Enum Searcher 

pub enum Searcher {
    Random,
    Sobol,
    Tpe,
    Cmaes,
    IsotropicEs,
}

Which searcher a run drives.

The three from the sampler-quality gate (SamplerKind) plus the two CMA-ES-specific ones: the real [Cmaes] and the covariance-free ablation (IsotropicEs) its regression sentinel measures against.

Sobol, TPE and CMA-ES run their earlier defaults (each promoted setting’s with_*(false): sorted-name QMC, univariate TPE, and CMA-ES without Optuna initialization). Those are the configurations the oracle’s measured table and floors were taken on, and the ones the rotation argument needs: the default TPE is multivariate, so it can represent a rotated valley, and the ablation mirrors the earlier CMA-ES initialization.

Variants

Random

Uniform independent sampling — the baseline, and axis-aligned by definition.

Sobol

The Sobol low-discrepancy sequence — space-filling, and also axis-aligned.

Tpe

The univariate Tree-structured Parzen Estimator (the earlier default, every promoted setting off): the strong learning baseline, whose densities are per-dimension, so it has no representation of a rotated valley.

Cmaes

The real CMA-ES, without Optuna initialization (the earlier default).

IsotropicEs

CMA-ES with the covariance update deleted (C ≡ I) — the sentinel.

Implementations

impl Searcher

pub const ALL: [Self; 5]

Every searcher, in a fixed order.

pub const fn label(self) -> &'static str

The lowercase label used in reports and assertion messages.

pub const fn models_dimensions_independently(self) -> bool

Whether this searcher models the dimensions independently — i.e. whether its search distribution has no off-diagonal terms.

This is the property the rotated fixture attacks, and it is true for four of the five: uniform random and Sobol draw each coordinate on its own, TPE fits a 1-D Parzen density per dimension, and the ablation’s covariance is the identity by construction. Only [Cmaes] can represent a valley that runs at an angle to the axes.

Note the ablation is isotropic, not literally axis-preferring — a spherical distribution has no favourite direction at all. What it shares with the other three is the thing that matters here: no way to encode a correlation between two parameters.

pub fn build(self) -> Arc<dyn Sampler>

Builds a fresh sampler of this kind.

Trait Implementations

impl Clone for Searcher

fn clone(&self) -> Searcher

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for Searcher

impl Debug for Searcher

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Eq for Searcher

impl PartialEq for Searcher

fn eq(&self, other: &Searcher) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for Searcher