[Ipopt] convergence problem

[Andreas Waechter]

Hi Juan,

What you describe usually shouldn't happen, and it's difficult to know 
what is going on.  The first thing to do is always to check that the 
derivatives are correct (using the derivative checker).

Andreas

Andreas Waechter

Associate Professor
Department of Industrial Engineering and Management Sciences
McCormick School of Engineering
Northwestern University
Evanston, IL 60208
USA


On 02/08/2012 08:23 AM, Juan Carlos Lopez Alfonso wrote:
> Hi all,
>
> I have a serious problem with my optimization algorithm.
>
> Supose that I have a problem where the convergence to the optimal
> solution (and unique, because is convex) using IPOPT is achieved. This
> convergence is achieved in several iterations, but it is always
> achieved even when I change the options "tol" and "acceptable_tol" for
> example 1e-10. Now I have removed a constraint and my problem still
> have a unique solution, but the convergence is no achieved when the
> tolerance are less than 1e-5. I have implemented the same problem in
> matlab using the optimization toolbox and here for these two cases the
> convergence is achieved when I change the tolerances.
>
> The problem is the follows: when I set for example the tolerances
> app->Options()->SetNumericValue("tol", 1e-8);
> app->Options()->SetNumericValue("acceptable_tol", 1e-7);
>
> the algorithm starts to decrease the function value, but when almost
> is achieved the desired result (which I know comparing with matlab),
> come several iterations (maybe 10) increasing the function value and
> again iterations decreasing the function value, and so on.
>
> I think that I have a problem with the step calculation or any
> tolerance related with this. What it is clear si that the problem is
> related with some options of IPOPT (ah, I am using mumps) because all
> is compared with matlab which work perfect.
>
> Please, could anyone give me suggestions
> thank in advance and best regards
> Juan Carlos
> _______________________________________________
> Ipopt mailing list
> Ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/ipopt