Column Generation example in PySCIPOpt

[belzebuu]

Hi,

I would like to implement a column generation approach in PySCIPOpt for teaching purposes.

I have seen from tests/test_pricer.py that it is possible to use the callback function includePricer(). But I would like the students to work on a lower level to understand better
the technique.

I would like to follow the approach of: examples/unfinished/cutstock.py

However, I understood that relax() and column() are not available/implemented yet.

I could resort to reimplementing from scratch the restricted master problem every time a new variable is added. Thus no warm restart.

Is there another way I could mimic at this stage the implementation of cutstock.py. For example using addConstrCoeff()??

Thank you!
Best regards,
Marco

[fserra]

sorry we never replied. is this still relevant?

[belzebuu]

Yes 😊

…

-- Marco Chiarandini, Asc. Prof. Department of Mathematics and Computer Science, University of Southern Denmark, Campusvej 55, DK-5230 Odense M, Denmark. Phone: +45 6550 4031 marco@imada.sdu.dk<mailto:marco@imada.sdu.dk>, http://www.imada.sdu.dk/~marco From: fserra <notifications@github.com> Reply to: SCIP-Interfaces/PySCIPOpt <reply@reply.github.com> Date: Thursday, 28. May 2020 at 21:11 To: SCIP-Interfaces/PySCIPOpt <PySCIPOpt@noreply.github.com> Cc: Marco Chiarandini <marco@imada.sdu.dk>, Author <author@noreply.github.com> Subject: Re: [SCIP-Interfaces/PySCIPOpt] Column Generation example in PySCIPOpt (#281) sorry we never replied. is this still relevant? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub<#281 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ADDRXCHLPT3OOBJ5YHPRVLTRT2ZPHANCNFSM4HOMQPBA>.

[Joao-Dionisio]

Hey @belzebuu :) I don't know if the code is still useful for you, but I think I did what you wanted.

Here is the code:
cutstock.zip

Let me know what you think!

[belzebuu]

Dear @Joao-Dionisio,

yes! this is what I wanted. It will be useful for my teaching purposes.

I have read the discussion around the workaround on line 48. Maybe if you publish this example under the finished ones, you could avoid redefining y below on line 73, as it maybe confusing.

I will then reconsider pyscipopt for the new term!
Thanks a lot for keeping this in mind!

[Joao-Dionisio]

I'm very glad I could help @belzebuu , I hope your students end up liking pyscipopt!

As for publishing the code, it is not in SCIP's philosophy to use it as a black box solver. The proper way would be to use a pricer as you first mentioned. If it interests you, you can read the discussion on Pull Request #617 . Thank you for the recommendation on the y variables, though, you're absolutely right.

Good luck for the new term!

[belzebuu]

Thank you @Joao-Dionisio for making me aware of that discussion. Yes, it is of interest for me. I have checked test_pricer and took note that, that is the "SCIP way". I am not so familiar with SCIP after all. I agree that in test_pricer there is more emphasis on the LP vs SCIP distinction and that it is for this reason instructive. What I fear might be more problematic to teach is the inheritance from the class Pricer. I will give it some more thoughts but nice to have now both possibilities.