kumulant

warmStart

fun BernoulliArm.Companion.warmStart(global: BernoulliSumResult, shrinkage: Double = 1.0): BernoulliArm(source)

Warm-started BernoulliArm from a global Bernoulli snapshot.

Builds an Arm spec from a pooled global snapshot. shrinkage in [0, 1] scales how much of the global evidence is counted as the arm's prior pseudo-count: 0 collapses to the bare prior; 1 treats every global observation as if it had been seen by this arm. The same convention applies to every warmStart overload below.

Univariate hierarchical pooling has no clean general API (see the discussion next to com.eignex.kumulant.bandit.contextual.RegressionContextualBandit for the contextual case); these per-arm-type helpers cover the Result types where shrinkage has a uniform pseudo-count interpretation.

fun MeanArm.Companion.warmStart(global: WeightedMeanResult, shrinkage: Double = 1.0): MeanArm(source)

Warm-started MeanArm from a global weighted-mean snapshot.

fun NormalArm.Companion.warmStart(global: WeightedVarianceResult, shrinkage: Double = 1.0): NormalArm(source)

Warm-started NormalArm from a global weighted-variance snapshot. The arm's prior variance is preserved from the global; only the prior weight is shrunk.

fun LogNormalArm.Companion.warmStart(global: WeightedVarianceResult, shrinkage: Double = 1.0): LogNormalArm(source)

Warm-started LogNormalArm from a global weighted-variance snapshot on the log scale (caller is responsible for ensuring the snapshot is over ln(reward)).

fun MomentsArm.Companion.warmStart(global: MomentsResult, shrinkage: Double = 1.0): MomentsArm(source)

Warm-started MomentsArm from a global moments snapshot.