rlevo_evolution/

fitness.rs

//! Fitness evaluation traits and adapters.
//!
//! Two traits model the two evaluation shapes strategies expect:
//!
//! - [`FitnessFn`] — evaluates a single member. Callers hand the fitness
//!   function a host-side genome row (typically `Vec<f32>`) and receive a
//!   scalar. Useful for simple benchmarks and for unit-testing operators.
//! - [`BatchFitnessFn`] — evaluates an entire population in one call and
//!   returns a device-resident `Tensor<B, 1>` of shape `(pop_size,)`. This
//!   is the hot path — strategies call it once per generation.
//!
//! Two adapters bridge host-side scalar fitness code into [`BatchFitnessFn`]:
//!
//! - [`FromFitnessEvaluable`] — wraps
//!   `rlevo-core::fitness::FitnessEvaluable<Individual = Vec<f64>, Landscape = L>`.
//!   Use this when an evaluator and a landscape type are already defined
//!   separately (e.g. `RastriginEvaluator` + `RastriginLandscape`).
//! - [`FromLandscape`] — wraps `rlevo-core::fitness::Landscape` directly.
//!   Use this when the landscape is self-evaluating (Sphere, Ackley, Rastrigin)
//!   and a separate evaluator shim would add no value.
//!
//! Both adapters pull each population row to host as `f32`, widen to `f64`,
//! evaluate on the CPU, and re-upload the results as a `Tensor<B, 1>`.
//! Purpose-built batched-on-device landscapes should implement
//! [`BatchFitnessFn`] directly to avoid that round-trip.

use burn::tensor::{Tensor, TensorData, backend::Backend};

use rlevo_core::fitness::{FitnessEvaluable, Landscape};
use rlevo_core::objective::ObjectiveSense;

/// Single-member fitness evaluation.
///
/// Implementors may hold mutable state (e.g. a counter for number of
/// evaluations) and are therefore `&mut self`.
pub trait FitnessFn<G>: Send {
    /// Evaluates one genome and returns its scalar fitness.
    fn evaluate_one(&mut self, member: &G) -> f32;
}

/// Batched fitness evaluation over a population genome container `G`.
///
/// The returned tensor has shape `(pop_size,)` on the supplied device.
/// Implementors must preserve row order — `fitness[i]` refers to the
/// individual at row `i` of `population`.
pub trait BatchFitnessFn<B: Backend, G>: Send {
    /// Evaluates every member of `population` and returns a fitness tensor in
    /// the objective's **natural** value space (no hand-negation).
    ///
    /// The returned `Tensor<B, 1>` has shape `(pop_size,)` and is placed on
    /// `device`. Row order is preserved: `fitness[i]` corresponds to the
    /// individual at row `i` of `population`. Cost objectives return their
    /// natural cost; the harness reconciles direction via [`sense`](Self::sense).
    ///
    /// The returned tensor **may contain `NaN` or `±∞`** — implementors are not
    /// required to sanitize. The
    /// [`EvolutionaryHarness`](crate::strategy::EvolutionaryHarness) canonicalizes
    /// and then sanitizes (ADR 0034) before any [`Strategy::tell`](crate::strategy::Strategy::tell),
    /// so a non-finite fitness cannot poison selection or best-so-far tracking on
    /// harness-driven runs.
    fn evaluate_batch(
        &mut self,
        population: &G,
        device: &<B as burn::tensor::backend::BackendTypes>::Device,
    ) -> Tensor<B, 1>;

    /// The optimisation direction of this objective.
    ///
    /// This is the **single source of truth** the
    /// [`EvolutionaryHarness`](crate::strategy::EvolutionaryHarness) reads to
    /// reconcile a cost objective into the engine's canonical (maximise) space.
    /// It is **required, with no default**, so a reward/accuracy objective
    /// cannot silently inherit the wrong direction by omission — declare it
    /// explicitly ([`ObjectiveSense::Maximize`] for a reward,
    /// [`ObjectiveSense::Minimize`] for a cost). The bundled landscape adapters
    /// ([`FromLandscape`], [`FromFitnessEvaluable`]) forward the landscape's
    /// declared sense.
    fn sense(&self) -> ObjectiveSense;
}

