Regularization weight factor (wfmin/wfinit) #591
Comments
|
the dynamic regularization solve is some tricky business and I think the "regularization continue" option in pest_hp might be to help with this issue. I wonder if you take some smaller steps on the way to phimlim if that might help? So maybe a larger eigthresh and/or larger lambda values so slow down the first couple of upgrades since, in many cases, that first upgrade can be pretty aggressive... |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
This is not so much a pyemu question and more a general PEST question, But I don't know anyone else to ask.
I have a highly parameterized groundwater model and I do parameter estimation in PEST++ with preferred homogeneity regularization. According to what I’ve read (Pest manual, roadmaps etc.) the best practice is, in short, to run the inversion with a low PHIMLIM/PHIMACCEPT first and then increase the two iteratively until parameters are realistic and the fit is ok.
After doing this I reached a PHIMLIM that I feel should be obtainable without unrealistic parameters, but PEST would still reduce the regularization weight factor and reach my target PHI in a few iterations while giving me weird parameter values.
So as an experiment I tried to keep my PHIMLIM/ACCEPT at the level I felt should be obtainable, but increased my starting weight factor (WFINIT) and minimum weight factor (WFMIN) to enforce a higher level of regularization.
This took PEST a lot more iterations and it tried to transgress the lower WFMIN limit, but it reached the acceptable PHI and the parameter values were okay.
So my question is what do I miss out on by doing it the “experimental way” rather than the best practice. Is it problematic and why?
Cheers Kim
The text was updated successfully, but these errors were encountered: