Struct CmsaEsState 

pub struct CmsaEsState<B: Backend> { /* private fields */ }

Generation state for CmsaEs.

All adaptive quantities live here (not in CmsaEsConfig) so instances stay lock-free across parallel runs. Linear-algebra state — the mean and covariance — is held host-side as Vec<f32>; only the offspring population crosses to the device.

Implementations

impl<B: Backend> CmsaEsState<B>

pub fn try_new( mean: Vec<f32>, cov: Vec<f32>, sigma: f32, offspring_sigmas: Vec<f32>, generation: usize, best_genome: Option<Tensor<B, 2>>, best_fitness: f32, ) -> Result<Self, ConfigError>

Assembles a CMSA-ES state, checking the distribution parameters are dimensionally consistent and normalizing cov to exact symmetry.

The supplied cov is symmetrized in place via crate::ops::linalg::symmetrize before construction. The in-loop covariance blend in tell already preserves bit-exact symmetry — IEEE-754 multiplication is commutative and the two triangle entries C[i,j] / C[j,i] accumulate the identical rank-μ terms in the identical order — so caller-supplied construction is the only asymmetry entry point. Normalizing it here mirrors pycma practice and the ADR 0034 sanitize-at-chokepoint convention.

Errors

Returns a ConfigError if mean is empty, if cov is not D × D row-major (D = mean.len()), or if sigma is not strictly positive and finite. No length constraint is imposed on offspring_sigmas: it may be empty (the pre-ask state) or any length — tell falls back to sigma for any missing entry.

pub fn mean(&self) -> &[f32]

Distribution mean m, length D.

pub fn cov(&self) -> &[f32]

Covariance matrix C, row-major D × D.

pub fn sigma(&self) -> f32

Global step size σ̄.

pub fn offspring_sigmas(&self) -> &[f32]

Per-offspring step sizes σᵢ, carried ask → tell (empty before the first ask).

pub fn generation(&self) -> usize

Completed-generation counter.

pub fn best_genome(&self) -> Option<&Tensor<B, 2>>

Best-so-far genome (shape (1, D)), or None before the first tell.

pub fn best_fitness(&self) -> f32

Best-so-far (canonical, maximise) fitness.

Trait Implementations

impl<B: Clone + Backend> Clone for CmsaEsState<B>

fn clone(&self) -> CmsaEsState<B>

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<B: Debug + Backend> Debug for CmsaEsState<B>

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

Formats the value using the given formatter.