/// Adapter from `FitnessEvaluable` to [`BatchFitnessFn<B, Tensor<B, 2>>`].
///
/// Each row of the population is pulled to host, converted to `Vec<f64>`,
/// and passed to the underlying evaluator with the configured landscape.
/// Fitness is computed on the host and then re-uploaded as a single
/// `Tensor<B, 1>`.
///
/// # Precision
///
/// Populations are read as `f32` and widened to `f64` for the evaluator
/// call; the returned `f64` fitness is narrowed back to `f32` before it
/// is uploaded as a `Tensor<B, 1>`. Fitness values that exceed `f32`
/// range (or rely on sub-ulp precision) will lose information at the
/// narrowing step. Purpose-built batched-on-device landscapes should
/// implement [`BatchFitnessFn`] directly to avoid the round-trip.
///
/// # Type Parameters
///
/// - `FE`: Concrete [`FitnessEvaluable`] implementation.
/// - `L`: Landscape type; must match `FE::Landscape`.
///
/// # Panics
///
/// `evaluate_batch` panics if the supplied population tensor is not rank
/// 2, or if its data cannot be read as `f32` (e.g. an integer backend).
#[derive(Debug)]
pub struct FromFitnessEvaluable<FE, L> {
    evaluator: FE,
    landscape: L,
    sense: ObjectiveSense,
}

impl<FE, L> FromFitnessEvaluable<FE, L> {
    /// Builds the adapter from an evaluator and a landscape, defaulting the
    /// objective sense to [`ObjectiveSense::Minimize`] (the cost convention a
    /// [`FitnessEvaluable`] follows).
    ///
    /// Use [`with_sense`](Self::with_sense) to declare a maximisation objective
    /// (reward, accuracy) explicitly.
    pub fn new(evaluator: FE, landscape: L) -> Self {
        Self::with_sense(evaluator, landscape, ObjectiveSense::Minimize)
    }

    /// Builds the adapter with an explicit [`ObjectiveSense`].
    pub fn with_sense(evaluator: FE, landscape: L, sense: ObjectiveSense) -> Self {
        Self {
            evaluator,
            landscape,
            sense,
        }
    }

    /// Returns a reference to the wrapped landscape.
    pub fn landscape(&self) -> &L {
        &self.landscape
    }
}

impl<FE, L, B> BatchFitnessFn<B, Tensor<B, 2>> for FromFitnessEvaluable<FE, L>
where
    B: Backend,
    FE: FitnessEvaluable<Individual = Vec<f64>, Landscape = L> + Send,
    L: Send + Sync,
{
    fn evaluate_batch(
        &mut self,
        population: &Tensor<B, 2>,
        device: &<B as burn::tensor::backend::BackendTypes>::Device,
    ) -> Tensor<B, 1> {
        let dims = population.dims();
        assert_eq!(dims.len(), 2, "population tensor must be rank 2");
        let pop_size = dims[0];
        let genome_dim = dims[1];

        let flat = population
            .clone()
            .into_data()
            .into_vec::<f32>()
            .expect("tensor data must be readable as f32");
        debug_assert_eq!(flat.len(), pop_size * genome_dim);

        let mut fitness = Vec::with_capacity(pop_size);
        let mut individual = Vec::with_capacity(genome_dim);
        for row in 0..pop_size {
            individual.clear();
            let start = row * genome_dim;
            individual.extend(
                flat[start..start + genome_dim]
                    .iter()
                    .map(|&v| f64::from(v)),
            );
            let f = self.evaluator.evaluate(&individual, &self.landscape);
            #[allow(clippy::cast_possible_truncation)]
            fitness.push(f as f32);
        }

        let data = TensorData::new(fitness, [pop_size]);
        Tensor::<B, 1>::from_data(data, device)
    }

    fn sense(&self) -> ObjectiveSense {
        self.sense
    }
}

/// Adapter from [`Landscape`] to [`BatchFitnessFn<B, Tensor<B, 2>>`].
///
/// Use this when the landscape carries its own `evaluate(&[f64]) -> f64`
/// (Sphere, Ackley, Rastrigin) so the example does not need a separate
/// `FitnessEvaluable` shim. Each row is pulled to host as `f32`, widened
/// to `f64`, evaluated, and re-uploaded as a `Tensor<B, 1>` — same
/// precision caveats as [`FromFitnessEvaluable`] apply.
///
/// # Panics
///
/// `evaluate_batch` panics if the supplied population tensor is not rank
/// 2, or if its data cannot be read as `f32` (e.g. an integer backend).
#[derive(Debug)]
pub struct FromLandscape<L> {
    landscape: L,
    sense: ObjectiveSense,
}

impl<L: Landscape> FromLandscape<L> {
    /// Builds the adapter from a self-evaluating landscape, taking the
    /// objective sense from the landscape's [`Landscape::sense`] (which
    /// defaults to [`ObjectiveSense::Minimize`]).
    pub fn new(landscape: L) -> Self {
        let sense = landscape.sense();
        Self { landscape, sense }
    }

