Originally posted by: lyxthe

I have a MILP scheduling problem (7726 rows, 3892 columns, and 28427 nonzeros),  I use cplex with MIPEmphasis to 1. 

In a first execution of my problem, I set EpGap to 0.05 and the execution stoped in arround 150s. In the log I can see the solution found have a gap to 2.28%

In a second execution of my problem, I don't set the EpGap parameter and let it by default. After 800s I cut the execution myself, and see in the log that the best feasible solution found yet has a 8.78% gap.

Questions :

In the first execution, dispite the EpGap parameter set to 0.05, why did the program stop with a gap to 2.28% ?

In the second execution, why, after 800s, the best solution found has a gap to 8.78% while in the first execution the best solution found after 150s is well better with a gap to 2.28% ? Nothing changed between the two executions excepted the EpGap parameter.

Thanks for your attention.

Note that the EpGap perameter is not given in percent but as a ratio between primal and dual bound. See also the reference documentation. The reference documentation even gives 0.05 as an example and says that this means to stop within a gap of 5%. So stopping the first run at a gap of 2.28% is prefectly fine, I think.

Do you change any other parameters than EpGap and MIPEmphasis? Can you please double check that everything but EpGap is exactly the same in the two runs? Could you show us the log files for both solves?

Originally posted by: lyxthe

HI, thanks for your answer. Of course I've just read the first question and I realize it is fine for me too :)

For the second one here are my log, if it can help :)

Just to explain there are a few differences between the numbers in the logs and those I asked about, it is because when I wrote the question, I didn't have the logs under the eyes and so didn't write the good numbers. But it is those executions that I spoked about.

Thanks for your attention.

Sorry for the very long delay. What you observe is unexpected. The two log files start to differ with the following line:

log1: *  7694  5393      integral     0      400.0000      291.0000    78983   27.25%
log2: *  7694  5392      integral     0      400.0000      291.0000    78983   27.25%

The expected behavior is that both lines are equal. Once you see two different log lines everything is possible, in particular the difference in the solution process you observe.

This looks like a bug in CPLEX. What version of CPLEX do you use? Would be able to disclose your model so that we can take a look at what is going wrong?