 PM, Bruno Hebling Vieira <
notifications@github.com> wrote:

@topepo https://github.com/topepo Well, I have no theory to back it,
1E-2 is the default. tau 10^(-4:4) brings great variability in iris. See
these results

y = c(0.01342282, 0.01342282, 0.006711409,0.01342282,0.02013423,0.02684564,0.03355705,0.04026846,0.1073826,0.3288591,0.3288591)
x = 10^(-5:5)

The region around 10^(-4:-2) is the optimum.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#415 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/AFdy498bEhYei63RVbrEIC5102HyP14yks5qZ6uNgaJpZM4IOmez
.

[topepo]

on CRAN