[Insight-users] LBFGS vs LBFGSB

[Martin Urschler]

hi,

can anybody explain me the difference between the two limited memory 
BFGS optimization methods which are implemented in the toolkit?
itkLBFGSOptimizer
itkLBFGSBOptimizer

I understand that the LBFGSB version lets the user specify constraints 
on the unknowns, however the deformable bspline registration example 
DeformableRegistration7 argues that "the LBFGSB algorithm is more 
appropriate for performing optimization in a parametric spaces of higher 
dimensions"
I don't see why this is true, IMHO both algorithms are useful for 
parameter spaces of higher dimensions, since both should be implemented 
in a "limited-memory" way (that's what the L is for, right?)
So, how to choose between them? What I find especially interesting in 
DeformableRegistration7 is that the bounds for LBFGSB are set up in a 
way that they will be ignored during the optimization, therefore 
essentially reducing LBFGSB to an LBFGS (with some additional storage 
requirement due to the more complicated implementation). I'm a little 
confused with this issue.

regards,
Martin

-- 
Martin Urschler
TU Graz - Institute for Computer Graphics & Vision
EMail: martin at urschler.info  ICQ: 41586273
Web: http://www.urschler.info
-----------------------------------------------------------
"Wehre nichts ab, was zu dir kommt, und halte nichts fest, was von dir 
geht."