Source-linked AI summary
Efficient, reliable and fast high-level triggering using a bonsai boosted decision tree
Vladimir Vava Gligorov, Mike Williams
TL;DR
Modern experiments need online triggers because collision rates and event sizes produce data volumes that cannot be stored or processed in full. This paper modifies BDTs into bonsai BDTs, which are efficient, fast, robust to instabilities, and demonstrated at LHCb.
Problem
High collision rates and large detector event sizes create data volumes that require online reduction, while many signals cannot be separated from backgrounds using L0, single-track, or cut-based multi-track methods.
Method
The paper introduces a modification of the standard BDT algorithm for high-level triggering and illustrates its operation using toy-model data and LHCb experience.
Results
The BBDT is more efficient than traditional cut-based approaches, nearly as stable as cuts while nearly as efficient as a BDT, and highly efficient across varied B-decay topologies.
Takeaways & Limitations
The BBDT is fit for online running in large-scale data-acquisition systems and should be considered for current or future HLT algorithms.
Takeaways & Limitations
The discretization must exceed detector resolution and expected online variations, and must reflect which training-signal PDF properties the BBDT should incorporate.
Abstract
from arXiv · showhide
High-level triggering is a vital component in many modern particle physics experiments. This paper describes a modification to the standard boosted decision tree (BDT) classifier, the so-called "bonsai" BDT, that has the following important properties: it is more efficient than traditional cut-based approaches; it is robust against detector instabilities, and it is very fast. Thus, it is fit-for-purpose for the online running conditions faced by any large-scale data acquisition system.
1. Introduction
Rising collision rates, luminosities, and detector event sizes have made online high-level triggering essential for reducing data to manageable volumes. The paper motivates an HLT approach that uses full-event information while addressing the limitations of hardware and cut-based triggers.
- Motivation: Modern accelerators and detectors produce data volumes too large to store and process without online trigger reduction.Triggers preferentially retain events considered signal for later analysis.
- Trigger architectures: L0 triggers are fast and reliably modeled but use only low-level detector information, limiting their effectiveness when signal traits are not simple or readily accessible.They can use single-subdetector information or simple inter-detector coincidences.
- Trigger architectures: HLT algorithms perform online event reconstruction and make decisions using full event information, often with simplified geometry or coarser granularity than offline reconstruction.These simplifications accommodate online timing constraints.
- Existing approach: Traditional HLT algorithms are cut-based, combining largely uncorrelated individual criteria through logical ANDs despite access to full detector information.A representative selection may require high transverse momentum, displacement, and invariant mass simultaneously.
- LHCb setting: At LHCb, data rates reached over 10 MHz and approximately 1 TB/s, while many signals could not be separated from backgrounds using L0, single-track, or cut-based multitrack algorithms.These conditions require substantial online data reduction before storage or analysis.
- Contribution: The paper presents a modified BDT HLT algorithm designed for LHCb but intended to illustrate a method usable by other experiments.The algorithm had run since the start of LHCb data taking in 2011, with the paper covering its design, toy-model benchmark, and LHCb performance.
2. The Bonsai Algorithm
The bonsai BDT adapts boosted decision trees for online triggering by discretizing variables, constraining tree growth, and enabling rapid response lookup. These design choices target detector stability, imperfect signal models, and stringent online speed requirements.
- Decision trees: Boosting methods such as bagging reduce overtraining by combining many decision trees trained on bootstrap samples.The BDT response is the fraction of trees assigning an event to a signal leaf.
- Design requirements: Multivariate classifiers define signal or keep regions from training data, but HLT use requires addressing stability, imperfect signal PDFs, and speed.Small keep regions can cause efficiency oscillations as detector conditions vary.
- Core idea: The bonsai BDT addresses these concerns by discretizing every variable, limiting split locations and giving the tree grower control over its structure.This constrained growth motivates the name bonsai BDT.
- Discretization choices: The discretization scale should exceed detector resolution and expected online variation, while reflecting which common signal traits the classifier should learn.For inclusive triggers, the discretization should discourage learning traits specific only to the training samples.
- Implementation: Finite discretization creates a finite number of possible keep regions, allowing BBDT responses to be stored in a one-dimensional array for extremely fast lookup.This replaces the large collection of conditional statements in a conventional BDT.
- Resulting properties: By construction, the BBDT is suited to HLT use and addresses the stated concerns surrounding multivariate online classification.Its design connects controlled region size with detector robustness and array lookup with online speed.
3. Toy Model Example
The toy model compares cut-based, BDT-based, and bonsai BDT HLT selections using simulated B-meson signals and several background types. Selections are optimized for a 100-fold background reduction, then evaluated for efficiency and stability under imperfect online conditions.
- Trigger discriminating variables: The simulation uses sums of track transverse momenta and impact parameters from three-track vertices as trigger inputs.All possible track combinations are considered, and an event passes when at least one combination satisfies the trigger algorithm.
- Toy model signals: Four- and five-body B-meson decays test whether training on one signal transfers to another.The five-body test is made harder by discarding its lowest-transverse-momentum track before forming trigger candidates.
- Toy model backgrounds: Three simulated backgrounds represent pure combinatorics, ghost tracks, and prompt charm, with ghost-plus-charm combinations identified as least abundant but most dangerous.Each background event contains 30 primary-originating tracks, 10% ghosts, and one D+ decay.
- Evaluation procedure: Selections are trained on the four-body signal and background, then validated on both signal topologies and background at a factor-of-100 background reduction.The optimal selection maximizes four-body signal efficiency subject to that reduction.
- Efficiency comparison: The BDT and BBDT are more efficient than cuts, while the BBDT is 3% less efficient than the BDT on four-body signals but 1% more efficient on five-body signals.The BBDT-based HLT is 22% higher than the cut-based one in the reported comparison.
- Stability comparison: The BBDT rate remains very close to the cut-based rate under imperfect online conditions, whereas the BDT rate increases six times more than the cut-based rate.The study links finer allowed split points to higher efficiency but greater sensitivity to online instabilities; the example uses 10 split points per variable and a smallest keep region 5σ wide.
4. Performance at LHCb
The paper reports the bonsai BDT's deployment in LHCb's topological trigger for B-physics selections. It remained stable across detector misalignments and collision-energy changes, while showing high efficiency across varied B-decay topologies.
- Deployed trigger: The deployed BBDT topological trigger reconstructs displaced two-, three-, and four-track vertices and evaluates their vertex and track properties.It targets B-meson decays producing charged tracks, including kaons, pions, protons, muons, and electrons.
- Operational performance: The trigger covered about 98% of the luminosity collected to date and served as LHCb's main trigger for B physics without fundamental changes or retunings.This operational record includes the stated deployment period.
- Operational performance: It proved robust against annual detector misalignments after hardware interventions and against the change from 7 to 8 TeV collision energy.The paper reports both effects as observed deployment conditions.
- Measured performance: Data-driven measurements found high efficiency across varied B-decay topologies without pathological structures indicative of overtraining or resolution effects.Simulation also verified maintained performance at 14 TeV and 25% greater per-bunch luminosity.
5. Summary
The BBDT is more efficient than traditional cut-based approaches, robust against detector instabilities, and very fast. These properties make it fit for online operation in large-scale data acquisition systems.
- BBDT is more efficient than traditional cut-based approaches.
- BBDT is robust against detector instabilities.
- BBDT is very fast and fit-for-purpose for large-scale online data acquisition.