獲得関数#

Layer 2: Compose

基底#

saealib.AcquisitionFunction

Abstract base class for acquisition functions (infill criteria).

実装#

saealib.ExpectedImprovement

Expected Improvement (EI) acquisition function.

saealib.BatchExpectedImprovement

Monte Carlo qEI with candidate-joint covariance and greedy set utility.

saealib.LowerConfidenceBound

Lower Confidence Bound (LCB) acquisition function.

saealib.MaxUncertainty

Acquisition function that maximizes predictive uncertainty (exploration).

saealib.MeanPrediction

Acquisition function based on predicted mean value (exploitation).

saealib.CORSDistance

CORS distance-constrained predicted-mean acquisition function.

saealib.acquisition.CORSReference

Prepared archive reference and beta for one CORS decision.

saealib.ProbabilityOfFeasibility

Probability of Feasibility (PoF) acquisition function.

saealib.ProductOfFeasibility

Joint probability of feasibility across all constraint columns.

saealib.EHVIAcquisition

Expected Hypervolume Improvement (EHVI) acquisition function.

saealib.ParEGOAcquisition

ParEGO acquisition function for multi-objective Bayesian optimisation.

saealib.SMSEGOAcquisition

SMS-EGO acquisition function for multi-objective Bayesian optimisation.
saealib.gp_ucb_beta_schedule(domain_size, delta=0.1)[ソース]#

Return the finite-domain GP-UCB beta_t schedule for use with beta_schedule.

Implements the finite-domain schedule beta_t = 2 log(domain_size * t^2 * pi^2 / (6 * delta)) [25], where domain_size is the cardinality |D| of a finite decision set. This helper does not implement the cited paper's separate continuous-domain schedule (a different formula, for a compact domain). Passing a domain_size for a continuous search space only substitutes a number into this finite-domain formula; it does not by itself grant the finite-domain theorem's regret guarantee -- see the "GP-UCB" algorithm page in the docs for the additional conditions that requires.

パラメータ:

  • domain_size (int) -- Cardinality |D| of the finite decision set (the cited bound's |D|). Must be an int and positive.

  • delta (float) -- Confidence parameter. Must satisfy 0 < delta < 1. Default: 0.1.

戻り値:

A picklable schedule(t) -> float, increasing in the round index t (t must be an int starting at 1; raises ValidationError for a non-integer t or t < 1).

戻り値の型:

callable

例外:

ValidationError -- If domain_size is not an int or positive, or delta is not in (0, 1).

CORSDistance#

CORSDistanceは、予測平均のスコアにCORSの距離制約を課します。prepare(archive, ctx)はctx.decision_countからbetaを選び、評価済み点と合わせて保持するCORSReferenceを返します。そのため最初の判断では探索パターンの先頭要素が使われ、以降の判断はパターンを巡回します。原法に忠実な標準構成では、判断1回につき1点だけを真の評価に送ります。たとえばPreSelectionStrategy(..., n_select=1)やTopKEvaluation(1)を使います。

delta=Noneのとき、距離スケールは現在の候補プールから、評価済みアーカイブへの最小距離の最大値として近似されます。固定のスケールを使いたい場合は、有限の正のdeltaを渡します。