Run local optimization with inital guess

[ahonorat]

Hello,

For some NLP model, I would like to run SCIP (from PySCIPopt) as a local optimizer only and I know a valid starting solution.
My goal is to quickly refine the starting solution to the closest local minima.

Is this possible ?

I've seen a very similar question for the SCIP API, but it's a bit old so things may have changed.
https://listserv.zib.de/pipermail/scip/2014-March/001862.html

Thank you

[Joao-Dionisio]

Hey @ahonorat ! I have no experience with this, but I asked Stefan Vigerske, and he said it might be possible, albeit complicated. For this, you'd need to get into the solving stage of SCIP (using an event handler), get the NLP relaxation (I'd need to interface SCIPgetNLPI()), and then solve the relaxation with an NLP solver attached to SCIP with a given starting point. For this, I assume you could use SCIPsetNLPInitialGuessSol(), but maybe it's only for the global problem. It would also need to be interfaced in PySCIPOpt.

[ahonorat]

Thank you @Joao-Dionisio , so it's theoretically possible in SCIP and not yet in PySCIPOpt, it's good to know.

I've read that you have less time to code this year, feel no pressure on this because we have other workarounds. Thank to your answer, at worst we can use SCIP directly instead of the Python interface.