Struct EvolutionSMC 

pub struct EvolutionSMC;

The tempered-SMC evolution driver.

Implementations

impl EvolutionSMC

pub fn run<P, L, R>( rng: &mut R, model: &EvolutionModel<P, L>, cfg: EvoSmcConfig, ) -> EvolutionPosterior<P::Genome>

where P: GenomePrior, L: GenomeLikelihood<P::Genome>, R: Rng,

Run tempered SMC targeting the Boltzmann posterior π ∝ p(x)·exp(f(x)) of model (β is supplied by fugue’s adaptive tempering; model’s own β setting is ignored here by construction).

cfg.crossover builds a structure-safe SharedSiteCrossover kernel (see CrossoverConfig for what it can and cannot exchange on a variable-structure prior); None runs per-particle rejuvenation only.

impl EvolutionSMC

pub fn run_with_kernel<P, L, R, K>( rng: &mut R, model: &EvolutionModel<P, L>, cfg: EvoSmcConfig, kernel: &mut K, ) -> EvolutionPosterior<P::Genome>

where P: GenomePrior, L: GenomeLikelihood<P::Genome>, R: Rng, K: PopulationKernel<P::Genome>,

Like EvolutionSMC::run, but with an explicit population kernel (e.g. a [fugue::CrossoverKernel] with a subtree_crossover_mask for grammar-driven tree genomes). cfg.crossover is ignored.

impl EvolutionSMC

pub fn anneal<P, L, R>( rng: &mut R, model: &EvolutionModel<P, L>, cfg: EvoSmcConfig, beta_max: f64, anneal_steps: usize, ) -> EvolutionPosterior<P::Genome>

where P: GenomePrior, L: GenomeLikelihood<P::Genome>, R: Rng,

Optimizer mode: run tempered SMC to the posterior (β = 1), then keep annealing the ladder toward beta_max, concentrating the population on the maximizers of the likelihood/fitness.

The continuation is built from fugue’s exported primitives and keeps every invariant of the tempering loop: at each rung the particles are incrementally reweighted by Δβ·(log_likelihood + log_factors), normalized, systematically resampled to uniform weights, and rejuvenated with π_β-invariant MH (plus the structure-safe SharedSiteCrossover sweep when cfg.crossover is set — see CrossoverConfig; for a model-aware kernel such as subtree_crossover_mask use EvolutionSMC::anneal_with_kernel). The rung schedule is geometric from 1 to beta_max over anneal_steps rungs.

The returned population approximates π_{β_max} ∝ p(x)·L(x)^{β_max}, which for large beta_max concentrates on the optima — a principled, uncertainty-aware replacement for a classic GA on single-objective problems. log_evidence reflects only the β ≤ 1 ladder (evidence is defined at the posterior).

With rejuvenation_steps == 0 and no kernel nothing moves a particle past β = 1: each rung only reweights and resamples, so the population collapses onto duplicates of the fittest posterior particles. Keep at least one rejuvenation step (or a kernel) when annealing.

pub fn anneal_with_kernel<P, L, R, K>( rng: &mut R, model: &EvolutionModel<P, L>, cfg: EvoSmcConfig, beta_max: f64, anneal_steps: usize, kernel: &mut K, ) -> EvolutionPosterior<P::Genome>

where P: GenomePrior, L: GenomeLikelihood<P::Genome>, R: Rng, K: PopulationKernel<P::Genome>,

Like EvolutionSMC::anneal, but with an explicit population kernel applied both inside the β ≤ 1 ladder and at every annealing rung — the optimizer-mode counterpart of EvolutionSMC::run_with_kernel (e.g. a [fugue::CrossoverKernel] with a subtree_crossover_mask to anneal a grammar prior with subtree grafts). cfg.crossover is ignored; pass [fugue::NoKernel] for rejuvenation only.