[Coin-ipopt] IPOPT within any MINLP solvers?

[kevin.c.furman at exxonmobil.com]

Hello,

With only about 10 binary variables, might I suggest just writing a simple
decomposition based approach using outer approximation for solving the
MINLP?  It could be written in AMPL and call any MILP solver for the master
problem and then IPOPT for the NLP subproblems.  There are variations for
convex and nonconvex MINLPs.

Suggested reading:
Duran & Grossmann, Math. Prog. (1986)
Viswanathan & Grossmann, Computers & Chemical Engineering (1990)
Fletcher & Leyffer, Math. Prog. (1994)
Kesavan et al., Math. Prog. (2004)

Regards,

Kevin Furman



Kevin C. Furman, Ph.D.
Corporate Strategic Research
ExxonMobil Research & Engineering
1545 Route 22 East
Annandale, NJ 08801
(908) 730-2281
kevin.c.furman at exxonmobil.com





Date: Tue, 15 Feb 2005 16:39:30 -0500
From: "Frank J. Iannarilli" <franki at aerodyne.com>
Subject: Re: [Coin-ipopt] IPOPT within any MINLP solvers?
To: Andreas Waechter <andreasw at watson.ibm.com>
Cc: coin-ipopt at list.coin-or.org
Message-ID: <40DB94537C26A2035C3B7441 at beautz2.aerodyne.com>
Content-Type: text/plain; charset=us-ascii; format=flowed

--On Tuesday, February 15, 2005 4:22 PM -0500 Andreas Waechter
<andreasw at watson.ibm.com> wrote:

<snip>
>
> Just out of curiosity, would your MINLP problems be convex?  Also, would
> you want to provide you model in AMPL, or would you want to use some
C/C++
> interface?
>

Hi Andreas,

The formulation I'm using at the moment is linear objective with quadratic
non-linear constraints -- it's actually a least-absolute-deviations (LAD)
fitting, with binary variables added to do some variable selection.  So the

objective is the sum of deviations, and the constraints use the binary
variables to zero-out selected deviations (with respect to hidden variables

of the optimization).  Oh, so to answer your question:  convex.

I'd personally be most happy with an AMPL interface.

Good luck with your IPOPT/CBC MINLP effort!  You might periodically update
this maillist as to your progress :-)

BTW, in the meantime, to solve my problem, is it worth my contemplating a
quick and dirty hooking up of IPOPT to some B&B code?  I presently have
"only" about 10 binary variables.  If so, any basic pointers?  Or is this
silly?

Thanks!



Frank J. Iannarilli,  franki at aerodyne.com
Aerodyne Research, Inc., 45 Manning Rd., Billerica, MA 01821 USA
www.aerodyne.com/cosr/cosr.html