[ITK-users] L-BFGS optimizer's in versor space

[zoro]

Dear all,

I tried to implement L-BFGS optimizer in the versor space and registration
framework with the algorithm as below:

http://en.wikipedia.org/wiki/Limited-memory_BFGS

L-BFGS produced the direction and base on that, I used the line-search to
update parameters ( number of parameters is 6, anyway). But sometimes,
direction goes to irregular places where translation direction is extremely
large.  And the convergence is not good. It does not show the effect of BFGS
but only line-search.

I thought that with constraint of versor space (in unit cube surface), it's
not straight as the algorithm above. For example, in versor space (vec_a -
vec_b) actually equal to composition of  vec_a and Conjugate of vec_b.
Calculating (vec_a - vec_b) in simple way (in the algorithm above) leads to
problem of irregular jumping in parameters space.

Is anyone who deals with L-BFGS in versor space? 
I will appreciate if you give me some advice.

Thank you!
Regards,




--
View this message in context: http://itk-insight-users.2283740.n2.nabble.com/L-BFGS-optimizer-s-in-versor-space-tp7585916.html
Sent from the ITK Insight Users mailing list archive at Nabble.com.