Crate parzen 

High-performance Tree-structured Parzen Estimator optimization.

use parzen::{
    CategoricalDistribution, Direction, Distribution, SearchSpace, Study,
    TpeSampler, TpeSamplerConfig,
};

let mut space = SearchSpace::new();
space.add("x", Distribution::Categorical(CategoricalDistribution::new(5)?))?;
let sampler = TpeSampler::new(TpeSamplerConfig::performance(42).startup_trials(5))?;
let mut study = Study::new(Direction::Maximize, sampler, space)?;

for _ in 0..20 {
    let x = study.suggest_categorical("x")?;
    study.complete_trial(if x == 2 { 1.0 } else { 0.1 })?;
}
assert!(study.best_value().is_some_and(|value| value > 0.5));

Explicit parameter groups use one trial-aligned mixture component for the entire vector. Their joint likelihood is logsumexp(log(weight[k]) + sum_d log(kernel[d][k](x[d]))), preserving correlations that independent marginal models discard. Integer and stepped distributions integrate each Gaussian kernel over the selected grid cell instead of treating a discrete value as a continuous point.

HistoryPolicy::Bounded keeps a fixed-size exact best set, recent bad observations, and a deterministic reservoir, so estimator state and incremental update work do not grow with completed-trial count. Raw trial records remain complete. HistoryPolicy::Full retains exact full-history ranking and therefore has linear storage and model-construction costs.

Structs

CategoricalDistribution

A categorical distribution represented by choice indices.

FloatDistribution

A bounded floating-point distribution.

GroupId

Stable explicit multivariate-group identifier.

IntDistribution

A bounded integer distribution.

ParamId

Stable parameter identifier within one search space.

ParameterRef

Read-only information about a registered parameter.

Params

Iterator over one trial’s parameter names and values.

SearchSpace

A validated collection of parameter distributions, conditions, and groups.

Study

A sequential Bayesian optimization study.

TpeSampler

Seeded Tree-structured Parzen Estimator sampler.

TpeSamplerConfig

Validated sampler configuration.

TrialId

Stable identifier assigned to a completed trial.

TrialInput

Owned input for injecting a completed trial.

TrialRecord

Owned, serializable representation of a completed trial.

TrialRef

Borrowed, allocation-free view of a completed trial.

Trials

Exact-size, double-ended iterator over completed trials.

Enums

Condition

A condition controlling whether a parameter is active.

Direction

Optimization direction.

Distribution

A categorical, floating-point, or integer search distribution.

FloatScale

Scale used by a floating-point distribution.

GammaStrategy

Strategy mapping applicable observation count to good-trial count.

HistoryPolicy

Amount of estimator history retained.

IntScale

Scale used by an integer distribution.

ModelStrategy

Independent or explicit-group multivariate modeling.

ParamValue

A typed parameter value.

ParzenError

An invalid configuration, search space, trial, or study operation.

WeightStrategy

Observation weighting strategy.