Source-linked AI summary
LHCb Topological Trigger Reoptimization
Tatiana Likhomanenko, Philip Ilten, Egor Khairullin, Alex Rogozhnikov, Andrey Ustyuzhanin, Michael Williams
TL;DR
LHCb reoptimizes its Run 2 topological trigger to improve selection efficiency across B-hadron decays while meeting online-processing constraints. The study compares multivariate classifiers and deployment strategies, selecting MatrixNet for the topological line. The reoptimized trigger improves efficiency over Run 1 by 15-60% at 2.5 kHz and 50-80% at 4 kHz, and the presented HLT scheme is intended for Run 2.
Problem
The Run 2 topological trigger must efficiently select diverse B-hadron decays, including modes absent from training, while operating within trigger output-rate and online-processing constraints.
Method
The study compares multivariate classifiers, tests generalization from six topological training modes to all available signal modes, and evaluates MatrixNet deployment through BBDT conversion and post-pruning.
Results
15-60% efficiency improvement was obtained at 2.5 kHz and 50-80% at 4 kHz relative to Run 1, with MatrixNet selected for the topological line.
Takeaways & Limitations
The presented HLT scheme will be applied in Run 2, with the preferred online MatrixNet implementation depending on background efficiency.
Abstract
from arXiv · showhide
The main b-physics trigger algorithm used by the LHCb experiment is the so-called topological trigger. The topological trigger selects vertices which are a) detached from the primary proton-proton collision and b) compatible with coming from the decay of a b-hadron. In the LHC Run 1, this trigger, which utilized a custom boosted decision tree algorithm, selected a nearly 100% pure sample of b-hadrons with a typical efficiency of 60-70%; its output was used in about 60% of LHCb papers. This talk presents studies carried out to optimize the topological trigger for LHC Run 2. In particular, we have carried out a detailed comparison of various machine learning classifier algorithms, e.g., AdaBoost, MatrixNet and neural networks. The topological trigger algorithm is designed to select all "interesting" decays of b-hadrons, but cannot be trained on every such decay. Studies have therefore been performed to determine how to optimize the performance of the classification algorithm on decays not used in the training. Methods studied include cascading, ensembling and blending techniques. Furthermore, novel boosting techniques have been implemented that will help reduce systematic uncertainties in Run 2 measurements. We demonstrate that the reoptimized topological trigger is expected to significantly improve on the Run 1 performance for a wide range of b-hadron decays.
1. Introduction
LHCb’s trigger has hardware and software stages, with HLT2 able to perform near-offline reconstruction and use multivariate algorithms. The note describes reoptimized HLT2 topological lines for Run 2, which efficiently trigger most n-body hadronic B decays.
- The trigger comprises a hardware L0 stage followed by two software high-level stages, HLT1 and HLT2.
- HLT2 processes few enough events to perform reconstruction similar to offline processing and use multivariate algorithms.
- The reoptimized HLT2 topological lines target efficient triggering of most hadronic B decays with at least three bodies.
2. HLT LHC Run 2 scheme
Run 2 adds a two-body secondary-vertex line to HLT1 and strengthens the topological HLT line with multivariate reconstruction of two-, three-, and four-track vertices. The study uses simulated 13-TeV B decays and generic proton-proton backgrounds.
- HLT LHC Run 2 scheme: Run 2 HLT1 combines a one-track line with a two-body secondary-vertex line because the tracking threshold enables efficient secondary-vertex selection.
- HLT LHC Run 2 scheme: The one-track line targets tracks with very high transverse momentum or large displacement from the primary vertex.
- HLT LHC Run 2 scheme: The two-body secondary-vertex line uses multivariate analysis to identify pairs of tracks forming a vertex.
- HLT LHC Run 2 scheme: The improved topological line uses a more powerful multivariate analysis on fully reconstructed events to find two-, three-, and four-track vertices.
- HLT LHC Run 2 scheme: Training uses simulated 13-TeV B decays for signal and generic Pythia 13-TeV proton-proton collisions for background.
3. Multivariate Analysis
The multivariate study evaluates classifier choices and trigger-line designs under output-rate constraints. MatrixNet is selected for the topological line, whose Run 2 efficiency improves substantially over Run 1 across the tested training modes.
- Classifier evaluation: A trigger event is accepted when at least one secondary vertex passes the classifier, while output-rate limits translate into background-efficiency constraints.For example, a 2 kHz output rate corresponds to a 0.2% false-positive rate.
- Classifier evaluation: Classifiers train on half the data and test on the other half; the topological line trains on six specific modes but tests all available signal modes.
- HLT “1-track” line: MatrixNet, logistic regression, and neural networks define candidate decision boundaries for the one-track line, with MatrixNet performing best in experiments.A simpler hyperbolic boundary is suitable for online processing.
- HLT “2 body SV” line: Removing the corrected-mass cut and classifier input has negligible performance impact for the two-body secondary-vertex line.The removals reduce systematic uncertainties and support exotic searches.
- HLT topological line: The topological line reconstructs tracks with PT > 200 MeV, rejects high trackχ2/ndof ghosts, and suppresses prompt particles using impact-parameter χ2 requirements.The line remains inclusive and allows omission of one or more B daughters when forming a trigger candidate.
- HLT topological line: A hierarchical blend feeds predictions from separate 2-, 3-, and 4-body MatrixNet classifiers into a final MatrixNet model and improves several training modes.
- HLT topological line: The selected MatrixNet model improves efficiency over Run 1 by 15-60% at 2.5 kHz and 50-80% at 4 kHz for six training modes.
4. Online processing
Online deployment requires converting or pruning MatrixNet because boosted decision trees are too slow for trigger processing. Both approaches accelerate prediction but reduce signal efficiency, and the preferred method depends on output rate.
- Online processing: Boosted decision trees are unsuitable for online trigger processing because of their low speed.
- Online processing: MatrixNet can be converted into bonsai boosted decision-tree format, while post-pruning reduces several thousand trees to a few hundred.
- Online processing: Both BBDT conversion and post-pruning significantly speed prediction but reduce MatrixNet signal efficiencies.BBDT deployment is additionally limited by the RAM size of its lookup table.
- Online processing: The preferred online model depends on background efficiency, or equivalently the selected output rate.
5. Conclusion
The Run 2 topological trigger was reoptimized, achieving substantial efficiency improvements at both tested output rates. Classifier preference depends on the selected background efficiency.
- The preference between BBDT and post-prunning depends on background efficiency, so ROC ordering varies with output rate.
- 15-60% efficiency improvement was obtained for 2.5 kHz output rate.
- 50-80% efficiency improvement was obtained for 4 kHz output rate.