How are the automatic values of hyper-parameters in Matlab Regression Learner determined?

Using Matlab regression learner one can choose the auto option for the values of the various hyper-parameters such as epsilon and Kernel scale mode in SVM's. In this case is stated that if auto is chosen the app uses a heuristic procedure to select the kernel scale. Also the same applies in the Gaussian Processes. When Kernel scale mode is set to Auto, it is stated that the app uses a heuristic procedure to select the initial kernel parameters. -What is the heuristic procedure followed? -Are the values given optimised? -If they are why the "tips" encourage the user to give values manualy?

[Bernhard Suhm]

Are you just trying to understand what's going on, or do you have evidence it's not working as designed?

[Georgios Etsias]

It is important to know if the selected parameters are the optimal ones or I should do an optimization of my own!

Answer 1

[Ilya]

If you type

edit classreg.learning.svmutils.optimalKernelScale

in your MATLAB session and hit Return, the editor will bring up the code for that heuristic procedure.

You won't know if these parameters are optimal or not without doing optimization. These are based on a guess. The guess is often good but it can fail from time to time.