Learning to Trigger: Reinforcement Learning at the Large Hadron Collider

Zixin Ding, Shaghayegh Emami, Giovanna Salvi, Cecilia Tosciri, Abhijith Gandrakota, Ngadiuba Jennifer, Nhan Tran, Christian Herwig, David Miller, Yuxin Chen

TL;DR

This paper introduces a novel reinforcement learning (RL) approach, Group-Filtered Policy Optimization (GFPO), to dynamically adjust trigger thresholds in real-time at the Large Hadron Collider (LHC). By adapting to changing conditions, GFPO significantly improves the efficiency of detecting rare physics events while maintaining crucial background rate constraints, demonstrating its effectiveness on both simulated and real LHC data.

What Problem Does This Paper Address?

The Large Hadron Collider (LHC) generates an immense amount of data, necessitating a sophisticated real-time "trigger" system to filter events and select those worthy of further analysis. This trigger system relies on carefully tuned thresholds that are typically static and hand-crafted. However, the conditions within the LHC (like "pileup" – the number of simultaneous collisions – and detector performance) can drift over time. This drift can make static thresholds suboptimal, leading to missed physics signals or an overload of uninteresting background events. Manually re-tuning these thresholds is a costly and time-consuming process. The paper addresses the challenge of creating an autonomous system that can adapt these trigger thresholds in real-time to maintain optimal performance under these dynamic conditions.

Methods

The core of the paper's contribution lies in its novel RL approach called Group-Filtered Policy Optimization (GFPO). Instead of traditional RL methods that can struggle with the strict rate constraints and non-stationary environments of the LHC, GFPO introduces a crucial "feasibility-first" filtering mechanism.

Here's a breakdown:

  • Casting the Problem as RL: The problem of adjusting trigger thresholds is framed as a sequential decision-making problem, where an RL agent learns to make decisions (adjusting thresholds) over time to maximize a reward (signal efficiency) while adhering to constraints (background rate).
  • Novel State Representation: The agent doesn't just look at the immediate event but uses a sequence of recent events and control features (like rate error, pileup information, and threshold geometry) to build a richer understanding of the current state. This is achieved using a recurrent neural network (RNN) architecture.
  • Group-Filtered Policy Optimization (GFPO): This is the key innovation. GFPO addresses the challenge of sampling actions (threshold adjustments) that might violate the strict background rate constraints.
  • Feasibility Filtering: Before evaluating the "goodness" of potential threshold adjustments, GFPO first filters them to ensure they are likely to meet the background rate target. This is crucial because in the LHC, violating the rate budget can lead to data loss.
  • Two Variants (GFPO-F and GFPO-FR): * GFPO-F: Prioritizes minimizing the deviation from the target background rate.
  • GFPO-FR: Prioritizes signal efficiency within the feasible region, allowing for slightly more rate deviation.
  • Comparison to Baselines: The paper compares GFPO to several established methods, including PID controllers, DSPOT (a statistical method), and other RL algorithms like DQN and GRPO.

Datasets

The research utilizes two primary types of data:

  • Monte Carlo (MC) Simulation Data: This data is generated to emulate realistic collider conditions and is used for training and initial evaluation of the RL agents. Specific datasets include background events and simulated signals (tt and h→4b).
  • Real CMS Collision Data: Data from the CMS experiment at the LHC (specifically Run 283408) is used for evaluating the performance of the trained agents in a real-world scenario. This includes both background and anomaly detection training data.
  • Anomaly Detection Benchmarks: To demonstrate the generalizability of their approach beyond particle physics, the authors also evaluate their methods on two standard anomaly detection datasets:
  • UNSW-NB15: A network intrusion detection dataset.
  • Numenta Anomaly Benchmark (NAB): A benchmark for streaming anomaly detection on time-series data.

