Expand description
Topology-evolving neuroevolution (NEAT) — the host-side graph data model.
This module holds the building blocks shared by NEAT: the graph genome
(topology), the per-run innovation registry (innovation), speciation
(species), and the phenotype builders (phenotype). The
NeatStrategy custom
harness that drives them lives in crate::algorithms::neuroevolution::neat.
These types are graphs, not tensors: unlike the
Strategy genomes elsewhere in the crate,
TopologyGenome is plain host-side data (and is Clone). A genome is
scored either one at a time, through the interpreted InterpretedBuilder
reference path, or a whole population at once, through the device-batched
DensePaddedEvaluator; the two agree to float epsilon.
§Orientation
NEAT is maximization (higher fitness is better) — matching the crate-wide
maximise convention. Species best/stagnation tracking, fitness sharing,
and the GraphFitnessFn
seam all treat higher as better; a cost objective is reconciled into
canonical space by the harness/adapter chokepoint, not by hand here.
Re-exports§
pub use innovation::InnovationRegistry;pub use innovation::NodeSplit;pub use phenotype::BatchPhenotypeEvaluator;pub use phenotype::DensePaddedEvaluator;pub use phenotype::InterpretedBuilder;pub use phenotype::InterpretedPhenotype;pub use phenotype::Phenotype;pub use phenotype::PhenotypeBuilder;pub use species::Species;pub use species::SpeciesId;pub use species::allocate_offspring;pub use species::compatibility_distance;pub use species::remove_stagnant;pub use species::speciate;pub use topology::ActivationFn;pub use topology::ConnectionGene;pub use topology::InnovationId;pub use topology::NodeGene;pub use topology::NodeId;pub use topology::NodeKind;pub use topology::TopologyGenome;
Modules§
- innovation
- Innovation-number bookkeeping — the per-run registry that assigns the historical markers NEAT crossover aligns on.
- phenotype
- Phenotype construction — turning a
TopologyGenomeinto a callable network. - species
- Speciation — compatibility distance, the representative-assignment pass, fitness sharing, stagnation, and offspring apportionment.
- topology
- NEAT topology genome — a direct graph encoding of node and connection genes.