kumulant
kumulant/com.eignex.kumulant.bandit.univariate/CompositeArm

CompositeArm

@Serializable
@SerialName(value = "CompositeArm")
data class CompositeArm(val subArms: List<CompositeSubArm>) : Arm<ResultList<Result>> (source)

Composite Arm built from N independent sub-arms. Each observation is routed through every sub-arm's CompositeSubArm.valueExpr / CompositeSubArm.weightExpr / CompositeSubArm.filter AST before being fed to the corresponding per-sub-arm accumulator. The composite result is a ResultList of the sub-snapshots; pair with CompositePosterior to combine sub-arm draws into a single score.

Plumbing is AST-driven (no lambdas) so the whole spec is wire-serializable. Per-sub-arm encode is still applied after valueExpr; e.g. with LogNormalArm

  • valueExpr = X the lognormal's own encode = ln runs, so don't double-log via valueExpr = Log(X).

Example; zero-inflated lognormal revenue:

val ziln = CompositeArm(listOf(<br>    CompositeSubArm(BernoulliArm(), valueExpr = IfExpr(X gt 0.0, Const(1.0), Const(0.0))),<br>    CompositeSubArm(LogNormalArm(), filter = X gt 0.0),<br>))<br>val score = CompositePosterior(<br>    subPosteriors = listOf(BetaPosterior, LogNormalGammaPosterior),<br>    combine = V(0) * V(1),<br>)<br>val bandit = MultiArmedBandit(nbrArms = 4, policy = ThompsonSampling(ziln, score))
Content copied to clipboard

Constructors

Link copied to clipboard
constructor(subArms: List<CompositeSubArm>)

Types

Link copied to clipboard
object Companion

Companion host for factory helpers.

Properties

Link copied to clipboard
val subArms: List<CompositeSubArm>

Sub-arms whose stats receive routed observations.

Functions

Link copied to clipboard
open override fun createStat(): SeriesStat<ResultList<Result>>

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

Link copied to clipboard
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.

CompositeArm

constructor(subArms: List<CompositeSubArm>)(source)

createStat

open override fun createStat(): SeriesStat<ResultList<Result>>(source)

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

subArms

val subArms: List<CompositeSubArm>(source)

Sub-arms whose stats receive routed observations.