Key Insights and Ideas

  • Feasibility-First Filtering is Crucial: The most significant insight is that simply optimizing for reward (signal efficiency) isn't enough in constrained environments like the LHC trigger. Prioritizing actions that satisfy the hard constraints (background rate) before optimizing for performance is essential for robust and effective control.
  • RL Can Outperform Static and Simple Adaptive Methods: The paper clearly shows that RL, particularly GFPO, can achieve better performance than traditional hand-tuned thresholds and simpler adaptive controllers like PID, especially under challenging conditions like pileup drift.
  • Sim-to-Real Transfer is Promising: The trained RL agents, without fine-tuning on real data, perform remarkably well when deployed on actual LHC collision data. This suggests that well-designed RL policies can generalize effectively from simulations to real-world systems.
  • The "Zero-Feasible" Problem: A key challenge identified is that in certain situations, the sampled actions might all violate the rate constraint. Standard RL methods can struggle here, but GFPO's filtering mechanism helps mitigate this issue.
  • Dynamic Thresholding is the Future: The success of GFPO highlights the potential for more intelligent, adaptive trigger systems in high-energy physics and other data-intensive scientific fields.

Key Results

  • Improved In-Band Rate: GFPO significantly increases the fraction of time the trigger system operates within the desired background rate tolerance band.
  • On MC data: 48% improvement for HT trigger, 28% for AD trigger.
  • On real CMS data (without fine-tuning): 56% improvement for HT trigger, 28% for AD trigger.
  • Increased Signal Efficiency: Alongside improved rate control, GFPO also achieves higher signal efficiency.
  • On MC data: up to 2% cumulative gain on in-tolerance intervals.
  • On real CMS data: further signal efficiency gains observed.
  • Robustness to Distributional Shift: GFPO demonstrates strong performance even when deployed on real data without fine-tuning, indicating its robustness to the differences between simulated and real-world conditions.
  • First RL-Based Trigger Control on Real LHC Data: This paper marks a significant milestone as the first demonstration of RL-based trigger control using actual LHC collision data.

Limitations and Future Work

  • Single Scalar Threshold: The current framework controls a single scalar threshold per trigger. This limits the ability to independently optimize for different signal types that might require different threshold settings. Future work could explore multi-dimensional threshold control.
  • Coupled Signal Efficiencies: The paper notes that signal efficiencies for different physics processes are coupled by the shared rate budget. While the reward mixing parameter offers some control, it's not a perfect solution.
  • Limited to CMS Data: The evaluation was primarily conducted on CMS Open Data. Generalization to other LHC experiments (like ATLAS or LHCb) is not yet tested.
  • Future Work: The authors suggest extending the framework to multi-trigger coordination with a joint threshold vector under a shared budget, which would be even more critical for feasibility-first filtering. They also point to the potential for this approach in other domains beyond HEP.

Summary updated

Abstract

High-throughput scientific facilities such as the Large Hadron Collider depend on real-time event filtering (triggering) under tight constraints on bandwidth, latency, and storage. In practice, trigger menus are largely static and hand-tuned and can become suboptimal as detector conditions, pileup, and background composition drift over time. We cast online threshold tuning as a sequential decision-making problem: a reinforcement learning agent ingests streaming summaries of recent rates and signal-sensitive features and updates trigger thresholds to maximize signal efficiency while tracking a target background rate within a tolerance band. We adapt Group-Filtered Policy Optimization (GFPO) to streaming control and introduce two variants (GFPO-F, GFPO-FR) that enforce background rate feasibility during training. On a benchmark that emulates realistic collider operation, we study two representative triggers: a total transverse energy (HT) trigger sensitive to pileup variation, and an anomaly-detection (AD) trigger based on reconstruction loss for rare or non-standard signatures. On Monte Carlo streams, our agent increases the fraction of in-tolerance time intervals by 48% (HT) and 28% (AD), with a cumulative gain of up to 2% in signal efficiency on those in-tolerance intervals. Transferring from simulation to real collision data (CMS Run 283408), the same agent, without fine-tuning, achieves a 56% (HT) and 28% (AD) in-tolerance improvement over baselines, with further signal-efficiency gain on both triggers. To our knowledge, this is the first demonstration of RL-based trigger control on real Large Hadron Collider collision data. Code is available at https://github.com/Zixind/GFPO_LHC.
BibTeX
@misc{ding2026learning,
  author = {Zixin Ding and Shaghayegh Emami and Giovanna Salvi and Cecilia Tosciri and Abhijith Gandrakota and Ngadiuba Jennifer and Nhan Tran and Christian Herwig and David Miller and Yuxin Chen},
  title = {{Learning to Trigger: Reinforcement Learning at the Large Hadron Collider}},
  year = {2026},
  eprint = {2606.23993},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2606.23993},
}
Figure