[Insight-users] Some questions regarding the lbfgsb optimizer

[Erik Türke]

Erik Türke wrote:
> Hi!
>
> At the moment i am dealing with the LBFGSB-Optimizer.
>
> I kinda understand how it works as a quasi-newton-optimizer, but i 
> have some question regarding the parameters which can be set.
>
> - SetCostFunctionConvergenceFactor:
>
> This defines when the algorithm terminates.
> It is said that the algorithm terminates, " when the reduction in cost 
> function is less than factor * epsmch"
> Is the reduction of the cost function referred to 1 iteration? If the 
> difference between the value of the cost function of iteration k and 
> k+1 is less
> factor * epsmch then algorithm terminates?
> And what exactly is meant by "mechine precision" ?
>
> - SetProjectedGradientTolerance:
> What exactly is the difference between ProjectedGradientTolerance and 
> GradientMagnitudeTolerance, used in the RegularStepGradientOptimizer
>
> - SetMaximumNumberOfEvaluations:
> What evaluation is meant by this parameter?
>
> - SetMaximumNumberOfCorrections:
> What correction is meant by this parameter?
>
> And what is the use of the boundSelect, lowerBound and upperBound 
> parameters? Following the examples, i always set them to 0.
>
> And how is the step length defined? Is it set to 1? I am asking this, 
> because regarding the nonlinear optimization problem, i did not find 
> any strategy of choosing a appropriate step length.
>
> Sorry for that many questions, but i have not found the code for the 
> algorithm itself.
>
> Many thanks!
>
> Regards, Erik!
>
> _____________________________________
> Powered by www.kitware.com
>
> Visit other Kitware open-source projects at
> http://www.kitware.com/opensource/opensource.html
>
> Kitware offers ITK Training Courses, for more information visit:
> http://www.kitware.com/products/protraining.html
>
> Please keep messages on-topic and check the ITK FAQ at:
> http://www.itk.org/Wiki/ITK_FAQ
>
> Follow this link to subscribe/unsubscribe:
> http://www.itk.org/mailman/listinfo/insight-users
Hi!

I have been working on getting informations regarding my questions.

At least i have found out what the bound parameters of the optimizer do 
and what is meant by epsmch.
And is it right, that the correction and evaluation parameters have 
something to do with the approximation of the hessian matrix?

Thanks for your help!

Regards, Erik!