Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero...

24
Bruce Knuteson Quaero (I search for, I seek) Multivariate Workshop, Caltech, Feb 11 2008 The problem The solution

Transcript of Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero...

Page 10: Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero algorithm overview (you wish to test a hypothesis H ) Quaero@H1 S. Caron, B. Knuteson

Bruce Knuteson

Quaero@D0RunIDØ CollaborationPhys.Rev.Lett.87:231801,2001

Quaero@H1S. Caron, B. KnutesonEur.Phys.J.C53:167-175,2008

10

Page 11: Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero algorithm overview (you wish to test a hypothesis H ) Quaero@H1 S. Caron, B. Knuteson

11

• Time budget is calculated

• H events are run through the detector simulation

• H, SM, data are partitioned into final states• Variables are chosen automatically • Binning is chosen automatically• A binned likelihood is calculated• Results from different final states are combined• Results from different experiments are combined• Systematic errors are integrated numerically• Result returned

Quaero algorithm overview(you wish to test a hypothesis H )

Page 12: Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero algorithm overview (you wish to test a hypothesis H ) Quaero@H1 S. Caron, B. Knuteson

Quaero adjusts its analysis strategy to

fit within time budget

Page 13: Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero algorithm overview (you wish to test a hypothesis H ) Quaero@H1 S. Caron, B. Knuteson

13

• Time budget is calculated

• H events are run through the detector simulation

• H, SM, data are partitioned into final states• Variables are chosen automatically • Binning is chosen automatically• A binned likelihood is calculated• Results from different final states are combined• Results from different experiments are combined• Systematic errors are integrated numerically• Result returned

Quaero algorithm overview(you wish to test a hypothesis H )

Page 14: Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero algorithm overview (you wish to test a hypothesis H ) Quaero@H1 S. Caron, B. Knuteson

Bruce Knuteson

TurboSim

A fast detector simulation that

tunes itself to any experiment’s

detailed detector simulation

Full simulation 100 seconds

TurboSim0.01 seconds

Bruce Knuteson

Page 15: Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero algorithm overview (you wish to test a hypothesis H ) Quaero@H1 S. Caron, B. Knuteson

15

• Time budget is calculated

• H events are run through the detector simulation

• H, SM, data are partitioned into final states• Variables are chosen automatically • Binning is chosen automatically• A binned likelihood is calculated• Results from different final states are combined• Results from different experiments are combined• Systematic errors are integrated numerically• Result returned

Quaero algorithm overview(you wish to test a hypothesis H )

Page 17: Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero algorithm overview (you wish to test a hypothesis H ) Quaero@H1 S. Caron, B. Knuteson

17

• Time budget is calculated

• H events are run through the detector simulation

• H, SM, data are partitioned into final states• Variables are chosen automatically • Binning is chosen automatically• A binned likelihood is calculated• Results from different final states are combined• Results from different experiments are combined• Systematic errors are integrated numerically• Result returned

Quaero algorithm overview(you wish to test a hypothesis H )

Page 18: Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero algorithm overview (you wish to test a hypothesis H ) Quaero@H1 S. Caron, B. Knuteson

18

Variables are chosen separately for each final state

Dimensional Rule of Thumb: > 10d events are needed to adequately

populate a d-dimensional space Corollary: analysis should be performed in a space of

dimensionality d = log10NMC Prescription: 1. Generate a long (but finite) list of relevant variables TeV: pT, ϕ, η, Δϕij, ΔRij, mij, mijk, mijkl

LEP: E, ϕ, θ, Δϕij, ΔRij, mij, mijk, mijkl

2. Order according to decreasing discrepancy (h vs b)

3. Use the first d variables in the list, removing highly correlated variables

Goals: speed, robustness, transparency

Page 19: Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero algorithm overview (you wish to test a hypothesis H ) Quaero@H1 S. Caron, B. Knuteson

19

• Time budget is calculated

• H events are run through the detector simulation

• H, SM, data are partitioned into final states• Variables are chosen automatically • Binning is chosen automatically• A binned likelihood is calculated• Results from different final states are combined• Results from different experiments are combined• Systematic errors are integrated numerically• Result returned

Quaero algorithm overview(you wish to test a hypothesis H )

Page 20: Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero algorithm overview (you wish to test a hypothesis H ) Quaero@H1 S. Caron, B. Knuteson

20

FewKDE

Typical kernel solution

1) place “bumps of probability” around each Monte Carlo point

2) sum these bumps into a continuous distribution

Time cost is O(N2) FewKDE

fit for parameters of a handful of Gaussians appropriately handle hard physical boundaries

Form a discriminant

Page 21: Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero algorithm overview (you wish to test a hypothesis H ) Quaero@H1 S. Caron, B. Knuteson

21

Expected # of events / unit of x

Expected evidence

OptimalBinning

Page 22: Quaero - California Institute of Technologyveverka/files/20080211_knuteson...2008/02/11  · Quaero algorithm overview (you wish to test a hypothesis H ) Quaero@H1 S. Caron, B. Knuteson

22

• Time budget is calculated

• H events are run through the detector simulation

• H, SM, data are partitioned into final states• Variables are chosen automatically • Binning is chosen automatically• A binned likelihood is calculated• Results from different final states are combined• Results from different experiments are combined• Systematic errors are integrated numerically• Result returned

Quaero algorithm overview(you wish to test a hypothesis H )