Slides here. Preliminary conclusion is that I can cut background to 1% for signal cut to ~25%. Not too good.

So, to answer the question I asked myself last time: (“is the reported 1% background the fraction (background after/ background before) or (background after/ total after)?” )

Answer is that it’s background after/background before, so at most I’m cutting out two orders of magnitude background. Considering that I have 3 1/2 orders to do total, I was hoping to do more with this bdt.

What I showed on the call is a subset of the automatic output from TMVA. I still need to figure out how to get my trained bdt to run independently of the training, so I can get real rates measurements and all my other normal plots.

I’m a little troubled by a combination of stuff, any of which would be fine on its own but they don’t seem to go together:

1. (1)my output btd score has a 3-peak structure in the background,
   

2. (2)I’ve isolates which variable creates that structure (when I take that var out of the training, the 3 peaks colapse to a single peaks)
   

3. (3)That variable, (it’s the number of hits in the veto region above -200) doesn’t have a 3 peak structure itself
   

The thought on the call was that the 3 peaks probably come from the ice structure, entering through the “VGZ” variable (depth of the first vertex guess) but somehow it only shows up in the NVeto variable interactions. To test this, I can try taking VGZ out of the training, then redoing the NVeto in/out test (“does the 3-peak structure go away when I take out NVeto, even after VGZ has been taken out).

Also, there’s the little matter of getting the root output histograms to looks how I think they should, which will show that I’m actually doing the hit cleaning I say I am.