Expand description
Tensor-native classical evolutionary algorithms for rlevo.
This crate ships the classical EA families — Genetic Algorithm (GA),
Evolution Strategy (ES), Evolutionary Programming (EP), Differential
Evolution (DE), and Cartesian Genetic Programming (CGP) — on top of the
Burn tensor abstraction, with GPU acceleration via custom CubeCL kernels
on hot paths.
§Surface area
strategy— the centralStrategytrait and theEvolutionaryHarnessadapter that wraps any strategy intorlevo-core::evaluation::BenchEnv.genome— zero-sized marker types (Real,Binary,Integer,Tree,Permutation) that parameterize the operator set.population—Population<B, K>, a thin wrapper aroundTensor<B, 2>carrying shape metadata.fitness—FitnessFn/BatchFitnessFn, theFromFitnessEvaluableadapter forrlevo-core::fitness::FitnessEvaluable, and theFromLandscapeadapter for landscapes that carry their ownevaluatemethod.observer—PopulationObserver/PopulationSnapshot/SharedPopulationObserver: structured per-generation callback for recorders that need more than the scalartracingevents (full fitness vector, best-individual index, lineage).rng— deterministic seed streams (splitmix64) for reproducibility.shaping— fitness shaping transforms (centered rank, z-score).ops— selection, crossover, mutation, and replacement operators.algorithms— concrete strategies.
Re-exports§
pub use observer::PopulationObserver;pub use observer::PopulationSnapshot;pub use strategy::EvolutionaryHarness;pub use strategy::Strategy;pub use strategy::StrategyMetrics;
Modules§
- algorithms
- Concrete evolutionary algorithms.
- fitness
- Fitness evaluation traits and adapters.
- genome
- Genome category trait and its zero-sized marker types.
- observer
- Per-generation observer surface for
EvolutionaryHarness. - ops
- Evolutionary operators.
- population
- Population containers.
- rng
- Deterministic seed derivation for strategies.
- shaping
- Fitness shaping transforms.
- strategy
- Central
Strategytrait and theEvolutionaryHarnessadapter.