Module mixture 

Online mixtures of probabilistic predictors (log-loss Hedge / Bayes, switching, MDL).

This module provides a small, rigorously correct toolkit for sequential model mixing. Predictors expose per-symbol log-probabilities, which allows principled Bayesian mixture updates and clean information-theoretic accounting.

Rate-Backend Mixtures

The mixture primitives here power RateBackend::Mixture, enabling Bayes, fading Bayes, switching, and MDL-style selectors to be used anywhere a rate backend is accepted.

Structs

BayesMixture

Exponential-weights Bayes mixture (log-loss Hedge).

ExpertConfig

Configuration for a mixture expert.

FadingBayesMixture

Exponential-weights Bayes mixture with exponential forgetting on weights.

MdlSelector

MDL-style selector: predicts with the current best expert (by cumulative loss).

SwitchingMixture

Switching mixture: allows occasional switches between experts.

Enums

MixtureRuntime

RateBackendPredictor

A concrete online predictor backed by a RateBackend configuration.

Constants

DEFAULT_MIN_PROB

Default minimum probability floor to avoid log(0).

Traits

OnlineBytePredictor

Trait for online byte-level predictors that expose per-symbol log-probabilities.