Closed pckroon closed 5 years ago
Also, if this works as intended/promised, we can consider simplifying Fit by making it always use trust-constr. Lastly, it may be worthwhile to also generate the Hessian analytically.
In light of #195, I think it is best that I first complete and merge that one, and then I can experiment with doing a rebase of this PR on that new master.
As a general question, a lot of the methods now inherit from BoundedMinimizer
, even though they do not support bounds according to the scipy docs. Why is that?
Waiting for #195 to be merged and rebasing on that works for me.
I was under the impression that all scipy minimizers could deal with bounds. But apparently not. Note that this does not lead to failing tests...
Merged. One thing this PR made very obvious though, is that we need to take another look at the minimizer structure. I'll open a new issue where we can discuss this.
Completely agree. Please copy some of the discussion/ideas we've had on that here to that issue. Should save us some work.
Oh, and thanks for the help in polishing this :)
You're welcome, implementing this was a good move.
This implements the trust-constr method from scipy. They advertise it as the most robust/flexible/powerfull minimizer in their arsenal. It's kind of tricky to make it play ball, but I think this should work in most cases.