Re: GSoC 2015: Optimization Package: Non-linear and constrained least squares lsqcurvefit, lsqlin, lsqnonlin

[Ben Abbott]

> On Apr 6, 2015, at 3:05 PM, Carnë Draug <address@hidden> wrote:
> 
> On 6 April 2015 at 19:57, Ben Abbott <address@hidden> wrote:
>>> On Feb 24, 2015, at 3:42 AM, Olaf Till <address@hidden> wrote:
>>> 
>>>> 
>>>> I am studying the Levenberg-Marquardt algorithm from [2].
>>> 
>>> This seems to be a general introduction to LM. An actual algorithm
>>> must also, among others, take measures to be numerically stable. It is
>>> usually best to start from one of the several already existing
>>> algorithms (I know this is your intention). In the optim package we
>>> have, among others, an SVD-based algorithm.
>> 
>> I came across an Levenberg-Marquardt implementation for Octave.
>> 
>>        https://sites.google.com/site/ulfgri/numerical/levmar
>> 
>> Perhaps it is useful.
>> 
> 
> leasqr [1] from optim package provides an implementation of
> Levenberg-Marquardt nonlinear least squares algorithm.
> 
> I have used it plenty of times as replacement for Matlab's nlinfit
> and get results which are comparable within machine precision.
> 
> Carnë
> 
> [1] http://octave.sourceforge.net/optim/function/leasqr.html

Admittedly, I've not used leasqr enough to be comfortable with all its 
capabilities.

For example, I hadn't realized that it handled constraints (linear, inequality, 
and box constraints are all supported).

Ben