[Insight-users] Parameter setting for the RegularStepGradientDescentOptimizer

[Lino Ramirez]

Hello,

I have been digging for some heuristics to set up the parameters of the
RegularStepGradientDescentOptimizer. So far, I have found some information
in this mailing list and the Software Guide (I posted some of the findings
in the ITK Wiki: http://www.itk.org/Wiki/ITK_Image_Registration). A
summary of the heuristics I have found are:

For the scale factors:
If the maximum expected absolute value for parameter "P" is about "v" then
one good initial estimate for the scale factor for that parameter is
1/(v*k)
With k being a suitable constant greater than 1. Luis Ibáñez usually
suggests k = 10

For the relaxation factor:
If metric is noisy then increase the relaxation factor

For the MaximumStepLength:
If the maximum expected absolute value for all the parameters being
optimized is about "vMax" then the MaximumStepLength should be set to a
value smaller than vMax/k
With k being a suitable constant greater than 1. Note that vMax must take
into account the scale factors.

For the MinimumStepLength:
If the registration stops too early then decrease the MinimumStepLength

I am trying to put these heuristics into practice in a sort of an
automated way (reducing user interaction), but I am not sure if the
heuristics are appropriate? Could someone tell me if the heuristics are
ok? Or could you suggest another set of heuristics? Moreover, I would
really appreciate some pointers to:

1.- defining noisy metric (I know I could do it if I see the plot, but I
would like to do it without user interaction)
2.- determining how much to increase the relaxation factor as a function
of the noise in the metric
3.- setting the constant k mentioned above
4- determining that the registration stops too early (again without user
interaction)

Thank you very much for any pointer

Sincerely,

Lino Ramirez

P.S. I will post on the WIKI any new information I can compile on the issue