Re: fmincon options playing

In article <ef2ff48.-1@xxxxxxxxxxxxxxxx>, "Max Touzel" <mtouzel@xxxxxxxxxxx> wrote:

Hi,
I'm trying to make my fmincon optimization more efficient by playing
around with the options, ie. seeing how loose I can make them and
still get a good answer. There are 5 things I am playing with
TolFun,
TolCon,
MaxIter,
MaxFunEvals,
and
the number of times I loop the entire routine (I have the
optimization in a loop where at the end I take the minimum over all
the iteration results).

Is there any convention for doing this. I want high accuracy, so I've
set my tolerances around 10^-15 to 10^-17. MaxIter and MaxFunEvals
are at 10^5. Then there's the loop. Are any of these values
ridiculous? Are they reasonable? I guess what might be most helpful
is a short breakdown of how these options work.
I assume that tolerances definitely increase the processing time
(more than the rest of the options?)
The loop seems like a fast way to go? (i.e. many loops)
Which one (tolerances or loops) would be the better one to make the
dominant part of the routine?
I don't know how MaxIter and MaxFunEvals affect the processing time.

again, is there any method to finding the best set of options? What
should I be looking for? e.g. consistent results and quick processing
time?


Sorry. But it really depends on your problem. If your
parameters are on the order of 1e-15, setting TolX
to 1e-15 may be a poor choice.

Likewise, if you have 100000 parameters to optimize,
then MaxFunEvals will need to be large, especially if
you let it use finite differences to generate the
gradients.

I have a chapter in my tips and tricks doc on some of
these issues, where I show what would happen as you
change some of these parameters.

John


--
The best material model of a cat is another, or preferably the same, cat.
A. Rosenblueth, Philosophy of Science, 1945

Those who can't laugh at themselves leave the job to others.
Anonymous
.

Relevant Pages

  • Disregard last msg(but check this one)
    ... I'm trying to make my fmincon optimization more efficient by playing ... optimization in a loop where at the end I take the minimum over all ... I don't know how MaxIter and MaxFunEvals affect the processing time. ...
    (comp.soft-sys.matlab)
  • fmincon options playing
    ... I'm trying to make my fmincon optimization more efficient by playing ... optimization in a loop where at the end I take the minimum over all ... I don't know how MaxIter and MaxFunEvals affect the processing time. ...
    (comp.soft-sys.matlab)
  • Re: looper pedal
    ... the audience, that the loop better be created live, otherwise its just ... guitar Karaoke. ... function, but no bpm readout), but for live playing it's unnecessary, ... two footswitch JamMan and it works fine. ...
    (rec.music.makers.guitar.jazz)
  • Re: looper pedal
    ... the audience, that the loop better be created live, otherwise its just ... guitar Karaoke. ... function, but no bpm readout), but for live playing it's unnecessary, ... two footswitch JamMan and it works fine. ...
    (rec.music.makers.guitar.jazz)
  • Re: looper pedal
    ... sequence of changes within a tune. ... the audience, that the loop better be created live, otherwise its just ... guitar Karaoke. ... function, but no bpm readout), but for live playing it's unnecessary, ...
    (rec.music.makers.guitar.jazz)