KSLCore/ksl.utilities.distributions.fitting.mixture.scoring/MixtureCriterionIfc

MixtureCriterionIfc

A criterion by which fitted mixtures are ranked.

Implementations must obtain the number of free parameters from the candidate rather than from an assembled mixture distribution. The library's mixture distribution reports one parameter per mixing weight, but only one fewer than that number are free, so reading the count from the assembled distribution overstates the dimension.

Implementations must also screen the log-likelihood for finiteness before using it. The library's information criterion helpers reject a non-finite log-likelihood by throwing, and a mixture produces one routinely: whenever a component assigns zero density to an observation it was not fitted to. Turning an expected modelling outcome into an exception mid-experiment is not acceptable, so the value is bounded and marked incomparable instead.

Inheritors

LogLikelihoodCriterion

HeldOutLogLikelihoodCriterion

Properties

name

abstract val name: String

A short name for the criterion, used as a column heading and a factor level.

smallerIsBetter

open val smallerIsBetter: Boolean

Whether smaller values indicate a better fit. Information criteria are smaller-is-better; a held-out mean log-likelihood is not. Ranking machinery must consult this rather than assuming a direction.

Functions

evaluate

abstract fun evaluate(candidate: MixtureCandidate, data: DoubleArray): MixtureCriterionValue

Evaluates the criterion for the supplied candidate against the supplied observations.

metric

abstract fun metric(): MetricIfc

A metric describing this criterion, for use with the library's multi-objective decision analysis machinery. The direction is taken from this interface.