[Insight-users] Criterion to stop 1+1_evolutionary optimizer?

[Li, George (NIH/NCI)]

Stephen,

Thanks for the input. I am using Mattes method
As it is a smoother metric, as you said. I have
Made an option in my program for users to choose 
Either the evolutionary optimizer or the regular
Step optimizer. Just for sake of code integrity, 
I would like to have both done in the same fashion. 

On the other hand, I would think that the multi-
Resolution strategy is not completely parallel 
To what the evolutionary Optimizer does. The 
Former is built mostly from performance concern
And avoiding local extremes may be a by-product,
While latter is one of the best optimizers in 
Terms of avoiding local traps and tolerating 
Noises. Therefore, it would be still beneficial 
To have the combined fitting approach.

Actually, I was thinking to combine these two
Optimizers and using them in tandem, in order to
Achieve some performance gain. But, the multi-
Resolution is a better one.

Regards,

George



-----Original Message-----
From: Stephen R. Aylward [mailto:aylward at unc.edu] 
Sent: Wednesday, March 16, 2005 6:23 PM
To: Li, George (NIH/NCI)
Cc: 'Luis Ibanez'; 'Insight-users at itk.org'
Subject: Re: [Insight-users] Criterion to stop 1+1_evolutionary optimizer?


Instead of adjusting the factors of the 1+1 optimizer for each level in 
the pyramid, I suggest adjusting the scalings of the parameters being 
optimized at each level.

Also, you might want to consider Mattes MI method with a gradient 
strategy instead of 1+1 when doing multi-scale optimization.  The 
justification is as follows:
* Mattes method provides effective gradients as it is a smooth metric. 
See the illustrations in the softwareGuide.
* The 1+1 optimizer is a pseudo-random search that helps avoid local 
extremes by not using gradient info.  A multi-scale approach is also 
used to avoid local extremes by essentially blurring the 
image/metric/gradient space.  So, most of the time you don't need to do 
both.   Specifically, consider using the Powel/Brent conjugate gradient 
optimizer with multi-scale registration using a Mattes MI metric.   See 
the imageRegTool application in the 
InsightApplications/LandmarkBasedMutualInformationRegistrationOrSomeSimilarL
ongName
directory (I really hate that directory name...)

Stephen

Li, George (NIH/NCI) wrote:
> Hi, Luis and all itk users:
> 
> I have found the multi-resolution samples. So, I am
> Now upgrading my registration program for a better and
> More attractive performance.
> 
> For OnePlusOneOptimizer, The growth factor seems to be
> A suitable variable to control the fitting speed for 
> Each pyramid. However, how to set a stopping criterion 
> Seems unclear to me.
> 
> Anyone has experience or idea on this matter?
> 
> Thanks,
> 
> George
> 
> 
> 
> -----Original Message-----
> From: Li, George (NIH/NCI)
> Sent: Tuesday, March 15, 2005 12:28 PM
> To: Li, George (NIH/NCI); 'Luis Ibanez'
> Cc: Insight-users at itk.org
> Subject: Registration performance
> 
> 
> Hi, Luis:
> 
> Sometime ago, you commented on improving performance
> As the following. I now understand what you meant by Multi-resolution. 
> However, is there any sample code For its implementation?
> 
> 
>>Trivial answer to your question about performance:
>>
>>            The way to improve performance
>>             is to use multi-resolution.
>>
>>You can register volumes of size 200x200x200 pixels
>>in about 20 seconds when using 3 levels of a multi- resolution 
>>pyramid, by subsampling by 2 at each level, in a typical Pentium 4 
>>machine at 2Ghz, and 512Mb of memory.
> 
> 
> Now, it seems that the registration performance has
> Become a big issue to me, and some 20 seconds for a
> Registration is much attractive, comparing with a few 
> minutes.
> 
> So, could you further provide some guidance on it?
> 
> Thanks,
> 
> George
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org 
> http://www.itk.org/mailman/listinfo/insight-users

-- 
===========================================================
Dr. Stephen R. Aylward
Associate Professor of Radiology
Adjunct Associate Professor of Computer Science and Surgery
http://caddlab.rad.unc.edu aylward at unc.edu
(919) 966-9695