Re: [eigen] How far along is the sparse module?

[Thomas Capricelli <orzel@xxxxxxxxxxxxxxx>]

In data mercoledì 09 settembre 2009 11:17:58, bp ha scritto:
> I'm implementing the Levenberg-Marquardt algorithm using Eigen and
> I've come to the point where I need to optimize the central piece of
> the update:
> (see http://www.ics.forth.gr/~lourakis/levmar/levmar.pdf appendix for
> pseudo-code)
> 
>     Scalar mu;    //damping factor
>     MatrixNxM J; //Jacobian
>     MatrixNxN H = 2*J*J.transpose();//pseudo Hessian
>     static const MatrixNxN I = MatrixNxN::Identity();
> 
>     MatrixNxN H_i;
>     (H+mu*I).computeInverse(&H_i);// TODO: something more intelligent
> 
> I need to take advantage of the fact that H is sparse, symmetric and
> positive definite.

Hi,

You might have noticed that i'm working on porting (c)minpack to eigen, and 
this includes a Levenberg-Marquardt algorithm which works pretty well even in 
corner cases.

It is not yet available for sparse matrices/vectors, but i think it will be 
generic enough when finished to handle those.

see my previous mail "port of (c)minpack to eigen : status" for more 
information.

best wishes,
Thomas
-- 
Thomas Capricelli <orzel@xxxxxxxxxxxxxxx>
http://www.freehackers.org/thomas