kumulant/com.eignex.kumulant.bandit.univariate/NormalArm

NormalArm

@Serializable

@SerialName(value = "NormalArm")

data class NormalArm(val priorMean: Double = 0.0, val priorWeight: Double = 0.02, val priorSquaredDeviations: Double = 0.02) : Arm<WeightedVarianceResult> (source)

Gaussian arm with a Normal-Gamma prior (unknown mean and variance). Tracks both the running mean and the sum of squared deviations, which gives NormalGammaPosterior enough to draw (mean, variance) jointly and gives the variance-aware policies (Greedy, EpsilonGreedy) a reasonable variance estimate.

The prior is parameterised as a Normal-Gamma:

priorMean is the prior mean reward.
priorWeight is the pseudo-weight of the prior: how many "phantom observations" the prior counts as. Higher = stronger prior, slower to move.
priorSquaredDeviations is the prior sum of squared deviations; tightens the prior on the variance. Higher = stronger belief that variance is large.

The default prior is mildly informative; priorWeight = 0.02 washes out after a few observations.

For multiplicative rewards (revenue, latency, anything where noise is log-normal rather than additive Gaussian), use LogNormalArm instead.

Constructors

NormalArm

constructor(priorMean: Double = 0.0, priorWeight: Double = 0.02, priorSquaredDeviations: Double = 0.02)

Types

Companion

object Companion

Factory entry-point for NormalArm (host for warmStart).

Properties

priorMean

val priorMean: Double

Prior mean reward.

priorSquaredDeviations

val priorSquaredDeviations: Double

Prior sum of squared deviations; tightens the prior on the variance.

priorWeight

val priorWeight: Double

Pseudo-weight of the prior seed.

Functions

createStat

open override fun createStat(): SeriesStat<WeightedVarianceResult>

Allocate a fresh per-arm accumulator already seeded with this arm's prior pseudo-counts.

encode

open fun encode(value: Double): Double

Map a raw observation onto the scale the stat accumulates. Identity by default; LogNormalArm overrides with ln so the underlying stat tracks the log-reward and the Normal-Gamma posterior fits the log-normal generative model.