From octave-sources-request at bevo dot che dot wisc dot edu Tue Aug  7 05:11:47 2001
Subject: Re: Optimization: Replace NaN by Inf
From: Jeremy DENISE <jeremy dot denise at inria dot fr>
To: Gabriel Arcos <gabrielarcos at cantv dot net>
CC: etienne at isr dot ist dot utl dot pt, octave-sources@bevo.che.wisc.edu
Date: Tue, 07 Aug 2001 12:11:37 +0200

Gabriel Arcos wrote:
> 
> Suppose you have the problem max F(x), and suppose F(x) have derivative
> DF(x). Now, suppose DF(x) can be expressed as DF(x)=P(x)/Q(x). Suppose F
> have his maximun at x=x0, so DF(x0) vanish.

> ----- Original Message -----
> From: Etienne Grossmann <etienne at anonimo dot isr dot ist dot utl dot pt>
> To: <help-octave at bevo dot che dot wisc dot edu>; <octave-sources@bevo.che.wisc.edu>
> Cc: <etienne at isr dot ist dot utl dot pt>; <jeremy.denise@inria.fr>
> Sent: Monday, August 06, 2001 1:14 PM
> Subject: Optimization: Replace NaN by Inf
> 
> >
> >   Hello,
> >
> >   when optimizing a function 'f', I am tempted to replace any NaNs
> > that it returns by Infs. This way, the returned value can be compared
> > w/ other values, and the optimization algorithm is happy.
> >
> >   Can anyone think of a reason not to replace NaNs by Infs inside an
> > optimization algorithm?
> >
> >   If not, I'll have my various optimization functions do the
> > replacement.
> >
> >   Etienne

Hi,

Well, I think something has to be cleared: this replacement would only
have to take place in the Nelder-Mead algorithm, where only the _values_
of the function to be minimized are used, not the first nor second
derivatives. This kind of trick is often used to solve a constrained
problem with algorithms primarily designed for unconstrained ones.

J