Re: nlinfit in octave

[Olaf Till]

On Fri, Jun 12, 2015 at 07:00:59PM -0700, AsmaA wrote:
> Hi Olaf,
> 
> I need some guidance on wrapping nlinfit on lsqcurvefit/lsqnonlin.
> 
> - Implementing RobustWgtFun, where weight functions can be specified to
> perform Robust regression 
> (weights computed from the specified function using the residual from every
> iteration) and a non-constant error model [1].
> 
> From what I have understood, the robust weight function/error model modifies
> the weights in every iteration and this feature has to be added in the
> backend. I'm not clear how Matlab does error parameter estimation for the
> error model and how this feature could be added.

Hi Asma,

the 'RobustWgtFun' feature is not in the algorithm of
'nonlin_residmin' and I currently haven't the time to try to add
it. Maybe later. I'm also not aware of a different algorithm in core
Octave or the optim package which implements a comparable feature. I
can only suggest to leave this feature not implemented in nlinfit for
now.

> -Setting options of nlinfit using statset instead of optimset or
> optimoptions [2]. The diarent functions to set similar options in nlinfit is
> because it is a part of statistics toolbox in Matlab.

Copy, and make the appropriate slight changes to, 'optimset.m',
'optimget.m', and '__all_opts__.m' from core Octave and rename them to
'statset.m', 'statget.m', and e.g. '__all_stat_opts__.m',
respectively. For now, these functions can probably be distributed
with the optim package. 'nlinfit' will have to include a line
e.g. like

## PKG_ADD: __all_stat_opts__ ("nlinfit");

I've done something similar in the database package, the respective
functions there are 'setdbopts', 'getdbopts', and '__all_db_opts__'.

Regards,

Olaf

-- 
public key id EAFE0591, e.g. on x-hkp://pool.sks-keyservers.net

signature.asc
Description: Digital signature