Struct PerturbedOptimizer 

pub struct PerturbedOptimizer<F>
where
    F: Fn(&[f64]) -> Vec<f64>,
{ /* private fields */ }

A differentiable wrapper around any black-box combinatorial optimizer.

Wraps a function optimizer: θ → y*(θ) and computes a smooth approximation E[y*(θ + σZ)] via Monte Carlo.

Type Parameter

• F – function type for the combinatorial optimizer, mapping &[f64] to Vec<f64>.

Implementations

impl<F> PerturbedOptimizer<F>

where F: Fn(&[f64]) -> Vec<f64>,

pub fn new(optimizer: F) -> Self

Create a new perturbed optimizer with default configuration.

pub fn with_config(optimizer: F, config: PerturbedOptimizerConfig) -> Self

Create a new perturbed optimizer with custom configuration.

pub fn forward(&mut self, theta: &[f64]) -> OptimizeResult<Vec<f64>>

Forward pass: compute E[y*(θ + σZ)] via Monte Carlo.

Samples n_samples perturbations Z_k ~ N(0, I) and returns the sample mean of the optimizer outputs.

Arguments

• theta – parameter vector (length d).

Returns

Expected optimizer output ŷ(θ) (length equal to optimizer output).

pub fn gradient(&self, theta: &[f64], dl_dy: &[f64]) -> OptimizeResult<Vec<f64>>

Gradient estimate via reparameterized covariance:

grad_theta L ~ (1/sigma) Cov[y*(theta + sigma*Z), Z] * dL/dy = (1/sigma^2*N) Sum_k (y_k - y_mean) * Z_k * dL/dy

This is an unbiased estimator when y* is the gradient of a linear function, and has lower variance than REINFORCE.

Arguments

• theta – parameter vector (length d).
• dl_dy – upstream gradient dL/dŷ (length = optimizer output length).

Returns

Gradient estimate ∇_θ L (length d).

pub fn reinforce_gradient( &self, theta: &[f64], dl_dy: &[f64], ) -> OptimizeResult<Vec<f64>>

REINFORCE (score-function) gradient estimator:

∇_θ L ≈ (1/σN) Σ_k L(y_k) Z_k

where L(y_k) = dL/dy · y_k (linear approximation to the loss).

Arguments

• theta – parameter vector.
• dl_dy – upstream gradient (defines the loss as L = dl_dy · y).

pub fn last_mean_output(&self) -> Option<Vec<f64>>

Access the cached mean output from the last forward pass.