    /// Builds the adapter with an explicit [`ObjectiveSense`], overriding the
    /// landscape's declared sense. Examples and showcases spell out
    /// [`ObjectiveSense::Minimize`] here so intent is visible at the call site.
    pub fn with_sense(landscape: L, sense: ObjectiveSense) -> Self {
        Self { landscape, sense }
    }

    /// Returns a reference to the wrapped landscape.
    pub fn landscape(&self) -> &L {
        &self.landscape
    }
}

impl<L, B> BatchFitnessFn<B, Tensor<B, 2>> for FromLandscape<L>
where
    B: Backend,
    L: Landscape,
{
    fn evaluate_batch(
        &mut self,
        population: &Tensor<B, 2>,
        device: &<B as burn::tensor::backend::BackendTypes>::Device,
    ) -> Tensor<B, 1> {
        let dims = population.dims();
        assert_eq!(dims.len(), 2, "population tensor must be rank 2");
        let pop_size = dims[0];
        let genome_dim = dims[1];

        let flat = population
            .clone()
            .into_data()
            .into_vec::<f32>()
            .expect("tensor data must be readable as f32");
        debug_assert_eq!(flat.len(), pop_size * genome_dim);

        let mut fitness = Vec::with_capacity(pop_size);
        let mut individual = Vec::with_capacity(genome_dim);
        for row in 0..pop_size {
            individual.clear();
            let start = row * genome_dim;
            individual.extend(
                flat[start..start + genome_dim]
                    .iter()
                    .map(|&v| f64::from(v)),
            );
            let f = self.landscape.evaluate(&individual);
            #[allow(clippy::cast_possible_truncation)]
            fitness.push(f as f32);
        }

        let data = TensorData::new(fitness, [pop_size]);
        Tensor::<B, 1>::from_data(data, device)
    }

    fn sense(&self) -> ObjectiveSense {
        self.sense
    }
}

/// Sanitizes one **canonical (maximise-space)** fitness value: `NaN →`
/// [`f32::NEG_INFINITY`], `+∞ →` [`f32::MAX`], everything else (including `−∞`)
/// passes through.
///
/// This is the crate-wide fitness-hygiene primitive and the single rule of the
/// canonical convention (ADR 0023 / ADR 0034):
///
/// - `NaN → −∞`: `−∞` is the worst value under the maximise convention, so a
///   sanitized `NaN` can never seed or displace a finite best-so-far. Rust's
///   `f32::NAN` is a *positive* NaN, so `total_cmp` would otherwise rank it as
///   the **maximum** (`rules.md` §3) — the exact inversion this prevents.
/// - `+∞ → f32::MAX`: a genuinely optimal individual (a landscape hitting its
///   optimum, an unbounded reward) still ranks top, but as a **finite** value —
///   so it cannot blow a population `mean`/`variance`/reward to `+∞`.
/// - `−∞` passes through: it is the worst-value sentinel *and* the
///   uninitialized `best_fitness_ever` seed, and it must stay non-finite so the
///   mean-over-finite statistic in
///   [`StrategyMetrics::from_host_fitness`](crate::strategy::StrategyMetrics::from_host_fitness)
///   can see and count it as a broken member.
///
/// Applied by [`BudgetedEval::eval`](crate::local_search::BudgetedEval) (every
/// probe, including the seeding eval), the searchers'
/// [`refine_with_known_fitness`](crate::local_search::LocalSearch::refine_with_known_fitness)
/// overrides, the EDA `tell` chokepoint
/// ([`crate::algorithms::eda::EdaStrategy`]), and every NaN-safe fitness sort
/// across the crate (selection, replacement, the ES/NEAT/ACO rankers). For the
/// finite benchmark landscapes the searchers ship against no branch is taken, so
/// it costs only one `is_nan` / `is_infinite` check.
pub(crate) fn sanitize_fitness(f: f32) -> f32 {
    if f.is_nan() {
        f32::NEG_INFINITY
    } else if f.is_infinite() && f.is_sign_positive() {
        // `f == f32::INFINITY` would trip the float-equality lint (rules §5/§8).
        f32::MAX
    } else {
        f
    }
}

