BooNE

Particle Identification (PID)

  • We use hit topology and timing to identify events. Particles produce Cherenkov light in our tank, as well as some scintillation light, dependent on particle type.
  • Two independent methods to identify electron neutrinos in MiniBooNE: Boosted Decision Trees, and Track Based. The two methods use different event reconstruction fitters.

Boosted Decision Trees (BDT)

  • Decision trees are similar to neural nets, but don't suffer from the same pathologies.
  • To form a decision tree, take a sample with some signal and background, and some variables. Cut on the variable which gives the most separation of signal and background, according to the Gini Index, and make that cut. Then apply subsequent cuts in the same fashion.
  • A leaf is a sample where no more cuts are made; it is designated 'signal' if it is predominantly signal, and 'background' if mainly background.
  • "Boosting" is a method to additionally separate signal from background by weighting of events. Increase the weight of misclassified events (e.g. background events on a signal leaf), and remake the tree. Do this hundreds or thousands of times. Sum over the trees, by counting events on signal leaves as +1, -1 otherwise. This forms the PID variable.

References:

  • Hai-Jun Yang, Byron P. Roe, Ji Zhu(U. Michigan), "Studies of Stability and Robustness for Artificial Neural Networks and Boosted Decision Trees ", Accepted by Nucl. Instrum. & Meth. A (2007)
  • Hai-Jun Yang, Byron P. Roe, Ji Zhu(U. Michigan), "Studies of boosted decision trees for MiniBooNE particle identification", Nucl. Instrum. & Meth. A 555 (2005) 370-385
  • Byron P. Roe, Hai-Jun Yang, Ji Zhu(U. Michigan), "Boosted decision trees, a powerful event classifier", Proceedings of PHYSTAT05(Statistical Problems in Particle Physics, Astrophysics and Cosmology), Oxford, UK, September 12-15, 2005.
  • Byron P. Roe, Hai-Jun Yang, Ji Zhu(U. Michigan), Yong Liu, Ion Stancu(U. Alabama), Gordon McGregor(Los Alamos National Lab), " Boosted decision trees as an alternative to artificial neural networks for particle identification", Nucl. Instrum. & Meth. A 543 (2005) 577-584
  • Yong Liu, Ion Stancu (UA),"Cascade Training Technique for Particle Identification", physics/0611267 (submitted to NIMA)

Track Based Analysis (TBA)

  • We use fitters to decide if a ring is more electron like vs muon like and electron like vs π0 like. Also, we apply an π0 mass cut to select electron neutrino events.