Struct BatchGraphFitness 

pub struct BatchGraphFitness<B: Backend, E> { /* private fields */ }

A GraphFitnessFn that scores the whole population in one device-resident pass via a BatchPhenotypeEvaluator, instead of the per-genome interpreted loop.

It is a drop-in alternative to an interpreted GraphFitnessFn in the same manual init → ask → evaluate → tell loop: the builder argument is ignored (the batched evaluator owns evaluation), so the harness stays evaluation-agnostic and NeatStrategy’s determinism is untouched.

The reducer maps one genome’s batch × action_dim output slab (row-major, as produced by BatchPhenotypeEvaluator::evaluate_population) to a single maximization fitness — the batched analogue of the per-genome scoring an interpreted GraphFitnessFn does on phenotype.forward(...).

Implementations

impl<B: Backend, E> BatchGraphFitness<B, E>

pub fn new( evaluator: E, obs: Tensor<B, 2>, reducer: impl Fn(&[f32]) -> f32 + Send + Sync + 'static, ) -> Self

Build a batched fitness from an evaluator, a shared [batch, obs_dim] observation tensor, and a per-genome reducer over the batch × action_dim output slab (row-major).

Trait Implementations

impl<B: Backend, E> Debug for BatchGraphFitness<B, E>

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

Formats the value using the given formatter.

impl<B: Backend, E: BatchPhenotypeEvaluator<B>> GraphFitnessFn<B> for BatchGraphFitness<B, E>

fn evaluate( &self, population: &[TopologyGenome], _builder: &dyn PhenotypeBuilder<B>, device: &<B as BackendTypes>::Device, ) -> Vec<f32>

Score every genome in population, returning one maximization fitness per genome (higher is better, matching the crate-wide canonical convention). Fitness sharing assumes the values are non-negative. The returned Vec has one entry per input genome, in order. Build each genome’s network with builder on device.