Gurobi finds non-optimal results for different solution paths
AnsweredHello,
I've encountered a problem with Gurobi finding an optimum that is far off from the correct optimum. The optimization model is a MIP with convex quadratic constraints.
The issue seems to be that Gurobi has a wrong BestBd value. This leads to Gurobi terminating early with a wrong optimum (~3826). The correct optimum would be ~5663.
Take a look at the MIP Log:By manually fixing some binary variables to their optimal results, Gurobi finds the correct optimum.
Interestingly, when Gurobi takes a different solution path (e.g. different Parameters or Seed), then the correct optimum is found sometimes. To my understanding, a different solution path affects the performance, but it should not result in a different optimum (except MIP Gap > 0). In my case, a different Seed value brings up a solution with binary variables at different values for each solution.
Testing the model (.lp files)
Unfixed (non-optimal): https://drive.google.com/file/d/1YTpS2aq5W7IjrmZHdq4y8Egfp8vWGmi2/view?usp=sharing
Fixed (optimal): https://drive.google.com/file/d/1zF_nNkq4jXc-sXE3CBhztM2iVN0wifKt/view?usp=sharing
Keep in mind that when testing the model on a different machine, the issue might not appear (because of a different solution path).
Version:
Gurobi Optimizer version 11.0.3 build v11.0.3rc0 (win64 - Windows 11.0 (22631.2))
Any help or comments are welcome. What might be causing this? A potential bug in Gurobi? Problems with the model formulation? Or is the behavior expected? Has anyone ideas for further investigation?
Thanks a lot!
- Julian
-
Hi Julian,
Thanks for all the info! We will take a look.
Cheers,
David0 -
Hi Julian,
We have identified the cause and will fix it in a future release.
Cheers,
David0 -
Hi David,
Thank you for the investigation. Looking forward for the fix.
- Julian
0 -
Hi Julian,
We have released Gurobi 12.0 which includes a fix for this bug.
I have already tested it again on this model but please feel free to download and test it.
Cheers,
David0
Please sign in to leave a comment.
Comments
4 comments