Struct DependencyChain 

pub struct DependencyChain;

MIMIC-style dependency-chain model for continuous spaces.

Implements ProbabilityModel by fitting a greedy first-order dependency chain over dimensions (see module docs for the algorithm, estimator regularisation, and references). Fitness is accepted but ignored; the fit is always unweighted.

fit is O(k · D²); sample is O(D) per individual.

Trait Implementations

impl Clone for DependencyChain

fn clone(&self) -> DependencyChain

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for DependencyChain

impl Debug for DependencyChain

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

Formats the value using the given formatter.

impl Default for DependencyChain

fn default() -> DependencyChain

Returns the “default value” for a type.

impl<B: Backend> ProbabilityModel<B> for DependencyChain

fn fit( &self, params: &Self::Params, prev: Option<&Self::State>, population: Tensor<B, 2>, fitness: Tensor<B, 1>, device: &<B as BackendTypes>::Device, ) -> Self::State

Fit the dependency-chain model to the selected population.

When prev = None returns the prior (natural-order chain, uniform init_mean / init_std, zero correlations). Otherwise:

1. Computes per-dimension MLE means and standard deviations (÷k variance, floored at min_variance).
2. Builds the full D × D Pearson correlation matrix.
3. Applies the |r| < 2/√k significance filter (see module docs for the estimator regularisation rationale) and clamps surviving correlations to [−0.9999, 0.9999].
4. Converts to mutual information MI = −0.5 · ln(1 − r²).
5. Builds the chain greedily: root = minimum-σ dimension, each subsequent link = unvisited dimension with the highest MI to the last chosen one.

The fitness argument is accepted but always ignored.

Panics

Panics if the population tensor cannot be read back as f32 (.expect("population tensor must be readable as f32")), or with an out-of-bounds index if the host buffer is shorter than k * d. Callers must therefore pass an f32, (k, d)-shaped population tensor. The unwrap() on the last greedy-chain iteration does not fire: at least one unvisited dimension remains when d > 1.

fn sample( &self, state: &Self::State, n: usize, rng: &mut dyn Rng, device: &<B as BackendTypes>::Device, ) -> Tensor<B, 2>

Draw n genomes by ancestral sampling along the fitted chain.

The root dimension is sampled from its marginal Gaussian; each subsequent dimension is sampled from the conditional Gaussian given its parent’s sampled value:

μ_cond = μ_c + r · (σ_c / σ_p) · (x_parent − μ_p)
σ_cond = σ_c · √(1 − r²)

All randomness is drawn from rng (host RNG only; never Tensor::random / B::seed). The returned tensor has shape (n, D). This is O(D) per individual drawn.

Panics

Does not panic under normal operation. The Normal::new calls are guarded: σ_c is floored at min_variance.sqrt() during fit, and r is clamped to [-0.9999, 0.9999] so 1 − r² ≥ 0.0002 > 0.

type Params = DependencyChainParams

Static, per-run model configuration (learning rates, priors, …).

type State = DependencyChainState

Evolving fitted statistics carried generation to generation.