[Coin-ipopt] Handle the problem without using automatic differentiation

[Andreas Waechter]

David,

> Thanks for your quick reply.
> I just have Ipopt Fortran version installed. According to what you said, I
> just need to provide the Jacobian, is that right?

In principle, yes.  But keep in mind that the code usually become less
efficient and robust if second derivatives are only approximated.
(Computing second derivatives by finite differences - if done exactly
enough - would probably improve robustness, but would probably also make
the whole thing quite slow).

One things to keep in mind is that you want to give Ipopt sparse Jacobian
(and Hessian) matrices, if your objective and constraint functions exhibit
such a sparse structure.  The dense the matrices are, the more expensive
the factorizations (and therefore iterations) become.

Andreas

>
> Thanks,
>
> David
>
>
> -----Original Message-----
> From: Andreas Waechter [mailto:andreasw at watson.ibm.com]
> Sent: Tuesday, December 20, 2005 5:10 PM
> To: Zhongzhou Chen
> Cc: coin-ipopt at list.coin-or.org
> Subject: Re: [Coin-ipopt] Handle the problem without using automatic
> differentiation
>
> Hi David,
>
> Ipopt does not approximate any derivatives by finite differences.  If you
> want to obtain derivatives that way, you could do so yourself in your code
> that implements the computation of the Jacobian and Hessian.
>
> At some point, Ipopt will have an option to approximate the Lagrangian
> Hessian by quasi-Newton methods (the Fortran version already has that, but
> the C++ version not yet).  But first derivatives will assumed to be
> available from the optimization problem.
>
> Regards
>
> Andreas
>
>
>
> On Tue, 20 Dec 2005, Zhongzhou Chen wrote:
>
> > Hi, all:
> >
> >
> >
> > I am trying to solve a large scale optimization problem. Automatic
> > differentiation is highly impossible due to the model structure.
> >
> > Is there an option to allow Ipopt to calculate the Jacobin and Hessian
> using
> > finite difference?
> >
> > Should I try other software, like Conopt?
> >
> >
> >
> > Thanks,
> >
> >
> >
> > David
> >
> >
> >
> >
>
>