/// Tensor-level [`sanitize_fitness`] for the driver chokepoints — a single
/// device op over a `(pop_size,)` **canonical-space** fitness vector.
///
/// Applies the same rule (`NaN → −∞`, `+∞ → f32::MAX`, `−∞` pass-through) to a
/// whole fitness tensor without a device→host→device round-trip, so the
/// [`EvolutionaryHarness`](crate::strategy::EvolutionaryHarness) and the
/// coevolution coupled-fitness path can sanitize on the hot path (ADR 0034).
///
/// Order matters: the `NaN → −∞` `mask_fill` runs first (so no `NaN` reaches the
/// clamp, which would propagate it), then `clamp_max(f32::MAX)` caps `+∞` while
/// leaving `−∞` and every finite value untouched. Mirrors the `is_nan` +
/// `mask_fill` + clamp idiom already used by `EdaStrategy::tell`'s gene backstop.
#[must_use]
pub(crate) fn sanitize_fitness_tensor<B: Backend>(fitness: Tensor<B, 1>) -> Tensor<B, 1> {
    let nan_mask = fitness.clone().is_nan();
    fitness
        .mask_fill(nan_mask, f32::NEG_INFINITY)
        .clamp_max(f32::MAX)
}

#[cfg(test)]
mod tests {
    use super::*;
    use burn::backend::Flex;
    type TestBackend = Flex;

    #[derive(Debug, Clone, Copy)]
    struct Sphere;

    struct SphereFit;
    impl FitnessEvaluable for SphereFit {
        type Individual = Vec<f64>;
        type Landscape = Sphere;
        fn evaluate(&self, x: &Self::Individual, _: &Self::Landscape) -> f64 {
            x.iter().map(|v| v * v).sum()
        }
    }

    #[test]
    fn from_fitness_evaluable_preserves_row_order() {
        let device = Default::default();
        let data = TensorData::new(
            vec![1.0_f32, 0.0, 0.0, 0.0, 2.0, 0.0, 0.0, 0.0, 3.0],
            [3, 3],
        );
        let pop = Tensor::<TestBackend, 2>::from_data(data, &device);

        let mut adapter = FromFitnessEvaluable::new(SphereFit, Sphere);
        let fitness = adapter.evaluate_batch(&pop, &device);

        let values = fitness
            .into_data()
            .into_vec::<f32>()
            .expect("fitness host-read of a tensor this test just built");
        assert_eq!(values.len(), 3);
        approx::assert_relative_eq!(values[0], 1.0, epsilon = 1e-6);
        approx::assert_relative_eq!(values[1], 4.0, epsilon = 1e-6);
        approx::assert_relative_eq!(values[2], 9.0, epsilon = 1e-6);
    }

    #[test]
    fn from_landscape_preserves_row_order() {
        let device = Default::default();
        let data = TensorData::new(
            vec![1.0_f32, 0.0, 0.0, 0.0, 2.0, 0.0, 0.0, 0.0, 3.0],
            [3, 3],
        );
        let pop = Tensor::<TestBackend, 2>::from_data(data, &device);

        let mut adapter = FromLandscape::new(SphereLandscape);
        let fitness = adapter.evaluate_batch(&pop, &device);

        let values = fitness
            .into_data()
            .into_vec::<f32>()
            .expect("fitness host-read of a tensor this test just built");
        assert_eq!(values.len(), 3);
        approx::assert_relative_eq!(values[0], 1.0, epsilon = 1e-6);
        approx::assert_relative_eq!(values[1], 4.0, epsilon = 1e-6);
        approx::assert_relative_eq!(values[2], 9.0, epsilon = 1e-6);
    }

    struct SphereLandscape;
    impl Landscape for SphereLandscape {
        fn evaluate(&self, x: &[f64]) -> f64 {
            x.iter().map(|v| v * v).sum()
        }
    }

    /// The scalar hygiene rule (ADR 0034): `NaN → −∞`, `+∞ → f32::MAX`, `−∞` and
    /// finite values pass through unchanged.
    #[test]
    fn sanitize_fitness_scalar_applies_canonical_rule() {
        // `−∞` sentinel: assert via `is_infinite`/sign, not float `==` (rules §5/§8).
        let nan_out: f32 = sanitize_fitness(f32::NAN);
        assert!(
            nan_out.is_infinite() && nan_out.is_sign_negative(),
            "NaN → −∞"
        );
        approx::assert_relative_eq!(sanitize_fitness(f32::INFINITY), f32::MAX);
        let neg_out: f32 = sanitize_fitness(f32::NEG_INFINITY);
        assert!(
            neg_out.is_infinite() && neg_out.is_sign_negative(),
            "−∞ passes through"
        );
        approx::assert_relative_eq!(sanitize_fitness(2.5), 2.5, epsilon = 1e-6);
        approx::assert_relative_eq!(sanitize_fitness(-7.0), -7.0, epsilon = 1e-6);
    }

