Originally posted by: AnirudhSubramanyam

Hi,

I've attached part of a screenshot of a node log during a particular MILP run implemented using the C API of V12.6 cplex. I'm using a cut callback function, lazy constraint callback function and a branch callback function in my implementation. The branch callback function may or may not create user defined branches. If it does, it will always create one and only one child branch, where several binary variables which are already at 0 or 1 (in the optimal LP solution of the parent node) are set/fixed to that value. If not, the default CPLEX selected branching is used.

The anomaly is as follows: consider the display logs for nodes 3, 4 and 7. The node objective value at each of these nodes (all of them have been created as a single child using my branch callback function as described above) is lower than the best bound value as reported by CPLEX.

Can anyone explain this?

Thanks,

Anirudh

What is the objective estimate you set for the newly created nodes in your branch callback? Is that just the objective value of the parent's LP relaxation or do you use something else there? If you use something else and that is larger than the actual LP relaxation of the child then this may explain the behavior you observe.

Originally posted by: AnirudhSubramanyam

I set the objective estimate to be the same as the parent.

Then this smells like a bug in CPLEX. Do you happen to have a small example code that reproduces this issue?

Originally posted by: AnirudhSubramanyam

Hi Daniel,

I have an LP file, on which if you run vanilla CPXXmipopt(), the above issue is reproduced, i.e., everything default CPLEX with no callbacks. Do you want me to send you the LP file?

Yes, that would be very helpful. Please send the file to daniel(dot)junglas(at)de(dot)ibm(dot)com. Thanks!