Originally posted by: GokhanCeyhan

Hi,

I am trying to solve an miqp problem attached with the specified params.prm file, but I get a singular basis warning and cplex cannot solve the root relaxation during the specified time limit.

Can you suggest any way to overcome this issue?

Thanks,

What version of CPLEX do you use and on how many threads? I could not reproduce your problem here with any of the recent CPLEX releases.

You may also want to set CPX_PARAM_MIPDISPLAY=4 to see the barrier log for the root relaxation solve.

Originally posted by: GokhanCeyhan

We use Cplex 12.8.0 without changing its default setting for the number of threads. I think it uses 8 threads.

Originally posted by: BoJensen

Just like Daniel, I would like to reproduce the run and understand what goes on first, but since you ask for a way to overcome the issue I suggest increasing the markowitz tolerance. Also you have some high numbers in the objective, looking into reducing these (if possible) may also help. Please send a log file.

Originally posted by: GokhanCeyhan

Hi,

I attached the log file now. With the previously attached params and model file, I can just reproduce this log with Cplex 12.8.0.0.

Regards,

Originally posted by: BoJensen

Thanks for the log. Can I ask you to set CPX_PARAM_MIPDISPLAY=4 as Daniel suggested and rerun ? I would like to see the log from the problematic root solve.

I can not reproduce the issue.

Originally posted by: GokhanCeyhan

By the way, increasing Markowitz tolerance to 0.99 solves the issue. However, what are the downsides of increasing this threshold? Can you suggest an appropriate threshold value?

Thanks,

Originally posted by: BoJensen

Increasing the markowitz tolerance to 0.99 in this case is rather extreme, I would raise it slightly to focus more on numerical stability in the LU factorization. Increasing this tolerance has the drawback that the LU factors can become more dense and the solution time increases.

Originally posted by: BoJensen

Try raising it to 0.1 or 0.2.

Originally posted by: GokhanCeyhan

You can find the extended log file now. ("log2.txt")

Thanks.

Originally posted by: BoJensen

Thanks, did using markowitz tolerance 0.1 or 0.2 help ?

Originally posted by: BoJensen

Which OS and machine are you running CPLEX on ?

Originally posted by: GokhanCeyhan

Yes, it helped with 0.1. Please find also the machine configurations in the attached file, "machine.PNG".

Originally posted by: BoJensen

I tried several ways to reproduce the issue, but failed. It's probably related to the numerics of your model, though the matrix coefficients doesn't look that bad, only the large objective coefficient sticks out. I suggest you raise the markowitz tolerance to a reasonable level such as 0.1. Should you experience this issue on other models, please share the model in the forum or with me directly bo.jensen (at) dk.ibm.com.

Originally posted by: GokhanCeyhan

Thank you very much, appreciate your help.