commons-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Thomas Neidhart <thomas.neidh...@gmail.com>
Subject Re: [math] Problem using the LevenbergMarquardtOptimizer
Date Sun, 03 Feb 2013 20:18:24 GMT
On 02/02/2013 12:24 PM, Gilles wrote:
> Hello.
> 
>> Dear all,
>> I'm using the commons math library for solving least-squares problems
>> using the LevenbergMarquardtOptimizer.
>> I have a problem when I specify bounds for the estimation (by means
>> of a SimpleBounds object). I don't understand why, but the estimate
>> seems to not care about the bounds.
>> I attach to the mail a simple code which gives the problem. This code
>> is inspired to the one published on the web-site for explaining how
>> the LevenbergMarquardtOptimizer works...
>> The code at the end perform two estimates:
>> 1) the first one without bounds is correct;
>> 2) the second one, instead, should take care of the bounds, but I
>> obtain the same values as in the first case... :-(
> 
> Actually, the "LevenbergMarquardtOptimizer" does not support bounds,
> even though the API accepts that you pass them.
> In fact, all "options" can be passed to all optimizers but not all
> optimizers will use all of the options, if you see what I mean. :-)
> 
> I agree that this is confusing. [The issue was already raised but
> perhaps you could file a report on the project's bug-tracking system
> so that we can gather all opinions on how to best deal with it.
> I chose the current behaviour because it allows a much simpler API.
> One way would be to just document which options are ignored by a
> given optimizer implementation (and that would leave room to add
> support later).]
> 
> Currently, only the "CMAESOptimizer" implementation supports bounds.
> However, Commons Math provides function wrappers which you can use to
> approximate support for simple bounds:
>  
> http://commons.apache.org/math/apidocs/org/apache/commons/math3/optim/nonlinear/scalar/MultivariateFunctionMappingAdapter.html
> 
>  
> http://commons.apache.org/math/apidocs/org/apache/commons/math3/optim/nonlinear/scalar/MultivariateFunctionPenaltyAdapter.html
> 
> [But note the caveat indicated there (which might require that you use
> a "custom" convergence checker).]

Well, while preparing the slides for FOSDEM I used the new optimization
API myself, and I have to say, I do not like it:

 * it is confusing (what is actually going on, aka which optimizer
   supports which options)
 * non-intuitive (which options are available)

I do not have a better idea right away, but it is a clear example that
such changes need community support!

Thomas

---------------------------------------------------------------------
To unsubscribe, e-mail: user-unsubscribe@commons.apache.org
For additional commands, e-mail: user-help@commons.apache.org


Mime
View raw message