Inverse Problems 6: Picking regularization parameter (L-curve+Morozov Discrepancy)
Вставка
- Опубліковано 8 вер 2024
- L-curve = "Pick point of maximal curvature of plot of data misfit v. regularization penalty norm" -- intuition is that you are making a compromise between data misfit and penalty.
Morozov discrepancy = "Find beta such that the data misfit is O(noise level) AND NO FURTHER" -- if you go further, you are overfitting.
L-curve is less theoretically robust, but it does not require prior knowledge of the noise level.
Morozov is more theoretically robust, but it requires prior knowledge of the noise level.