2013/2/15 Gael Varoquaux <[email protected]>: > On Fri, Feb 15, 2013 at 03:03:58PM +0100, Olivier Grisel wrote: >> > In my opinion, I am not sure that you want an interpolator that >> > interpolate the model parameters for specific alphas. You may want to >> > compute the RMSE at the knots of the path, because I think that this is >> > where they will be the minimum that you are looking for. > >> BTW is there any theoretical (geometric) guarantee that this is the case? > > Between knots, the coefficients are affine functions (theoretically > proven and well-known). Thus the squared residuals are affine functions > too. Thus they can be minimum only at one end of the affine functions, > i.e. a knot.
Alright for minimizing the RMSE on the training set but on the test set the error might be minimized in between kinks as on the on real distribution, the kinks location might be slightly off the positions found on the training set or am I making a mistake? -- Olivier http://twitter.com/ogrisel - http://github.com/ogrisel ------------------------------------------------------------------------------ Free Next-Gen Firewall Hardware Offer Buy your Sophos next-gen firewall before the end March 2013 and get the hardware for free! Learn more. http://p.sf.net/sfu/sophos-d2d-feb _______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