    /// The tensor sibling applies the identical rule element-wise, and — crucially
    /// — leaves `−∞` non-finite (it is not clamped to `−f32::MAX`) so downstream
    /// mean-over-finite logic can still detect and count it.
    #[test]
    fn sanitize_fitness_tensor_matches_scalar_rule() {
        let device = Default::default();
        let data = TensorData::new(
            vec![f32::NAN, f32::INFINITY, f32::NEG_INFINITY, 3.0_f32, -4.0],
            [5],
        );
        let t = Tensor::<TestBackend, 1>::from_data(data, &device);
        let out = sanitize_fitness_tensor(t)
            .into_data()
            .into_vec::<f32>()
            .expect("fitness host-read of a tensor this test just built");

        assert!(
            out[0].is_infinite() && out[0].is_sign_negative(),
            "NaN → −∞"
        );
        approx::assert_relative_eq!(out[1], f32::MAX); // +∞ → f32::MAX
        assert!(
            out[2].is_infinite() && out[2].is_sign_negative(),
            "−∞ passes through, stays non-finite"
        );
        approx::assert_relative_eq!(out[3], 3.0, epsilon = 1e-6);
        approx::assert_relative_eq!(out[4], -4.0, epsilon = 1e-6);
    }

    /// Regression test for the load-bearing `BatchFitnessFn` invariant
    /// documented in the fitness chapter of the user-book: `evaluate_batch`
    /// returns a `Tensor<B, 1>` of shape `(pop_size,)` with row order
    /// preserved.
    ///
    /// The population is deliberately **non-square** (`pop_size != genome_dim`)
    /// so a row/column transposition — reading the genome axis as the
    /// population axis — cannot hide behind a square shape, and the output
    /// length is asserted against `pop_size` (the rows), not `genome_dim`.
    /// Row `i` is `[i + 1, 0]`, so Sphere fitness is `(i + 1)^2`: a permuted
    /// mapping yields a different, detectable vector.
    #[test]
    fn from_fitness_evaluable_output_is_pop_size_shaped_and_row_aligned() {
        let device = Default::default();
        // 4 individuals, 2 genes each.
        let data = TensorData::new(vec![1.0_f32, 0.0, 2.0, 0.0, 3.0, 0.0, 4.0, 0.0], [4, 2]);
        let pop = Tensor::<TestBackend, 2>::from_data(data, &device);

        let mut adapter = FromFitnessEvaluable::new(SphereFit, Sphere);
        let fitness = adapter.evaluate_batch(&pop, &device);

        // Shape `(pop_size,)`: rank 1, exactly `pop_size` (4) elements —
        // not `genome_dim` (2).
        assert_eq!(fitness.dims(), [4]);

        let values = fitness
            .into_data()
            .into_vec::<f32>()
            .expect("fitness host-read of a tensor this test just built");
        for (i, &v) in values.iter().enumerate() {
            #[allow(clippy::cast_precision_loss)]
            let expected = ((i + 1) * (i + 1)) as f32;
            approx::assert_relative_eq!(v, expected, epsilon = 1e-6);
        }
    }

    /// Same invariant for [`FromLandscape`] — the two adapters carry
    /// independent copies of the row-walking loop, so each pins the shape and
    /// row alignment separately.
    #[test]
    fn from_landscape_output_is_pop_size_shaped_and_row_aligned() {
        let device = Default::default();
        // 4 individuals, 2 genes each — deliberately non-square.
        let data = TensorData::new(vec![1.0_f32, 0.0, 2.0, 0.0, 3.0, 0.0, 4.0, 0.0], [4, 2]);
        let pop = Tensor::<TestBackend, 2>::from_data(data, &device);

        let mut adapter = FromLandscape::new(SphereLandscape);
        let fitness = adapter.evaluate_batch(&pop, &device);

        assert_eq!(fitness.dims(), [4]);

        let values = fitness
            .into_data()
            .into_vec::<f32>()
            .expect("fitness host-read of a tensor this test just built");
        for (i, &v) in values.iter().enumerate() {
            #[allow(clippy::cast_precision_loss)]
            let expected = ((i + 1) * (i + 1)) as f32;
            approx::assert_relative_eq!(v, expected, epsilon = 1e-6);
        }
    }
}