definition of **s_max** and **s** ? about l1qr

[BayerSe]

Hi,

s_max is used to prematurely stop the algorithm when s reaches s_max. In fact, s is the regularization parameter.
Does this help?

[mahmoudalbardan]

thanks for the details.
However, I do not see why the regularization parameter s shows up in the prediction phase (argument of the predict function) not during training (in the fit function).
As I understand it, the regularization parameter is only used when training to estimate the coefficients, not for prediction.

Another question: if we want to access the final estimated coefficients, we can do mdl.beta[-1] ?

[BayerSe]

The algorithm of Li and Zhu (2008) fits the whole path, i.e. for all values of s. Therefore, it does not show up in the fit function - the only thing you can can control is some premature stopping via s_max. When it comes to prediction you have to make a choice - for which value of s do you want to get the prediction?

mdl.beta[-1] gives you the estimate of the coefficients for the largest value of s - which should be equivalent to the unpenalized QR estimator.

[mahmoudalbardan]

Thank you Bayer.
It is more clear right now.

[BayerSe]

Great!