Mathematica 9 is now available
Services & Resources / Wolfram Forums
-----
 /
MathGroup Archive
1999
*January
*February
*March
*April
*May
*June
*July
*August
*September
*October
*November
*December
*Archive Index
*Ask about this page
*Print this page
*Give us feedback
*Sign up for the Wolfram Insider

MathGroup Archive 1999

[Date Index] [Thread Index] [Author Index]

Search the Archive

Re: NonlinearFit

  • To: mathgroup at smc.vnet.net
  • Subject: [mg19700] Re: NonlinearFit
  • From: peter.buttgereit at pironet.de (Peter Buttgereit)
  • Date: Sat, 11 Sep 1999 16:36:05 -0400
  • Organization: communication works
  • References: <7r24q2$6qd@smc.vnet.net>
  • Sender: owner-wri-mathgroup at wolfram.com

[This followup was posted to comp.soft-sys.math.mathematica and a copy 
was sent to the cited author.]

In article <7r24q2$6qd at smc.vnet.net>, virgil.stokes at neuro.ki.se says...
> In reference to Statistics `NonlinearFit` (vers. 1.4, author: John M. Novak)
> for Mathematica Version 3.0.
> 
> I would like to be able to replace the minimization criterion
> that is currently in this package (least-squares). I am trying to
> find a different type of minimization criterion -- one that
> performs better (hopefully) for my particular problem.
> 
> Is it possible to use this package such that one can override the
> current least-squares criterion with a user defined criterion?
> 
> -- Virgil
> 
> 
> 
Dear Virgil

  As like as William I would be curious about Your ideas other

than least-squares optimisation...

  The L-M-Algorithm is essentially least-squares based -- there

is no way to change this criterion.

Also, using absolute distances (or whatever) will complicate

confidence interval estimation considerably.

  If You are unhappy with least-squares I suppose You have noisy

data or some problems with "outliers".

In this case You can either "clean up" Your data in advance of

fitting or use a suitable transformation.

I.e. using least-squares fitting on the logarithms of Your data

will minimize over the sum of squares of logarithms of the 

deviations -- noisiness and outlier problem are somewhat 

reduced if the data model allows for the transformation.

   Further, You can define Your own criterion this way:

Fit "the model" using L-M (i.e.)

Define a criterion to be minimised/maximised based on the residuals

(sum, sum of squares, whatever suitable) 

and then use FindMinium[] to optimise over Your criterion.

For me this method was a affordable solution occasionally ( must

have been posted in this group some years ago).

KR,

Peter


  • Prev by Date: Re: Finding a relative prime
  • Next by Date: Manipulating a list of pairs
  • Previous by thread: Re: NonlinearFit
  • Next by thread: wrong bounding box