user_bound_scale and 0.11 vs. 0.12

[infinity77]

Dear HiGHS community,

I’m solving a large number of relatively small LPs (by HiGHS standards), one after the other. Up to now I’ve been using 0.11 and everything seems to be fine. To note: I use user_bound_scale for all my problems, and then “unscale” the solution by the same quantity.

Today I’ve compiled 0.12, to the metal like 0.11, using CMake and Visual Studio 2022. To my surprise, I’m getting ridiculously enormous/unexpected results.

Did anything change in the handling of user_bound_scale? I’ve seen a note in the release notes, but I’m confused as it seems related to MIP problems. I only have LPs.

Am I supposed to still unscale the results of solve?

Thank you in advance for your help.

[jajhall]

The solutions are definitely unscaled in v1.12 - which explains the warning messages if they don't satisfy the primal or dual feasibility tolerances.

Your experience does seem to imply that in v1.11 such unscaling wasn't done.

Apologies, I should have stated this in the release notes.

[infinity77]

No apologies necessary @jajhall . I may just have misinterpreted the release notes. I’ll make some tests and report back. I’m not expecting earth-shattering differences between 0.11 and 0.12 as my LPs are still way too small to affect runtime between minor versions.

[jajhall]

For LPs nothing has changed with the user scaling - other than unscaling the solution.

If you build with our new interior point solver (HiPO) you may notice a decent performance gain

[infinity77]

Thank you @jajhall . Does the new HiPO solver require any special additional libraries?

Edit: looks like it does, Metis and BLAS. That’s going to be fun on Windows.

[jajhall]

Yes, but it does work. Only a big gain if your LP problems are large