RL Agent Optimizes Event Triggering at Large Hadron Collider

High-energy physics experiments, like those at the Large Hadron Collider (LHC), generate immense amounts of data, necessitating real-time event filtering or "triggering" under strict bandwidth and storage constraints. Current trigger systems are often static and manually tuned, becoming suboptimal as experimental conditions change. This study frames online trigger threshold tuning as a sequential decision-making problem, using a reinforcement learning (RL) agent. The agent processes streaming summaries of data rates and signal-sensitive features, then adjusts trigger thresholds to maximize signal efficiency while keeping background rates within a specified tolerance. The adapted Group-Filtered Policy Optimization (GFPO) method, with new variants, was tested on a benchmark emulating collider operations and then on real collision data from the CMS experiment. The RL agent increased the time intervals within tolerance by 48% for a transverse energy trigger and 28% for an anomaly detection trigger, yielding cumulative signal efficiency gains. This groundbreaking work represents the first successful application of RL for trigger control on actual LHC collision data.