Gaussian transformation turns variable to NAN when training BDT

Dear experts,

I am trying to train a BDT for signal to background discrimination and have recently started getting issues when adding the “number of jets” variable which can be either 2 or 3 in the case at hand. When trying to train the BDT I use the options “Transformations=I;D;P;G;D” when creating theTMVA factory. Something seems to go wrong however at the Gaussian transformation step, where the number of jets variable is always turned into “nan”, even though the original variable in the tree is always just either 2 or 3. When removing the gaussian transformation this problem disappears and the BDT can be trained without issues. Does anyone have an idea of what might be going wrong or what I am missing? Thanks in advance for the help!

(some output from TMVA that might help: )
Variable Mean RMS [ Min Max ]
: numberOfJets: 2.4977 0.49999 [ 2.0000 3.0000 ]

Preparing the Decorrelation transformation
: numberOfJets: 4.6591 1.0000 [ -6.9688 11.714 ]

Preparing the Principle Component (PCA) transformation
: numberOfJets: 7.7092 66.976 [ -588.72 764.58 ]

Preparing the Gaussian transformation…
Preparing the Decorrelation transformation…
: numberOfJets: -2.3513e+06 -nan [ -2.3513e+06 -2.3513e+06 ]




can you provide the data set and eventually also the macro reproducing this problem ?
This would be helpful in finding the cause and then solving this issue

Best Regards


Also, I would like to add that however it does not make much sense a Gaussian transformation for that variable. I would add that transformation only for the variables that need it. This is possible as explained at page 42 of the TMVA Users Guide