Bayesian optimization beats TPE for ACTS track-seeding parameters on Open Data Detector
Bayesian optimization improves ACTS seeding auto-tuning by locating high-performing configurations faster, supporting larger parameter spaces, and surfacing explicit trade-off surfaces that experts can select from after the fact. The work remains a single-detector, fixed-budget comparison on simulated ODD events.
The study replaced the existing Optuna TPE auto-tuner with Bayesian acquisition functions inside the full ACTS reconstruction chain. Under identical search ranges the two Bayesian methods reached competitive efficiency-fake-duplicate-runtime trade-offs earlier than TPE or random search; the best Bayesian run was then allowed to explore a fifteen-parameter space and finally multi-objective Expected Hypervolume Improvement was used to map the Pareto front without scalar weights. All candidate points were evaluated end-to-end and validated on disjoint events.
Lauren Tompkins: Within 18 months at least two LHC experiments will publish results replacing TPE-based ACTS tuning with Expected Hypervolume Improvement on real collision data.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2608.14714)
- [2]ACTS framework paper(https://doi.org/10.1016/j.nima.2021.165599)