[Couenne] A problem where Couenne appears to fail?

[Tony Kelman]

I remember having seen a bug or two with scaling in Couenne on maximization problems. Maybe change the sign of your objective in your modeling environment and compare the three where all are minimizations?



On Thu, Sep 22, 2016 at 2:10 AM -0700, "A. Torgovitsky" <atorgovitsky at gmail.com<mailto:atorgovitsky at gmail.com>> wrote:

Hello,

I have been comparing KNITRO, Couenne and BARON for some sample problems related to my research.

The problems are mixed integer nonlinear programs. However, the only non-integer non-convexities are through bilinear terms in the objective. So, my understanding is that both Couenne and BARON should be able to solve this problem to global optimality.

I've attached one example of such a problem to this email. (Note that the problems are maximizations not minimizations.)

When I solve the problem with the latest versions of KNITRO, Couenne and BARON I get the following (feasible) optima:

BARON: .00202
KNITRO: .00198
Couenne: -.00202

Is there some sort of bug here, or am I misunderstanding something?

My initial concern was that Couenne < KNITRO, which shouldn't occur if Couenne is finding the global optimum. However when I was writing this email I realized that the value Couenne returns is -1* the value that BARON returns. That seems like an odd coincidence -- is this maybe just some sort of output bug?

Thanks,
Alex
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/couenne/attachments/20160922/35a89028/attachment.html>