Module local_search 

Host-side local-search refinement for memetic algorithms.

A memetic algorithm (MA) interleaves a population-level evolutionary strategy with a per-individual local search that polishes promising genomes between generations. This module defines the local-search seam — the LocalSearch trait — together with the four reference searchers (HillClimbing, NelderMead, SimulatedAnnealing, RandomRestart) that a MemeticWrapper can compose with any existing Strategy.

Design seam

Local search here is host-side and gradient-free. Searchers operate on a flat Vec<f32> genome and a single-member FitnessFn, never touching device tensors or computing gradients. This keeps refinement trivially reproducible and independent of the Burn backend: it is pure host arithmetic plus calls to the supplied fitness function.

Stochasticity convention

Every searcher takes a &mut dyn Rng and all randomness flows through it. Searchers never seed the process-wide backend RNG (B::seed + Tensor::random) and never reach for rand::rng() / thread-local RNGs — that would race the global Flex mutex under the parallel test runner and destroy reproducibility (see crate::rng for the host-RNG convention). A memetic wrapper derives a dedicated stream via seed_stream(base, generation, SeedPurpose::LocalSearch) and threads it in here.

Evaluation budget

Each refine call is bounded by Params::max_iters total FitnessFn::evaluate_one calls (including the mandatory first re-evaluation of the input genome). The shared BudgetedEval helper enforces this so the bound holds structurally even on flat landscapes where no probe ever improves.

refine_with_known_fitness skips that seeding eval when the caller already knows the input’s fitness, so the same max_iters then buys one extra probe evaluation. The seeding eval consumes no rng, so skipping it never shifts a searcher’s random stream.

Re-exports

pub use hill_climbing::HillClimbVariant;

pub use hill_climbing::HillClimbing;

pub use hill_climbing::HillClimbingParams;

pub use nelder_mead::NelderMead;

pub use nelder_mead::NelderMeadParams;

pub use random_restart::RandomRestart;

pub use random_restart::RandomRestartParams;

pub use simulated_annealing::CoolingSchedule;

pub use simulated_annealing::SimulatedAnnealing;

pub use simulated_annealing::SimulatedAnnealingParams;

Modules

hill_climbing

Coordinate-wise hill climbing.

nelder_mead

Nelder–Mead downhill simplex.

random_restart

Random-restart meta-search.

simulated_annealing

Simulated annealing.

Traits

LocalSearch

A gradient-free, host-side local search over real-valued genomes.