[Insight-users] Re: [Insight-developers] a small bug initkConjugateGradientOptimizer

Karthik Krishnan Karthik.Krishnan at kitware.com
Wed Jun 22 17:08:29 EDT 2005



Einstein, Daniel R wrote:

> Anish,
>  
> As far as I can tell, all of the optimization algorithms from Netlib 
> are local. Global optimization is considerably harder, and requires 
> much more crunching. Examples are simulated annealing, multi-starts, 
> i.e. stochastically sampling the solutions space, particle swarm 
> methods, and sequential response surfaces. I am new enough to ITK that 
> I cannot say which if any of these might usefully be implemented in ITK.

http://itk.org/Insight/Doxygen/html/classitk_1_1SPSAOptimizer.html was 
recently added

> Particle swarm methods are an interesting option because they are so 
> easy to program. Be advised, however, there are no global methods that 
> guarantee conversion.
>  
> Dan
>  
>
> Daniel R Einstein, PhD
> Biological Monitoring and Modeling
> Pacific Northwest National Laboratory
> P.O. Box 999; MSIN P7-59
> Richland, WA 99352
> Tel: 509/ 376-2924
> Fax: 509/376-9064
> _daniel.einstein at pnl.gov_ <mailto:daniel.einstein at pnl.gov>
>
>  
>
> ------------------------------------------------------------------------
> *From:* insight-users-bounces+daniel.einstein=pnl.gov at itk.org 
> [mailto:insight-users-bounces+daniel.einstein=pnl.gov at itk.org] *On 
> Behalf Of *Ashish Poddar
> *Sent:* Wednesday, June 22, 2005 10:58 AM
> *To:* Luis Ibanez
> *Cc:* insight-users @ itk. org
> *Subject:* [Insight-users] Re: [Insight-developers] a small bug 
> initkConjugateGradientOptimizer
>
> Hi,
>  
> Most of the optimizers which I have came across help in finding the 
> local minima around the given initial approximation. But in that case 
> I mostly end up in a wrong place. Is there any algorithm which helps 
> to scan the global space somehow and help in determining the global 
> minima.
>  
> Other problem that I am facing with the conjugate gradient method is 
> that the scales do not work for conjugate gradient optimizer. The 
> scales are only taken into account in the very first initialization 
> step and are never considered again in any of the latter iterations. I 
> want to fix some of the parameters by setting the scale to some 
> extreme value (I used to set 100 or something for regular step 
> gradient descent optimizer and it used to serve the purpose very 
> conveniently).
>  
> any help will be highly appreciated,
> with regards,
> Ashish.
>
>  
> On 5/18/05, *Luis Ibanez* <luis.ibanez at kitware.com 
> <mailto:luis.ibanez at kitware.com>> wrote:
>
>
>     Hi Ashish,
>
>     The Conjugate Gradient method is only convenient when the cost
>     function has smooth second derivatives.  If your cost function
>     is noisy, is is unlikely that this optimizer will behave nicely.
>
>     Note that is is common to find that Image Metrics are rather
>     noisy functions.
>
>
>
>        Regards,
>
>
>           Luis
>
>
>
>     --------------------
>
>     Ashish Poddar wrote:
>
>     > Hi,
>     >
>     > I also am right now struggling with the initialization options
>     for the
>     > Conjugate Gradient for which I could not find any examples. While
>     > searching I came across an example for Levenberg Marquardt Optimizer
>     > which seems to be having similar interface as that of conjugate
>     > gradient optimizer. However the similar initialization did not
>     worked
>     > for Conjugate gradient. If someone can point out any reference for
>     > Conjugate Gradient, it would be great.
>     >
>     > Earlier I was using regular step gradient descent optimizer with
>     these
>     > parameters:
>     > Transform - Centered Affine
>     > Scale for first 9 parameters - 1.0
>     > Scale for next 6 parameters - 0.0001
>     > Number of Iterations - 400
>     > Minimum Step Length - 0.0001
>     > Maximum Step Length - 0.005
>     >
>     > Any help will be highly appreciated,
>     > with regards,
>     > Ashish.
>     >
>     >
>     >
>     > On 5/17/05, Ashish Poddar <ahpoddar at gmail.com
>     <mailto:ahpoddar at gmail.com>> wrote:
>     >
>     >>Hi Luis,
>     >>
>     >>Thank you for the quick action. probably similar change is required
>     >>for Levenberg Marquardt Optimizer too.
>     >>
>     >>with regards,
>     >>Ashish.
>     >>
>     >>On 5/16/05, Luis Ibanez <luis.ibanez at kitware.com
>     <mailto:luis.ibanez at kitware.com>> wrote:
>     >>
>     >>>Hi Ashish,
>     >>>
>     >>>Thanks for pointing this out.
>     >>>
>     >>>You are right, the GetValue() method should be const.
>     >>>
>     >>>A fix has now been committed to the CVS repository.
>     >>>
>     >>>Please let us know if you encounter any other problem.
>     >>>
>     >>>    Thanks
>     >>>
>     >>>       Luis
>     >>>
>     >>>----------------------
>     >>>Ashish Poddar wrote:
>     >>>
>     >>>>hi,
>     >>>>
>     >>>>I am not sure whether it qualifies as a bug or not, but surely
>     affects
>     >>>>the re-usability and pluggability model of ITK Library.
>     >>>>
>     >>>>the GetValue() function in ConjugateGradientOptimizer class
>     currently is
>     >>>>
>     >>>>MeasureType GetValue();
>     >>>>
>     >>>>but in case of RegularStepGradientDescentOptimizer class its
>     defined by macro as
>     >>>>
>     >>>>MeasureType GetValue() const;
>     >>>>
>     >>>>which is an interface mis-match... This I encountered when i
>     replaced
>     >>>>regular step gradient descent optimizer by conjugate gradient
>     >>>>optimizer. In the observer I was using a const reference of the
>     >>>>optimizer and displaying the value (just the example which is
>     >>>>available for the same nothing new =D)...
>     >>>>
>     >>>>with regards,
>     >>>>Ashish.
>     >>>>
>     >>>
>     >>>
>     >>--
>     >>Ashish Poddar
>     >>Have an acceptable reason for accepting anything.
>     >>Y:ashish_poddar | MSN:ashish_poddar at yahoo.com
>     <mailto:MSN:ashish_poddar at yahoo.com>
>     >>
>     >
>     >
>     >
>
>
>
>
>
>
> -- 
> Ashish Poddar
> Have an acceptable reason for accepting anything.
> Y:ashish_poddar | MSN:ashish_poddar at yahoo.com 
> <mailto:MSN:ashish_poddar at yahoo.com>
>
>------------------------------------------------------------------------
>
>_______________________________________________
>Insight-users mailing list
>Insight-users at itk.org
>http://www.itk.org/mailman/listinfo/insight-users
>  
>


More information about the Insight-users mailing list