Struct BayesianSampler 

pub struct BayesianSampler { /* private fields */ }

Bayesian optimization sampler using Tree-Parzen Estimator (TPE).

For the first n_startup trials, samples randomly. After that, uses the history of (params, metric) to model “good” vs “bad” parameter distributions and samples from the “good” distribution.

This is a simplified TPE: it splits trials into top/bottom quantiles and samples from the top quantile’s parameter distributions.

Implementations

impl BayesianSampler

pub fn new(n_trials: usize, n_startup: usize, seed: Option<u64>) -> Self

A TPE sampler producing n_trials configurations, the first n_startup (floored at 2 — TPE needs history to split) sampled randomly. None seed defaults to 42, keeping runs reproducible.

pub fn record(&mut self, params: HashMap<String, Value>, metric: f64)

Record a completed trial’s result (for informing future samples).

Trait Implementations

impl Sampler for BayesianSampler

fn record_result(&mut self, params: &HashMap<String, Value>, value: f64)

Completed-trial feedback — this is what makes TPE model-based instead of degenerating to random search.

fn sample( &mut self, space: &SearchSpace, trial_index: usize, ) -> Result<Option<HashMap<String, Value>>>

Sample the next set of parameters. Returns None when exhausted.

fn n_trials(&self) -> Option<usize>

Total number of trials this sampler will produce (if known).

fn prepare(&mut self, _space: &SearchSpace)

Called once before the trial loop with the resolved search space. Lets samplers precompute state — e.g. GridSampler resolves its dimension grid here so n_trials is correct before the first sample.