Release?
pkofod opened this issue · 3 comments
I'm rewriting Optim, and part of that is making trust region methods available. I wanted to try to use TRS for the subproblem, but realized that it's not release yet. Is it your intention to release it "soon"?
[I'm currently resorting to implementing More Sorensen from the paper, but I'm having some difficulties matching their claims about performance, so I expect something is off in my implementation]
Hey Patrick, thanks for the interest in this repo! Yes, I will aim on releasing it this by the start on next week.
An unrelated question: would you be interested in support of ellipsoidal norms, i.e.
minimize ½x'Px + q'x
subject to x'Cx ≤ r
or does the standard 2-norm (x'x ≤ r) would cover your needs? If you are interested in ellipsoidal norms would the simple option of a diagonal C be enough?
Ellipsoidal norms would be great, and yes I think diagonal C should be sufficient.
I actually just tried TRS against my own More Sorensen, and I seem to get the same solutions on the different (easy and hard case) problems I've tried. It's just using way too many iterations. You havn't tried to implement M&S yourself by any chance? Doesn't have to be julia. It's just that gqtpar is GPL so I'm trying not to peak under the hood :)
Had a look at the reference paper for this package, and it seems like a great approach. Looking forward to testing it out !
Thanks! No, unfortunately I haven't tried More Sorensen.