| optimization using IDL [message #7002] |
Mon, 16 September 1996 00:00  |
Karl Young
Messages: 31
Registered: April 1996
|
Member |
|
|
I know this has been discussed before and I apologize for any
redunancy. I am finding IDL's curvefit inadequate for the types
of optimization I need to do. For one thing I need to do
constrained optimization. I hacked curvefit to do a kludgy
version of constrained optimization and a coleague put
together an even better version but I'm at the point at which
I really need an "industrial strength" constrained optimizer.
(I went through the section in Bevington's book from which
IDL's curvefit was culled and there didn't seem to be any really
obvious ways to extend that algorithm to do the kind of
constrained optimization that you find in modern textbooks on
optimization, i.e. I generally couldn't find any way of
extending standard Marquardt-Levinson type optimization, but
maybe I'm reading the wrong books)
So my question is; has anybody been in this situation with
IDL and if so how did they deal with it ? E.g. has anybody
written more sophisticated optimization code in IDL, do
people generally call external Fortran or C routines from
commercial or shareware packages, ... ? Any comments or
suggestions greatly apreciated,
-- Karl Young karl@gojira.berkeley.edu
|
|
|
|
| Re: optimization using IDL [message #7078 is a reply to message #7002] |
Fri, 20 September 1996 00:00  |
PREUSSER
Messages: 8
Registered: March 1993
|
Junior Member |
|
|
In article <323DE338.41C67EA6@gojira.berkeley.edu>, Karl Young <karl@gojira.berkeley.edu> wrote:
> I know this has been discussed before and I apologize for any
> redunancy. I am finding IDL's curvefit inadequate for the types
> of optimization I need to do. For one thing I need to do
> constrained optimization. I hacked curvefit to do a kludgy
> version of constrained optimization and a coleague put
> together an even better version but I'm at the point at which
> I really need an "industrial strength" constrained optimizer.
> (I went through the section in Bevington's book from which
> IDL's curvefit was culled and there didn't seem to be any really
> obvious ways to extend that algorithm to do the kind of
> constrained optimization that you find in modern textbooks on
> optimization, i.e. I generally couldn't find any way of
> extending standard Marquardt-Levinson type optimization, but
> maybe I'm reading the wrong books)
>
> So my question is; has anybody been in this situation with
> IDL and if so how did they deal with it ? E.g. has anybody
> written more sophisticated optimization code in IDL, do
> people generally call external Fortran or C routines from
> commercial or shareware packages, ... ? Any comments or
> suggestions greatly apreciated,
>
> -- Karl Young karl@gojira.berkeley.edu
Karl,
I have to make two comments on your request:
1) I have developped a curve fitting program running with PV-Wave
Advantage that allows generating starting values close to
the desired solution, so that often you do not need to apply
special constraints. At the moment, it is limited to fit with
a sequence of Gaussians and a quadratic polynomial. But you
are able to keep any parameter constant, that is, force it to
a certain value. The called solver is that of PV-Wave Advantage,
the Levenberg-Marquardt from Minpack. Anybody interested can
download the latest version from the following address:
http://www.fhi-berlin.mpg.de/~grz/pub/preusser.html
2) If you still need to use a solver with constraints of "industrial
strength", I recommend you to contact
WTI@aol.com or http://members.aol.com/WTI
They are a company specializing in constrained optimization, and
offer industrial solution at resonable prices. It is a group
split from IMSL.
Hope that will help you and others interested in this topic.
--
Dr. Albrecht Preusser | Basic Engineering:
Gemeinsames Rechenzentrum | ----------------------
Fritz-Haber-Institut der MPG | Nothing for nothing
Faradayweg 4-6 | Nothing for everything
D-14195 Berlin (Dahlem) | Nothing for ever
Phone: +49-30-8413-3220 | ----------------------
http://www.fhi-berlin.mpg.de/~grz/pub/preusser.html
|
|
|
|
| Re: optimization using IDL [message #7096 is a reply to message #7002] |
Tue, 17 September 1996 00:00 |
Amara Graps
Messages: 24
Registered: June 1996
|
Junior Member |
|
|
Karl Young wrote:
>
> I know this has been discussed before and I apologize for any
> redunancy. I am finding IDL's curvefit inadequate for the types
> of optimization I need to do. For one thing I need to do
> constrained optimization. I hacked curvefit to do a kludgy
> version of constrained optimization and a coleague put
> together an even better version but I'm at the point at which
> I really need an "industrial strength" constrained optimizer.
