Photon Not Meeting 27 th July 2017 1 TMVA Classification Can now - - PowerPoint PPT Presentation

1

Photon Not Meeting

27th July 2017

2

TMVA Classification

• For several TMVA methods

varied the cut on the response distribution to maximize signal/sqrt(background).

• Largest

signal/sqrt(background)) method is the adaptive BDT method:

–

Signal efficiency: 33 %

–

Background efficiency: 0.07 %

–

signal/sqrt(background): 0.33

• Can now extract the response variable output by TMVA to cut on my events

and scale selected signal and background events to determine signal/sqrt(background).

• Variables used in TMVA can be found here:

http://www.hep.manchester.ac.uk/u/murrells/slides/photonmeeting_20_07_201 7/variables

3

Missing Variable Problem and Correlations

• Most variables are

uncorrelated excluding track variables.

• Track correlations result

from vertices with no associated track where all track variables are set to some “undefined” value e.g. 0, -1, -2, etc.

• Attempted solution:

classify vertices with no tracks and vertices with > 0 tracks separately.

Track variables Correlation matrix for signal sample, background looks similar.

4

w and w/o Track Classifications: Correlations

• Run separate classifications for vertices with associated tracks and

without associated tracks. In the no track case the corresponding track variables are removed from consideration.

• This completely removes the previously seen track variable

correlations.

With associated tracks Without associated tracks

5

w and w/o Track Classifications: Maximizing signal/sqrt(background)

• Vary cuts for both responses over a 2d space to find cut combination that

maximizes signal/sqrt(background).

• w and w/o track classification improves the performance of most methods

but worsens the performance of the BDT adaptive method which is the best performing method for this classification also.

With associated tracks Without associated tracks

6

Classification Summary

• Note that for the work shown TMVA was run without applying event

weights (i.e. POT scaling) to each sample.

• Applying these weights causes most methods to perform worse (POT

scaling is still applied when calculating signal/sqrt(background)).

• Classifications:

– Single:

• Efficiency – signal: 32.6 % background: 0.07 %
• signal/sqrt(backround) = 0.333.
• Largest background eliminated signal efficiency: 1 %.

– w and w/o track:

• Efficiency – signal: 31.5 % background: 0.07 %
• signal/sqrt(backround) = 0.332.
• Largest background eliminated signal efficiency: 3 %.
• Note: instances of eliminated background more likely due to be

relatively low simulated statistics compared to what would actually be seen in the detector.

7

Future Work

• Found a bug where a variable was not being correctly set to

“undefined” with no associated tracks present. Fixing this may cause the w and w/o track classification to perform better.

• Now that conversion to LArLite is no longer a factor I will run over

a larger set of simulated BNB + cosmic and cosmic in-time background samples for better statistics.

• Generate MCC8.1 signal sample with simulated cosmics. An

MCC8.1 version of in-time cosmics does not exist so may need to generate one of those as well.

• Implement shower dE/dx.