[Ipopt] Very slow convergence of IPopt

Andreas Waechter andreasw at watson.ibm.com
Tue Dec 1 14:41:02 EST 2009

Hi Jan,

If you see so many iterations, it probably means that your Hessian matrix 
is not correct.  Have you tried Ipopt's derivative checker?

If you set the print_level large enough, Ipopt will print out the values 
of the matrices as well, and you could compare them between the original 
model and the one where you use the automatic differentiation code.



On Tue, 1 Dec 2009, Jan gekheid wrote:

> Hello Andreas,
> I'm using the Java interface of IPopt and I've implemented an automatic
> differentiation library (JEP Java Expression Parser) to automatically do the
> differentiation of the objective function and the constraints. I implemented
> this in the given HS071 example so that I can check the results. All the
> automatically differentiated formulas come out the same as in the original
> problem. The model and the results are attached to this mail.
> It gives the right results, but the convergence is very slow. It takes more
> than 3000 iterations and even then it stops because the maximum number of
> iterations is exceeded. It is exactly the same problem as the original (as
> for the formulas and starting conditions), except that the values for the
> jacobian etc are calculated with the automatically differentiated functions.
> This implementation is just to see whether the combination of using IPopt en
> JEP is possible (which it is apparently). But I want to implement a much
> larger problem with very much larger formulas for the hessian entries. It
> would take days to compute just one problem if it takes this much time. Is
> there anything I can do about that?
> With kind regards,
> Jan

More information about the Ipopt mailing list