Fwd: [eigen] Non linear equations solver

[koldo.ramirez@xxxxxxxxx]

Hello list

I have been searching MINPACK ports to C and it seems the most complete is CMinpack (http://devernay.free.fr/hacks/cminpack.html). Its license is the original Minpack license, so it lets to be used in proprietary programs.

It includes non linear equation solver and 
nonlinear least squares too. 

To have a first contact I have compiled the samples and run well.

What do you think about including it in Eigen ?

Best regards
Koldo

---------- Forwarded message ----------
From: <koldo.ramirez@xxxxxxxxx>
Date: 2009/7/21
Subject: Re: [eigen] Non linear equations solver
To: eigen@xxxxxxxxxxxxxxxxxxx


Hello Keir

N-R uses a multidimensional secant method using global strategies for zero finding based on line searches. N-R indicates that hook step and dogleg step methods "have a reputation for robustness even when starting far from the desired zero or minimum". They seem to do line searches when far to the zero and cuadratic when close to the zero so that the convergence is better.

In some sets of equations I use I have found convergence problems that sometimes slow the program and others avoid to solve the equations. This is an important problem so I really need something better.

Do you have a test program so that I can try it?

Best regards
Koldo

By the way: I have seen the YouTube "Matchmoving with libmv: Now with Suzanne!" in http://www.youtube.com/watch?v=hswISipCv2M, and it is amazing!. Excelent job.

2009/7/21 Keir Mierle <mierle@xxxxxxxxx>

On Mon, Jul 20, 2009 at 7:14 AM, <koldo.ramirez@xxxxxxxxx> wrote:

Hello Keir

I have read a little bit and Levenberg-Marquardt and Powell's dogleg seems to be a better solution than the programed in N-R.

It depends on the problem. It's usually best to try both.

 

Is it possible to use libmmv non linear equations solver?. Is it possible to be used without Qt4?

It's a nonlinear minimizer, not an equation solver (though it may work depending on your case). The Qt4 code is not needed. However, the libmv minimizer is really just a toy. It hasn't had any real world use beyond the unittests.

If you were to use it, we're happy to accept patches to improve it.

Thanks,
Keir

Best regards
Koldo

2009/6/29 Keir Mierle <mierle@xxxxxxxxx>

I implemented a primitive Levenberg-Marquardt and Powell's dogleg minimizer using Eigen:

http://code.google.com/p/libmv/source/browse/trunk#trunk/src/libmv/numeric

They work on simple problems but are not industrial strength. However, I would be happy to help with extending these if necessary.

Keir

On Mon, Jun 29, 2009 at 8:22 AM, Koldo Ramirez <koldo.ramirez@xxxxxxxxx> wrote:

Hello all

I would like to ask you how to focus non linear equations solving in Eigen. I mean:

1. Is there somebody working on it now?

2. Did you know libraries or code to do it that could match well with Eigen? (including similar license)
I have seen something about DASPK (for DAE solving)

3. The last solution I have is to port the solver I use that comes in a 90% from Numerical Recipes source. But I would prefer to include something with a LGPL like license.

Best regards
Koldo