Author 
Message 
John Herman science forum beginner
Joined: 11 Jul 2005
Posts: 28

Posted: Sat Jul 15, 2006 2:33 pm Post subject:
Optimization of noisy functions



I'm trying to sort out a model matching problem with noisy data using
optimization . I've been looking at several methods. I went out to the
decision tree on plato and looked at what is available. The most promising is
probably SNOBFit but I'm limited to the C prrogramming language (f2c is
possible) by the platform. Peter says NEWUOA really isn't useful for
optimizing with noisy data. Implicit Filtering looks sort of like Truncated
Newton with numerical derivatives. I'm at a loss for where to go from here.
The other possibility is to move to curve fitting. 

Back to top 


Peter Spellucci science forum Guru
Joined: 29 Apr 2005
Posts: 702

Posted: Mon Jul 17, 2006 3:24 pm Post subject:
Re: Optimization of noisy functions



In article <T07ug.28362$uy3.9276@tornado.socal.rr.com>,
John_W_Herman@yahoo.com (John Herman) writes:
Quote:  I'm trying to sort out a model matching problem with noisy data using
optimization . I've been looking at several methods. I went out to the
decision tree on plato and looked at what is available. The most promising is
probably SNOBFit but I'm limited to the C prrogramming language (f2c is
possible) by the platform. Peter says NEWUOA really isn't useful for
optimizing with noisy data. Implicit Filtering looks sort of like Truncated
Newton with numerical derivatives. I'm at a loss for where to go from here.
The other possibility is to move to curve fitting.

you want nonlinear least squares fitting? why not LEVMAR or gaussfit ?
if derivatives (of the _model_!) should be available. I understand your
problem now as "fit model to noisy _data_ , not a noisy function evaluation?
Newuoa uses _interpolation_ of the function values, (for you?
the sum of squared deviations between model and data) hence, if the function
evaluation itself is noisy, this will promote this noise and possibly produce
lots of spurious local minima.
but if the function evaluation is noise free , it is a viable approach to
derivative free minimization.
hth
peter 

Back to top 


John Herman science forum beginner
Joined: 11 Jul 2005
Posts: 28

Posted: Fri Jul 21, 2006 12:13 pm Post subject:
Re: Optimization of noisy functions



Thank you, Peter.
In article <e9ga3t$frk$1@fb04373.mathematik.tudarmstadt.de>,
spellucci@fb04373.mathematik.tudarmstadt.de (Peter Spellucci) wrote:
Quote: 
In article <T07ug.28362$uy3.9276@tornado.socal.rr.com>,
John_W_Herman@yahoo.com (John Herman) writes:
I'm trying to sort out a model matching problem with noisy data using
optimization . I've been looking at several methods. I went out to the
decision tree on plato and looked at what is available. The most promising
is
probably SNOBFit but I'm limited to the C prrogramming language (f2c is
possible) by the platform. Peter says NEWUOA really isn't useful for
optimizing with noisy data. Implicit Filtering looks sort of like Truncated
Newton with numerical derivatives. I'm at a loss for where to go from here.
The other possibility is to move to curve fitting.
you want nonlinear least squares fitting? why not LEVMAR or gaussfit ?
if derivatives (of the _model_!) should be available. I understand your
problem now as "fit model to noisy _data_ , not a noisy function evaluation?
Newuoa uses _interpolation_ of the function values, (for you?
the sum of squared deviations between model and data) hence, if the function
evaluation itself is noisy, this will promote this noise and possibly produce
lots of spurious local minima.
but if the function evaluation is noise free , it is a viable approach to
derivative free minimization.
hth
peter 


Back to top 


Google


Back to top 



The time now is Thu Apr 26, 2018 5:47 am  All times are GMT

