Presolving

[gdalle]

• Papilo.jl
• QuadraticModels.jl

[klamike]

@gdalle Planning to address this next. I looked into Papilo and it is definitely more advanced than what QM implements. The PDLP paper shows that it can bring a decent speedup too. Another thing I didn't realize earlier is that even QM is not MIT, it is MPL 2.0.

I'm thinking I can set up a benchmark script that uses the MPS file API and lets you enable/disable each Papilo feature, to see which are the most impactful. Then we can implement the most important ones here so we can call them directly. But we can only look at the papers when writing the code since we want to release it under MIT. What do you think?

[gdalle]

I think it might be a colossal waste of our time to reimplement presolve from scratch. We could just give people the option to call Papilo.jl during MILP construction if they supply MPS files? Papilo's license is not MIT but that doesn't mean our own code can't be, we just need to avoid redistributing it as a binary. To be on the safe side, we can put that add-on in a package extension so that any users have to voluntarily install Papilo to use it

[klamike]

Yes, that makes much more sense.. just thought you wanted everything MIT (and frankly I don’t know much about licenses). I’ll make a PR for the extension

[gdalle]

I think requiring users to explicitly install and opt into any non-MIT deps is a good middle ground.
You should also chat with Mathieu Besançon, he was planning to work some more on Papilo.jl (can't tag him on a private repo)

[klamike]

https://github.com/klamike/PaPILO_Presolve_jll some progress, should have some time next week to finish wiring it up

[gdalle]

Any reason not to use the existing https://github.com/scipopt/PaPILO.jl or https://github.com/JuliaBinaryWrappers/SCIP_PaPILO_jll.jl ?

[klamike]

Mostly just that I’ve been looking for an excuse to figure out the binarybuilder/jll side of Julia and this seemed like a good candidate. It would avoid the MPS file intermediary. I’m thinking if it works well I can PR upstream, CoolPDLP would just have an extension for PaPILO.jl

[klamike]

Some more progress https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/compare/mk/presolve?expand=1

[gdalle]

I think @matbesancon might be interested