Документ взят из кэша поисковой машины. Адрес оригинального документа : http://acat02.sinp.msu.ru/presentations/bhat/acat_moscow_pushpa.pdf Дата изменения: Sat Jul 6 17:22:18 2002 Дата индексирования: Mon Oct 1 20:32:29 2012 Кодировка: ISO8859-5

Pushpa Bhat, Fermilab
1


Richard Feynman at the Thinking Machines, Inc. (1983) The schematic representation of the Connection Machine that Feynman helped design, inspired the new ACAT logo. Feynman worked out in some detail the program for computing Hopfield's neural network on the Connection Machine Feynman also worked on cellular automata-based programs on the connection machine
June 24-28, 2002

Richard Feynman
ACAT2002, Moscow, Russia Pushpa Bhat 2


Run 2 Physics at Fermilab and Advanced Data Analysis Methods
Pushpa Bhat Fermi National Accelerator Laboratory
pushpa@fnal.gov or bhat@fnal.gov
ACAT 2002 Workshop June 24-28, 2002 Moscow, Russia
June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 3


Aerial View of Fermilab Complex

CDF

D0

Tevatron
Main Injector

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

4


The CDF Detector

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

5


The Upgraded Dи Detector

Inside collision hall before putting the wall
June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 6


Run 2 Status
Several Accelerator Upgrades including new Main Injector and antiproton Recycler Officially started in Spring 2001 s = 1.96 TeV (Run 1: 1.8 TeV) Commissioning & Performance Tuning Expected Integrated Luminosities/experiment
Run 2a: 2 fb-1 Run 2b: 15 fb
-1

Luminosity delivered so far: 55 pb-1
June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 7


Upgraded Detectors
CDF New Inner Tracking New Plug Calorimeter Upgraded Muon Detectors New Trigger and DAQ Large Central Tracking Volume Dи New inner tracking with 2T superconducting solenoid New Preshowers New Trigger and DAQ

Hermetic Compensated Uranium/LAr Calorimeter

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

8


A Rich Harvest of Physics in Run 2
The Standard stuff:
QCD Physics, Heavy Flavor, Electroweak

Top Physics and Evidence for single top production Important Precision Measurements W mass, top quark mass, cross sections Lots of interesting searches:
Higgs Boson Supersymmetry Strong Dynamics Exotics: Leptoquarks, etc. Extra Dimensions

Good prospects for discovering a low mass Higgs Boson, SUSY, and possible surprises!
June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 9


Standard Physics Signals
Central tracking allows significant improvement of momentum resolution for muons w.r.t. Run 1.
KS0+-

Entries / 20 MeV

300

Dи Run II Preliminary
mean = 3.095 Б 0.003 GeV = 0.130 Б 0.003 GeV

200

J/

100

0

2

3

4 5 + Mass ( Е , Е ) [GeV]

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

10


Standard Physics Signals
W's and Z's
Zee candidate mee = 93.2 GeV Zee selection

SMT+CFT Global tracks

We (no jets) selection
Events/2 GeV/c

Background mostly from 's, 0

CDF

2

MT (GeV/c2) June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 11


QCD Physics
Integrated Luminosity 1.9 pb
partially corrected dN/LdMjjd 1d 2 [nb/GeV]
partially corrected dN/Ldp Td [nb/GeV]
10

-1

Cone algorithm, R=0.7
1
-1

10

-1

Cone algorithm, R=0.7 | | < 0.5
-2

| | < 0.5

10

Inlusive jet pT spectrum

10

Dijet mass spectrum

10

-2

10

-3

10

-3

10

-4

Dи Preliminary
50 100 150 200 250 300 350 p T [GeV]

10

-4

Dи Preliminary
200 300 400 500 600 Mjj [GeV]

ЗVery preliminary corrections for jet energy
June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 12


eЕ Candidate Event
D Run 2 Preliminary
Electron MET

Top, SUSY, b-physics, Z() give rise to eЕ + X events

Muon

e ET = 13.9 GeV pT = 9.3 GeV = -0.425 = 1.251 Charge = -1

Е pT (toroid) = 16.4 GeV pT (central) = 6.3 GeV = -0.461 = 2.967 Charge = +1 MET = 6.0 GeV

