Originally posted by: EXCT_RALF_GOLLMER

Hi,

now I encountered an effect with CPLEX 12.6.2:

for a rather small problem problem with all stable bases the integer optimal objective value increases by  .0897512 % (from 8.9001517504e+03 to 8.9081397439e+03) and represolve is not carried out fully (at least it seems so by the output) when  the simplex feasibility tolerance is set from the default to the minimal value (1e-09).

Does this rather big difference make sense? To me it seems more like a bug in repeating presolve.

The attached file contains the model file, mip starts, the different cplex commands (Cplex.in*), the call sequence cplex_call, the solutions produced, and cplex.log containing the mip quality displays, too.

The two files outside of the tar show that the problem occurs only with mip emphasis 3. That run uses the same problem and mst files from the tar.

Best regards

Ralf Gollmer

This indeed seems to be related to represolve. After disabling represolve (set pre repeat 0) all input sequences produce the same objective function value here.

I will take a look at what is going on internally but from experience I know that it may take a long time to track this down. So if this is a problem for you right now, are you OK with disabling repeated presolve?

Originally posted by: EXCT_RALF_GOLLMER

Hi Daniel,

this is not really a problem for me, I will disable represolve (though this really helps with most problems to speed up solving).

The effect found might explain results we got for SIPLIB test problems having less objective values than the published ones.

It would be important to eliminate this flaw in CPLEX, but I fully understand that this will require thorough testing and that this might need considerable time.

This forum is the only way I know to submit a bug report for someone with an academic license, thus being without an active support contract.

Best regards

Ralf
 

Ralf, thank you for reporting this in the first place. I already filed an internal bug report after my previous post. This should make sure that this problem gets looked at. Please keep reporting bugs you find here, it is the official channel for people without a commercial contract. if you don't want to report bugs here, please feel free to directly send them my way: daniel(dot)junglas(at)de(dot)ibm(dot)com.

Originally posted by: EXCT_RALF_GOLLMER

Hi Daniel,

it seems the problem couldn't be pinpointed before the release of 12.6.3, since it's behaviour wrt. to this problem is exactly the same as in 12.6.2.

Best regards

Ralf

Hi Ralf, this issue is not yet supposed to be fixed in 12.6.3, it came in too late in the development cycle :-( So it is expected that you still see the same behavior.