[Coin-ipopt] RE: Coin-ipopt Digest, Vol 19, Issue 5

[damien at khubla.com]

Hi Ivan,

I've reached the same conclusions as you with L-M compared to SQP-based
algorithms with least-squares in previous employment.  Unfortunately, my
current employer doesn't have access to a L-M code in the short term, and
I'm on a tight timeline, so I was hoping that perhaps someone had a good
set of tuning factors for IPOPT.  On those rare least-squares problems we
have with nonlinear constraints, IPOT has been excellent, but the great
majority of the least-squares problems we have are pure least-squares with
no constraints.

Damien


> Hi Damien,
>
> I'm a user of Ipopt and we also do a lot of nonlinear least-squares
> fitting around here (including of the black-box variety as in your
> case), so you might be interested in our experience.  This is an issue
> that we come back to occasionally but I think we've come to one definite
> conclusion (I'd be very interested in other experiences of members of
> this list).
>
> Ipopt has done an excellent job for us in solving general nonlinear
> optimization problems.  But we've found that it's a bit of an unfair
> comparison when we benchmark it on least-squares problems versus a
> specialized algorithm like Levenberg-Marquardt *because of the data
> structures involved*.  For model-fitting, the latter will (for our
> benchmarks) always beat Ipopt and other IPM codes without exception, but
> again - this is an unfair comparison since L-M is highly specialized for
> these problems and IPMs are very general.  One place where L-M falls
> short (and where Ipopt excels) is in handling nonlinear constraints, but
> these usually don't show up for us in model-fitting since our parameters
> are typically decoupled.  Our L-M implementation is proprietary so I'm
> not familiar with free code that you can use, but I'm sure they do exist
> (implementation is pretty straight forward so you can experiment
> easily).  Just one suggestion if you decide to go this route - take
> advantage of the data structures in the L-M implementation, otherwise it
> really isn't that different from bound-constrained interior point
> methods.
>
> Now perhaps someone can actually answer your question ;) (about good
> parameter choices for Ipopt) - I'd like to hear that too.  Also very
> interested in others' experience re L-M vs Ipopt for LS.
>
> --Ivan.
>
> ----------------------------
> Ivan B. Oliveira
> SC12-205
> (408)765-0584
>
> -----Original Message-----
> From: coin-ipopt-bounces at list.coin-or.org
> [mailto:coin-ipopt-bounces at list.coin-or.org] On Behalf Of
> coin-ipopt-request at list.coin-or.org
> Sent: Thursday, June 15, 2006 9:00 AM
> To: coin-ipopt at list.coin-or.org
> Subject: Coin-ipopt Digest, Vol 19, Issue 5
>
> Send Coin-ipopt mailing list submissions to
> 	coin-ipopt at list.coin-or.org
>
> To subscribe or unsubscribe via the World Wide Web, visit
> 	http://list.coin-or.org/mailman/listinfo/coin-ipopt
> or, via email, send a message with subject or body 'help' to
> 	coin-ipopt-request at list.coin-or.org
>
> You can reach the person managing the list at
> 	coin-ipopt-owner at list.coin-or.org
>
> When replying, please edit your Subject line so it is more specific
> than "Re: Contents of Coin-ipopt digest..."
>
>
> Today's Topics:
>
>    1. Ipopt for nonlinear least-squares (damien at khubla.com)
>
>
> ----------------------------------------------------------------------
>
> Message: 1
> Date: Wed, 14 Jun 2006 12:34:59 -0600 (MDT)
> From: damien at khubla.com
> Subject: [Coin-ipopt] Ipopt for nonlinear least-squares
> To: coin-ipopt at list.coin-or.org
> Message-ID: <39255.142.179.218.30.1150310099.squirrel at mail.khubla.com>
> Content-Type: text/plain;charset=iso-8859-1
>
> Hi all,
>
> we've been using Ipopt with good success for the last few months with a
> view to replacing our commercial license for Dash.  So far it's been
> great, and we're extending the use to some nonlinear least-squares
> problems.  The performance is generally good, but I was wondering if
> anyone has any experience finding a good set of solver parameters that
> might do weel on least-squares problems.  We treat the problem to be
> fitted as a black-box and use numerical differences to calculate
> derivatives, and use the quasi-newton Hessian approximation.  Most
> problems have between 5 and 50 variables and 10 or less constraints.
>
> Damien
>
>
>
> ------------------------------
>
> _______________________________________________
> Coin-ipopt mailing list
> Coin-ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/coin-ipopt
>
>
> End of Coin-ipopt Digest, Vol 19, Issue 5
> *****************************************
>
> _______________________________________________
> Coin-ipopt mailing list
> Coin-ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/coin-ipopt
>