Change found hit bonus to 7.5 #241
Conversation
|
A bonus of 7.5 definitely puts us much closer to CMSSW in nLayers, completely within uncertainties for the barrel, and now only down on average 2 layers in the endcaps (compared to missing 2 layers in the barrel and 4 in the endcaps for the hit bonus of 5). This is definitely a better point to show to CMS during the Tracking POG meetings / CMS week. |
|
@areinsvo |
|
@slava77 just to point out that there is indeed something "special" with a valid bonus of 10, since the chi2cut is 15 and the invalid penalty is 5. So it's not unlikely that there is a "cliff in efficiency" nearby. See Allie's slides for a more complete discussion (https://indico.cern.ch/event/827099/contributions/3460338/attachments/1894190/3124594/OptimizeEfficiency_Aug2019.pdf) |
this implies a large phase space of candidates in the range between "validBonus + penalty" and the chi2 cut off of 15. Currently it's 12.5 vs 15. Naively, this phase space should be small though. |
|
For the sake of documentation: @slava77 the comparison between red and black/orange is not completely straightforward, because the hit bonus changed but we also changed from "V1" to "V2" of the candidate score. V1 had additional categories that tweaked the bonus or penalties, so in reality the bonus was at least 17 in each category. |
ah, good point. |
|
Have we decided we want to make this our baseline, given that we are using this for the Sep. 9 TRK POG meeting? Or are we waiting for further optimization? |
|
I believe this is a better baseline, and we can take it as a landmark for the presentation.
M.
…________________________________________
From: Kevin McDermott [[email protected]]
Sent: 05 September 2019 16:34
To: trackreco/mkFit
Cc: Subscribed
Subject: Re: [trackreco/mkFit] Change found hit bonus to 7.5 (#241)
Have we decided we want to make this our baseline, given that we are using this for the Sep. 9 TRK POG meeting? Or are we waiting for further optimization?
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub<#241?email_source=notifications&email_token=ABDV5QDEQ3HFCV2EMLFCNTDQIEKITA5CNFSM4IR3YFVKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD57KKQA#issuecomment-528393536>, or mute the thread<https://github.com/notifications/unsubscribe-auth/ABDV5QATUNJZAJPZKEX5LPTQIEKITANCNFSM4IR3YFVA>.
|
Changing the hit bonus to 7.5 maintains a track building efficiency comparable to CMSSW but improves the number of strip layers in the built tracks (compared to the hit bonus of 5). Obviously this still doesn't represent a fine tuning, but it is a better set of comparison plots to use for the POG update than my initial guess of 5. Mario said he might have time to work on a further optimization in the next week.
Standalone validation plots: http://areinsvo.web.cern.ch/areinsvo/MkFit/Benchmarks/PR240/standalone
MTV plots (clean version): http://areinsvo.web.cern.ch/areinsvo/MkFit/Benchmarks/PR240/plots_NewScore_Bonus7p5
MTV plots (messy version, with old mkFit score, new mkFit score with bonus of 5, and new mkFit score with bonus of 7.5): http://areinsvo.web.cern.ch/areinsvo/MkFit/Benchmarks/PR240/plots_compareScores/
A few notable plots include the efficiency comparison here and the comparison of the number of layers in the built tracks here.