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

13


Advanced Data Analysis Methods
It is well recognized by now that Advanced Multivariate & Statistical Data Analysis techniques are crucial for the success of our physics program Best use of data i.e., maximal use of available information is necessary to achieve
Optimal Separation of Signal and Background Optimal Measurements Accurate Estimation of Errors

The goal is to enable new discoveries, and produce results with better precision, robustness and clarity

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

14


Data Analysis Tasks
Particle Identification
e-ID, -ID, b-ID, e/, q/g

Signal/Background Event Classification
Signals of new physics are rare and small

Parameter Estimation
t mass, H mass, track parameters, for example

Function Approximation
Correction functions, tag rates, fake rates

Data Exploration
Data-driven extraction of information, latent structure analysis
June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 15


The Golden Rule
Keep it simple As simple as possible Not any simpler - Einstein

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

16


Some Multivariate Methods
Fisher Linear Discriminant Principal Component Analysis Independent Component Analysis Self Organizing Map Random Grid Search Probability Density Estimation Neural Network Support Vector Machine (FLD) (PCA) (ICA) (SOM) (RGS) (PDE) (NN) (SVM)

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

17


The Neural Network Revolution
Key factors responsible for the sweeping success of Neural Networks:
Power Ease of use Successful Applications

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

18


Examples from Run 1
Top Quark Discovery at Dи benefited from NN analysis: Comparisons helped arrive at optimized cuts Precision measurement of the top quark mass: used both NN and Bayesian analysis - the statistical uncertainty reduced by a factor of two Top in all-jets mode at Dи Limit on single top production cross section by Dи Top in all-jets mode and single top search by CDF World's best limit on 1st generation LQ mass by Dи And more ..

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

19


Measuring the Top Quark Mass
Discriminant variables

Dи Lepton+jets

The Discriminants

mt = 173.3 Б 5.6(stat.) Б 6.2 (syst.) GeV/c2

Fit performed in 2-D: (DLB/NN, mfit)
June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 20


Measuring the Top Quark Mass
old But there is more to be gained by using event by event signal probability distributions as a function of top mass and background probability, and building a likelihood for the sample, including matrix element Information.
Ps Ps + Pb

new

D=

In ensemble tests, the mass error (statistical) is a factor of 2 lower in the new method!
Pushpa Bhat 21

June 24-28, 2002

ACAT2002, Moscow, Russia


Top Physics in Run 2
Advanced methods will be used in a variety of studies both at CDF and Dи
All hadronic decay mode Tau decay modes Search for single top X tt

Recent studies at Dи show that further improvements in top mass measurement may be possible using fully probabilistic approach that exploits all features of individual events
June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 22


Where is Higgs?
6
theory uncertainty
had =
(5)

Higgs is at the heart of the EWSB pursuit Stringent constraints on the SM Higgs mass from LEP, SLD and Tevatron data
Precision EW measurements
80.6
LEP1, SLD Data - LEP2, pp Data 68% CL

0.02761Б0.00036 0.02747Б0.00012

4


2

2

80.5

100

400

mW [GeV]

Excluded 0 20

Preliminary

mH [GeV]

80.4


MW, mt measurements and correlation as predicted by EW theory for various mH suggest a low mass SM Higgs

80.3
mH [GeV] 114 300 1000 80.2 130 150 170

Preliminary

190

210

In many SUSY theories, mass of the lightest Higgs (h) < 150 GeV
June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 23

mt [GeV]


Discovering the Higgs Boson
The challenges are daunting! But using NN provides same reach with a factor of 2 less luminosity w.r.t. conventional analysis Improved bb mass resolution & b-tag efficiency crucial
Combined Results(WH+ZH) Required Luminosity (fb-1)
60

50

Standard cuts (5) NN cuts (5)

40

30

20

10

0

85

90

95

100

105

110

115

120

125

130

135
2

MH (GeV/c )

