"optimal with unscaled infeasibilities" AMPL with CPLEX

[BenP]

Hi all,

I'm running a MILP with a few thousands variables and constraints. I'm using the CPLEX solver and I often get a very unstable behavior, and the solution found is obviously not optimal.

When I run AMPL I typically get the message optimal with unscaled infeasibilities.
I know it's a CPLEX message, related to a ill defined problem, for example very different orders of magnitude for the coefficients in the constraints..

But how can I see these coefficients in AMPL? Is it possible to see explicitly of CPLEX solving matrix? Or any other ideas where I need to look for the source of the problem?
I've already used the display function of AMPL, for lb, ub, body, slack, dual values, but I didn't see anything problematic.

I'm writing my Master Thesis due in 4 weeks, and I've already spent weeks on this unstability behavior, so I thank you very much for your help =)

[Paul]

Numerical instability is not always caused by poor scaling, although that is a frequent contributor. It's not hard to get CPLEX to show you evidence that instability exists, but I don't know any way to get it to isolate a specific cause.

You can write AMPL statements to print max and min absolute values for the coefficients of each variable, which can help you find scaling problems in the data. Beyond that, you might be better served looking for remedies (for instance, the CPLEX numerical emphasis parameter) than causes.

Paul

[bp.benoi...@gmail.com]

Thanks Paul for your answer.
I tried to use the numericalemphasis  option but unfortunately it didn't give me a better solution. Any thing else, or a place where I could find a list of options that can improve the accuracy?

When you say values for the coefficients of each variable, are there the coefficients of the scaling? How can I display them?

[bp.benoi...@gmail.com]

It works!
Unfortunately I didn't explicitly find the problem. But by trying to rescale some parameters of some constraints, the optimal with unscaled infeasibilities message ended up disappearing, and the solution is now clearly optimal.

Thanks for your help,
BenP