> (I went through the section in Bevington's book from which
> IDL's curvefit was culled and there didn't seem to be any really
> obvious ways to extend that algorithm to do the kind of
> constrained optimization that you find in modern textbooks on
> optimization, i.e. I generally couldn't find any way of
> extending standard Marquardt-Levinson type optimization, but
> maybe I'm reading the wrong books)
Hi Karl,
Do you want to stick with the Marquardt algorithm or try another
algorithm? I'm aware of more sophisticated optimization techniques
(but have never needed them, and would only be able to give
references).
If you want to stick with the Marquardt algorithm, another thing to
try is to work with the lambda parameter. The Bevington/Numerical
Recipes/IDL routine uses a lambda that is the same value for all of
the free parameters. Make lambda an _array_ instead, with a slightly
different value for each of the free parameters.
If you are navigating a tricky Chi-square "terrain" where particular
parameters give really flat chi-squares and therefore, the iterations
go on forever, this technique helps one get out of that problem. One
would have to have a pretty good understanding of their function that
they are trying to fit, however.
For example, one could do something like the following:
;percentage of lambda to increase/decrease
del_flambda = [30.d0,30.d0,30.d0,30.d0,25.d0,25.d0]
And so inside of the CURVFIT routine, the flambda looks like:
flambda = flambda/del_flambda
You can also impose your your own min and max inside the CURVFIT
routine.
For example, let's say you have 6 free parameters v(0..5). You could
initially set up the minimums/maximums for them:
;Min/max values allowed for v()
minmax(0,0)=[v(0)/3,v(1)-3*cf,v(2)/3,.0015,v(4),0] ;min
minmax(0,1)=[v(0)*3,v(1)+3*cf,v(2)*3,.0027,v(4),0] ;max
mmn=2 ;Least #iterations to impose temporary minmax values
mmx=100 ;Max #iterations to impose temporary minmax values
(This is just an example of min/max's I've used for a curvefitting
problem a few years ago.)
Then use these inside of CURVFIT in the iterations loop, checking to
see if the value of the parameters have exceeded the min/max's.
The Marquardt curvefitting is also extremely sensitive to initial
guesses. That's another place where one's knowledge of their
curvefitting function is necessary. If one doesn't have good
initial guesses, they're in trouble. I've recently thought that
a genetic algorithm method might work well for coming up with
good initial guesses, but have never implemented anything like
that.
Are your derivatives accurate? That is a really easy place for
one to make a mistake, and will make the Marquardt algorithm
completely useless.
So I don't know if the above techniques are is the "industrial
strength" Marquardt curvefitting, that you were looking for, (and
I'm not even sure if these tricks make sense mathematically) but
these tricks worked pretty well for me fitting some complicated
functions that consisted of integrals, numerical derivatives, and
with up to 7 free parameters.
Amara
--
************************************************************ *****
Amara Graps amara@quake.stanford.edu
Solar Oscillation Investigations Stanford University
http://quake.stanford.edu/~amara/amara.html
************************************************************ *****
"Never fight an inanimate object." - P. J. O'Rourke
|
|
|
|
| Re: optimization using IDL [message #7098 is a reply to message #7002] |
Tue, 17 September 1996 00:00 |
rivers
Messages: 228
Registered: March 1991
|
Senior Member |
|
|
In article <323DE338.41C67EA6@gojira.berkeley.edu>, Karl Young <karl@gojira.berkeley.edu> writes:
> I know this has been discussed before and I apologize for any
> redunancy. I am finding IDL's curvefit inadequate for the types
> of optimization I need to do. For one thing I need to do
> constrained optimization.
...
> So my question is; has anybody been in this situation with
> IDL and if so how did they deal with it ? E.g. has anybody
> written more sophisticated optimization code in IDL, do
> people generally call external Fortran or C routines from
> commercial or shareware packages, ... ? Any comments or
> suggestions greatly apreciated,
I have worked with RSI to fix some of the limitations in curvefit (e.g. it now
longer needs analytical derivatives). For more sophisticated optimization I
have used CALL_EXTERNAL and the IMSL or NAG libraries.
____________________________________________________________
Mark Rivers (312) 702-2279 (office)
CARS (312) 702-9951 (secretary)
Univ. of Chicago (312) 702-5454 (FAX)
5640 S. Ellis Ave. (708) 922-0499 (home)
Chicago, IL 60637 rivers@cars3.uchicago.edu (Internet)
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
