UW classification: new background rejection trees

1
UW classificationnew background rejection trees

2
The eight selections (from Bill)
3
The prefilter cuts
Gamma classification Gamma classification
category prefilter remove if true
vertex-high AcdActiveDist gt -10 CalTrackAngle gt .5 CalTrackDoca gt 40
vertex-med AcdActiveDist gt -199 AcdRibbonActDist gt -1900 CalTrackDoca gt 200
vertex-thin AcdActiveDist gt -199 AcdRibbonActDist gt -1000
vertex-thick AcdUpperTileCount gt 0 AcdLowerTileCount gt 1 AcdRibbonActDist gt -1999
track-high CalTrackDoca gt 30    CalTrackAngle gt .3
track-med AcdActiveDist gt -199 AcdRibbonActDist gt -1900 CalTrackDoca gt 40     CalTrackAngle gt .5 CalXtalRatio gt .85
track-thin AcdActiveDist gt -199 AcdRibbonActDist gt -1999 CalTrackDoca gt 200    EvtECalTransRms lt .8
track-thick AcdActiveDist gt -199 AcdRibbonActDist gt -1999 AcdDoca lt 1999 CalTrackDoca gt 200    EvtECalTransRms gt 2.5 CalMaxXtalRatio gt .8 Tkr1FirstChisq gt 2.5    Tkr1ToTTrAve gt 2
4
Implementation in merit
file structure

• Each tree is described by two files
• dtree.txt ascii file with a list of weighted
  trees and nodes
• tree specify the weight to assign to the tree
• branch variable index, cut value
• leaf purity
• variables.txt list of the corresponding tuple
  variables
• Evaluation is by passing a vector of floats,
  ordered according to the variable list.
• Proposal to incorporate the prefilter cut in the
  tree description

5
Training details

• Weight signal and background to be the same
• Train on the EVEN events, with optional boosting
• Test with ODD events
• Save training and testing efficiency curves

6
Boosting what does it do?

• Nice interpolation for low background
• Not much improvement in actual separation (so far)

7
Preliminary single-tree background
8
What about the energy resolution?Validity
fractions
9
The fraction of time each estimate is best