Module bayesian 

Bayesian building blocks and an online operator-parameter tuner.

This module provides:

• Honest conjugate posteriors that say what they are: BetaPosterior (Beta-Bernoulli over a probability), GammaPosterior (Gamma-Exponential over a positive rate), and RunningLogMoments (a moment tracker for log-scale quantities — not a Bayesian posterior, and named accordingly).
• A ThompsonSamplingTuner: a genuine multi-armed-bandit tuner for GA operator parameters. Each tunable parameter is discretized into candidate values (“arms”); each arm carries a Beta posterior over P(offspring improves on parents | this value is used). Every generation an arm is Thompson-sampled per parameter and its value is applied to the operators; the observed improvement events are credited back to the arm that produced them.

The Beta draw taken during Thompson sampling is used only to pick an arm; it is never returned as the parameter value itself. This is the correct separation that the previous BayesianHyperparameterLearner got wrong (it sampled a P(improvement) posterior and used the draw as the mutation rate).

Structs

BanditArm

One discretized candidate value for a tunable parameter, with a Beta posterior over P(offspring improves on parents | this value is used).

BanditParameter

A single tunable parameter, discretized into arms and tuned by Thompson sampling over each arm’s Beta posterior.

BetaPosterior

Beta distribution posterior for a probability parameter (Beta-Bernoulli).

GammaPosterior

Gamma posterior for the rate λ of an Exponential(λ) likelihood.

RunningLogMoments

Running mean/variance of ln(x) for positive quantities such as step sizes.

ThompsonConfig

Configuration for a ThompsonSamplingTuner wired into a GA.

ThompsonSamplingTuner

Thompson-sampling multi-armed-bandit tuner over GA operator parameters.

TunerSnapshot

A snapshot of a tuner’s per-arm posterior means at one generation.

Constants

PARAM_CROSSOVER_PROB

Canonical parameter name for the whole-genome crossover probability arm set.

PARAM_MUTATION_RATE

Canonical parameter name for the per-gene mutation probability arm set.

Traits

TunableMutation

Mutation operators whose per-gene mutation probability can be set by an online tuner such as ThompsonSamplingTuner.