Docstoc

ttH-Analysis

Document Sample
ttH-Analysis Powered By Docstoc
					CMS Group Meeting – 22.11.05


                                        Update III:

                          ttH-Analysis
                                   with neural nets

    1. Overview
    2. Pairing
    3. Background Reduction
    4. Outlook




 Institut für Experimentelle Kernphysik (IEKP)        Dennis Schieferdecker
1. Overview                                                          1

Software and Hardware used for the analysis:
 NeuroBayes v1.3
 PAX v2.00.10
 ROOT v4.02/00
 Athlon64 (3,0 Ghz, 2GByte RAM)

Event samples used:
 Signal (ttH) and Background (ttbb) events from 30.9.05,
 including generated and reconstructed events
 (with matching information added later)
 Signal:
 86907 Events, 16601 good matched
 Background:
 83292 Events, 13829 good matched

  Institut für Experimentelle Kernphysik (IEKP)   Dennis Schieferdecker
2. Pairing – Using Cuts                                                                  2


Modification of analysis:
 Only those events are considered whose best combination has
 at least a certain net output value

-Generator Events-
Signal:
Cut:                              0,0         0,5     0,6     0,7     0,8
Correct Pairings:                 71,2%       79,1%   80,1%   82,7%   82,7%
Efficiency:                       100%        76,0%   70,2%   52,0%   52,0%
(normalized on cut value 0,0)


Background:
Cut:                              0,0         0,5     0,6     0,7     0,8
Correct Pairings:                 75,0%       82,1%   83,3%   84,2%   84,4
Efficiency:                       100%        86,7%   79,3%   60,5%   50,9%
(normalized on cut value 0,0)

   Institut für Experimentelle Kernphysik (IEKP)                      Dennis Schieferdecker
2. Pairing – Using Cuts                                                                  3


Modification of analysis:
 Only those events are considered whose best combination has
 at least a certain net output value

-Reconstructed Events-
Signal:
Cut:                              0,0         0,5     0,6     0,7     0,8
Correct Pairings:                 31,2%       31,7%   32,7%   36,2%   42,0%
Efficiency:                       100%        96,7%   85,1%   54,4%   23,5%
(normalized on cut value 0,0)


Background:
Cut:                              0,0         0,5     0,6     0,7     0,8
Correct Pairings:                 37,0%       38,2%   40,4%   46,2%   53,5%
Efficiency:                       100%        94,6%   79,5%   43,8%   16,6%
(normalized on cut value 0,0)

   Institut für Experimentelle Kernphysik (IEKP)                      Dennis Schieferdecker
2. Pairing – Using cuts                                                                                  4


Impact of cuts on invariant mass shapes:
    (for netout < 0.6, Generator Events)
             Background (bb mass)                                      Signal (Higgs mass)
#                                                          #

                  before cut


                                                               before cut                    after cut

                                     after cut




                                                 E [GeV]                                            E [GeV]

Results after applying cuts:
 Pairing efficiency can be improved, but
 Cuts often cut more signal than background events
 Cuts don’t change the shape of the invariant bb mass
     Institut für Experimentelle Kernphysik (IEKP)                               Dennis Schieferdecker
3. Background Reduction – Invariant t* mass                                             5


Invariant mass of t* as input variable:
  Asumption: net will learn generator set Higgs mass
  Test: check shape of invariant bb mass
               bb mass (with t*)                        bb mass (without t*)
#                                                   #


                 before cut                             before cut




                                                               after cut
                              after cut




                                          E [GeV]                                 E [GeV]
                                                              (on Generator Events)
Results:
 Using this variable leads to a shifting of the bb mass shape
    Institut für Experimentelle Kernphysik (IEKP)                    Dennis Schieferdecker
2. Background Reduction – Reconstructed Events                                     6


Net output using reconstructed events:
 (red: background, blue: signal)

  #                with t* mass                     #   without t* mass




                                           netout                             netout

Only "good matched" matched events where used, meaning:
 particle identities were taken from generator level
 No pairing code was used and not all reconstructed jets were
 considered
  Institut für Experimentelle Kernphysik (IEKP)                 Dennis Schieferdecker
2. Background Reduction – Reconstructed Events                                              7


Input variables for analysis on reconstructed events:
  Same input variables were used
  Comparisson plots of these variables for Generator &
  Reconstructed Events are shown on the next slides:



Invariant t* mass:
                     #       Generator Events                #   Reconstructed Events




                                                   E [GeV]                               E [GeV]

   Institut für Experimentelle Kernphysik (IEKP)                         Dennis Schieferdecker
2. Background Reduction – Reconstructed Events                                              8

Angle (t   , bb):   #             Generator Events               Reconstructed Events
        raen                                                 #




                                                     angle                               angle
Rapidity difference (t
                     raen, bb):
                     #            Generator Events           #   Reconstructed Events




                                                     ∆rap                                ∆rap

   Institut für Experimentelle Kernphysik (IEKP)                         Dennis Schieferdecker
2. Background Reduction – Reconstructed Events                                         9

∆R (t , bb):       #         Generator Events               Reconstructed Events
     raen                                               #
  using rapidity




∆R (t , bb):
     raen                                          ∆R                                ∆R
  using pseudo-rapidity
                  #          Generator Events           #   Reconstructed Events




                                                   ∆R                                ∆R

   Institut für Experimentelle Kernphysik (IEKP)                    Dennis Schieferdecker
2. Background Reduction – Reconstructed Events                        10


Comparisson of efficiency for Generator and Reconstructed Events:
                  Background
                    efficiency

  x: Generated with t*
  +: Generated w/o t*

  x: Reconstructed with t*
  +: Reconstructed w/o t*




                                                                  Signal
Further steps:                                                efficiency

 Use pairing analysis to provide particle identities for background
 reduction analysis
  Institut für Experimentelle Kernphysik (IEKP)     Dennis Schieferdecker
3. NB & MLP comparisson – Pairing                                             11


Comparisson for Pairing Analysis:
 Same analysis code was used, only the net class was changed

#                NeuroBayes                             ROOT MLP
                                                    #




                                           netout                           netout

(on Generator Events)

Results:
 MLP is inadequate for solving this problem
    Institut für Experimentelle Kernphysik (IEKP)           Dennis Schieferdecker
3. NB & MLP comparisson – Background suppression                               12


Comparisson for Background Suppression:
 Again same analysis code was used and net class was changed
                 NeuroBayes                              ROOT MLP
#                                                    #




                                            netout                           netout

(on Generator Events)

Results:
 MLP can solve this problem about as well as NB
    Institut für Experimentelle Kernphysik (IEKP)            Dennis Schieferdecker
4. Outlook                                                          13


Pairing:
 Consider all jets in reconstructed events,
 not just “good matched” ones



Background suppression:
 Use pairing results for particle selection
 Use ttjj background samples



Write diploma thesis:
 To be done till December 20th



  Institut für Experimentelle Kernphysik (IEKP)   Dennis Schieferdecker

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:9
posted:11/5/2011
language:English
pages:14