Closed odow closed 2 years ago
Hi Oscar, thanks for reaching out.
Yes, we can definitely have such a discussion at some point. Right now I'm still working on some improvements for my algorithm, since in its current form it is interesting from a theoretical perspective, but not really efficient computationally (I guess that's in some sense similar to what you experienced in your tests on Lagrangian cuts).
I've already had a look at a lot of the changes that you made and some of them I have already included in the branch that I'm currently working on. Your idea on using first-order methods looks interesting to pursue, I haven't really thought in this direction yet.
Nice. I saw your MP paper but I haven't had a chance to read it yet. At some point we should discuss what parts would be useful to integrate back into SDDP.jl.
I took a deeper look at the Lagrangian solver in SDDP.jl last year and made some changes: https://github.com/odow/SDDP.jl/blob/5ec8a2e4baa66453af480a539b3cdc3d7679c6fe/src/plugins/duality_handlers.jl#L145-L224
I'm sure there must be a good first-order non-smooth algorithm for convex functions out there, but I haven't had a good search. It doesn't need to be anything specific to math programming or Lagrangian duality.