[SCIP] Sub-optimal solution produced after upgrade to SCIP version 8.0.3

[Mexi, Gioni]

Hi Barry,

Thank you for the bug report. We looked into the issue, and it appears to be a valid bug when using "numerics/feastol = 1e-09" that is resulting in an invalid node cutoff.
It requires further investigation. We will reach out to you with a fix as soon as we have one.

Best regards,
Gioni

________________________________
Von: Scip <scip-bounces at zib.de> im Auftrag von Barry Hurley <b.hurley at keelvar.com>
Gesendet: Montag, 24. April 2023 15:06:30
An: scip at zib.de
Betreff: [SCIP] Sub-optimal solution produced after upgrade to SCIP version 8.0.3

Hi, we were looking at upgrading from SCIP version 7 to version 8.0.3 but one of our unit-tests is catching a regression where SCIP appears to be producing a sub-optimal solution and claiming optimality. We are using SCIP via the LP file interface, the files to reproduce are attached. problem.lp is the original problem, and problem-forced-solution.lp extends this with an additional constraint called `FORCED_SOLUTION` to force the optimal solution. Also including the parameter settings used.

The solution should have an optimal objective value of 115, but SCIP version 8.0.3 is incorrectly producing a solution with objective value 120, even with gap zero. The solution is valid, just sub-optimal. There is one difference in the assignment, binary variable X_2__1_ has an objective coefficient of 15, but SCIP 8 is incorrectly assigning binary variable X_3__1_ to 1 instead even though its objective coefficient is higher at 20.

Would you have any advice if there is something we are using SCIP incorrectly, or if this is a legitimate bug? Other than the attached parameter settings, we are using the pre-built binaries on ubuntu x86. Many thanks in advance.

Regards,
Barry Hurley
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.zib.de/pipermail/scip/attachments/20230424/ce72baa2/attachment.html>