I'm using the SVM Learner node, RBF kernel, and I'm setting gamma to various values. The resulting trained SVM has a gamma that differs from what I set it to. Perhaps I'm misunderstanding SVM fitting, but I believe that gamma should remain fixed at the value I specify. (It does this in R's e1071's svm, for example.)
As an example, I specified a gamma of 0.9, and the resulting gamma is 0.61728.... That was the result that was closest to the specified value -- other results were farther off.
This is with an SVM that is being trained on a training set of 10,000 rows -- 17 fields -- from a dataset with 40 million values. The target is a string that takes on two values, "0" or "1".
I am not sure how do you set the gamma parameter value (on the dialog for the RBF kernel there is only sigma), though the computed gamma value for the RBF model should always follow (you can set sigma, but not gamma for RBF): As I see the current implementation uses a different sign for gamma, but it seems to agree with the PMML specification. The value 0.617284... seems to agree to the formula of gamma = 1/(2*0.9*0.9), so I do not see a problem there either. Maybe an option to being able to specify this parameter in both form (or show the other form during configuration) would make things easier to understand (also explaining the kernel parameters for the different kernels would also help in my opinion).
(Accidental double post.)