Struct PerturbedOptimizer 

pub struct PerturbedOptimizer { /* private fields */ }

Differentiable argmax via additive Gaussian perturbations.

Estimates E_Z[argmax(θ + ε·Z)] where Z ~ N(0, I). The forward pass returns soft assignment probabilities; the backward pass returns an unbiased gradient estimate via the score-function estimator.

Implementations

impl PerturbedOptimizer

pub fn new(config: PerturbedOptimizerConfig) -> Self

Create a new perturbed optimizer with the given configuration.

pub fn forward<F>(&self, scores: &[F]) -> OptimizeResult<Vec<F>>

where F: Float + FromPrimitive + Debug + Clone,

Forward pass: estimate E[argmax(θ + εZ)] via Monte Carlo.

Returns a probability vector of length scores.len().

pub fn backward<F>( &self, scores: &[F], upstream: &[F], ) -> OptimizeResult<Vec<F>>

where F: Float + FromPrimitive + Debug + Clone,

Backward pass: gradient estimate via score-function / log-derivative trick.

Formula

dL/dθ ≈ (1 / (ε² · n)) Σ_i [<argmax(θ+εZ_i), upstream>] · Z_i

Arguments

• scores – original (unperturbed) scores.
• upstream – upstream gradient ∂L/∂p (same shape as forward output).

Returns

Gradient ∂L/∂θ of the same length as scores.

Trait Implementations

impl Clone for PerturbedOptimizer

fn clone(&self) -> PerturbedOptimizer

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for PerturbedOptimizer

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

Formats the value using the given formatter.