Abstract Every built-in pruner is a heuristic, and none states how often it prunes a trial that would have ended well. This pruner does. You pick alpha, and the expected fraction of trials that were good and pruned stays at or below alpha, under one condition: the trials it calibrates on and the trials it judges must come from the same random process.
It runs the first n_calibration completed trials unpruned, calibrates a threshold on their learning curves by conformal risk control, then freezes and prunes.
Abstract This sampler optimizes the objective function given input perturbations. For example, most industrial productions have their own tolerance, and any differences within this tolerance are considered acceptable. For this reason, even if we transfer the optimized result of a design simulation to its production, it is mostly impossible to reproduce the precise simulated design setup. This necessitates accounting for noisy input so that the deployed setup remains sufficiently performant even in the presence of arbitrary noise.
Abstract TuRBOSampler implements Bayesian optimization with trust regions. It places local trust regions around the current best solutions and fits Gaussian Process (GP) models within those regions. Operating within adaptive local regions reduces high-dimensional sample complexity, yielding accurate fits with fewer trials.
Please refer to the paper, Scalable Global Optimization via Local Bayesian Optimization for more information.
APIs TuRBOSampler(*, n_startup_trials: int = 4, n_trust_region: int = 5, success_tolerance: int = 3, failure_tolerance: int = 5, seed: int | None = None, independent_sampler: BaseSampler | None = None, deterministic_objective: bool = False, warn_independent_sampling: bool = True) n_startup_trials: Number of initial trials PER TRUST REGION.