Run II Higgs study hep-ph/0010338 (Oct-2000) P.C.Bhat, R.Gilmartin, H.Prosper, Phys.Rev.D.62 (2000) 074022 T.Han, A.S.Turcot , and R.-J.Zhang, Phys.Rev.D59(1999) See also Lev Dudko's talk in parallel session June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 24


Template Fitting
Nice application of NN by ALEPH in the search

ZH qq bb
Could be employed in top mass analysis and many other cases

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

25


Supersymmetry and Beyond
The trilepton channels may be clean modes but jetty SUSY/technicolor channels certainly will benefit from multivariate methods Run 1 searches for SUSY in the e+jets +MET channel and Technirho at Dи use NN Big help in tau channels Search for Extra Dimensions (no more fiction!) benefits from tight control of fake leptons/ photons (good ID and fake rate estimation)

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

26


eee Candidate Event

D Run 2 Preliminary

e1 ET = 17.9 GeV pT = 0.52 GeV = 0.43 = 5.42 Charge = +1 m
e1e2

e2 ET = 13.9 GeV pT = 10.9 GeV = -1.94 = 2.80 Charge = +1 m
e1e3

e3 ET = 13.2 GeV pT = 15.1 GeV = 1.06 = 5.72 Charge = -1 m
e2e3

Electron

Trilepton events are classical SUSY signature Electrons
Pushpa Bhat 27

= 55.7 m
e1e2e3

= 10.8

= 63.5

= 85.2 GeV/c2

MET =10.7 GeV

June 24-28, 2002

ACAT2002, Moscow, Russia


eЕЕ Candidate Event
D Run 2 Preliminary
MET

Muon system
e ET = 19.2 GeV = 0.40 = 0.63 No track match Е1 pT = 28.2 GeV = -0.10 = 6.20 Charge = -1 m
ЕЕ

Е2 pT = 9.82 GeV = -1.48 = 2.88 Charge = 1 = 41.5 GeV/c2

Muon

Electron Muon

MET =31.8 GeV

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

28


Multivariate Analysis Issues
Dimensionality Reduction
Choosing Variables optimally without losing information

Choosing the right method for the problem Controlling Model Complexity Testing Convergence Validation
Given a limited sample what is the best way?

Computational Efficiency

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

29


More Issues
Apart from the usual stuff, Quantifying correctness of modeling or goodness of learning (fit) Checking the robustness of results Abstracting the response function or the mapping function from Monte Carlo
Inverse Monte Carlo (?)

June 24-28, 2002

ACAT2002, Moscow, Russia

Pushpa Bhat

30


Multivariate Bayesian

Methods and Statistics

Both Ancient Concepts; A lot of new approaches, algorithms and applications Adaptive learning and Stochastic optimization revolutionized the landscape for multivariate analysis Some hard problems can't be solved without Bayesian thinking! New Kids on the Block in HEP; Concerns 1990's: Why NN or Why Bayesian? Now: Why NOT?
June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 31


Exploring Models: Bayesian Approach
Enables straight-forward and meaningful model comparisons. Allows treatment of all uncertainties in a consistent manner. Provides probabilistic information on each parameter of a model (SUSY, for example) via marginalization over other parameters Mathematically linked to adaptive algorithms such as Neural Networks (NN) Hybrid methods involving multivariate probability density estimation and Bayesian treatment can be very powerful
June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 32


Summary
Run 2 at Fermilab is well underway; CDF and Dи will record unprecedented amounts of data in the coming years: 2 fb-1 in Run 2a, > 15 fb-1 in Run 2b Use of advanced "optimal" analysis techniques will be crucial to achieve the physics goals Multivariate methods, particularly Neural Network techniques, have already made impact on discoveries and precision measurements and will be the methods of choice in future analyses Hybrid methods combining "intelligent" algorithms and probabilistic approach will be the wave of the future We hope to unravel some of nature's mysteries!
June 24-28, 2002 ACAT2002, Moscow, Russia Pushpa Bhat 33