Struct BayesianNetwork 

pub struct BayesianNetwork;

Bayesian-network model for binary spaces (BOA).

Implements ProbabilityModel by greedily learning a bounded-in-degree DAG over the binary genes with a BIC structure score, then ancestral-sampling from the fitted CPTs (see the module docs for the algorithm, the BIC rationale, bit-packing, and references). Fitness is accepted but ignored; the fit is always unweighted and non-incremental.

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

Trait Implementations

impl Clone for BayesianNetwork

fn clone(&self) -> BayesianNetwork

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for BayesianNetwork

impl Debug for BayesianNetwork

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

Formats the value using the given formatter.

impl Default for BayesianNetwork

fn default() -> BayesianNetwork

Returns the “default value” for a type.

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

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 Bayesian network to the selected population.

When prev = None returns the edgeless prior (natural order, empty parent lists, single-cell CPTs at init_prob); population and fitness are ignored. Otherwise the whole network is relearned from scratch (the fit is non-incremental — prev is the not-bootstrap signal only):

1. Bitizes the selected rows to {0, 1} via >= 0.5.
2. Greedily adds the highest-BIC-gain edge each round (subject to the max_parents cap and acyclicity) until no strictly-positive gain remains, using a D × D gain cache.
3. Estimates Laplace-smoothed CPTs from the final structure.
4. Computes a deterministic topological order (Kahn, min-index).

The fitness argument is accepted but always ignored.

Panics

In release builds, panics if the population tensor cannot be read back as f32 (.expect("population tensor must be readable as f32")).

In debug builds, additionally panics on the following debug_assert checks (all disabled in release):

• the population column count disagrees with params.genome_dim;
• params.max_parents >= usize::BITS (which would overflow the 1usize << q CPT table sizing);
• the closing topological order fails to cover all D nodes (guaranteed by the DAG invariant).

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

Draw n binary genomes by ancestral sampling along the topological order.

Each gene v is sampled from P(v = 1 | parents) read out of its CPT at the bit-packed parent configuration (parents already sampled, since the traversal follows the topological order). Exactly one rng.random::<f32>() call is consumed per gene regardless of structure, keeping RNG consumption stable. Host RNG only (never Tensor::random / B::seed). The returned tensor has shape (n, D) and contains only 0.0 and 1.0.

type Params = BayesianNetworkParams

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

type State = BayesianNetworkState

Evolving fitted statistics carried generation to generation.