[Insight-users] Optimizer and R+* Cost functions

[Luis Ibanez]

Hi Julien,


   You are correct in that most ITK optimzers
   assume that the domain is R. (all real numbers)


   Your options are:


   A) Use the LBFGSB optimizer, which provides bounds
      for the parameters of the optimization

   or

   B) In your cost function truncate or fold the domain


   or


   C) Create a variation of your current cost function,
      in which R- is supported but results in high penalties.


   or


   D) Customize your own optimizer.



     Regards,



         Luis


----------------------------
Julien Michel wrote:
> Dear ITK users,
> 
> I am trying to use ITK optimizers (mainly GradientDescentOptimizer) to 
> optimize my own cost function. unfortunately, it is only defined on R+*, 
> and the optimizer happens to end sometimes asking for negative or null 
> values. I made my cost function return zero in such cases, but then the 
> derivative is null (function is constant on R-) and the optimizer does 
> not evolve anymore ... Is there a way to restrain the optimizer search 
> domain ?
> 
> Regards,
> 
> Julien