Struct WolfeLineSearch 

pub struct WolfeLineSearch { /* private fields */ }

Wolfe line search with Armijo and curvature conditions.

Enforces both the Armijo condition (sufficient decrease) and the curvature condition (sufficient curvature):

Armijo:    f(x + α*d) ≤ f(x) + c₁*α*∇f(x)ᵀd
Curvature: |∇f(x + α*d)ᵀd| ≤ c₂*|∇f(x)ᵀd|

The curvature condition ensures the step size is not too small by requiring that the gradient has decreased sufficiently along the search direction.

Parameters

• c1: Armijo constant (typical: 1e-4), c₁ ∈ (0, c₂)
• c2: Curvature constant (typical: 0.9), c₂ ∈ (c₁, 1)
• max_iter: Maximum line search iterations

Example

use aprender::optim::{WolfeLineSearch, LineSearch};
use aprender::primitives::Vector;

let line_search = WolfeLineSearch::new(1e-4, 0.9, 50);

let f = |x: &Vector<f32>| x[0] * x[0];
let grad = |x: &Vector<f32>| Vector::from_slice(&[2.0 * x[0]]);

let x = Vector::from_slice(&[1.0]);
let d = Vector::from_slice(&[-2.0]);

let alpha = line_search.search(&f, &grad, &x, &d);
assert!(alpha > 0.0);

Implementations

impl WolfeLineSearch

pub fn new(c1: f32, c2: f32, max_iter: usize) -> WolfeLineSearch

Creates a new Wolfe line search.

Arguments

• c1 - Armijo constant (typical: 1e-4)
• c2 - Curvature constant (typical: 0.9)
• max_iter - Maximum iterations (typical: 50)

Panics

Panics if c1 >= c2 or values are outside (0, 1).

Example

use aprender::optim::WolfeLineSearch;

let line_search = WolfeLineSearch::new(1e-4, 0.9, 50);

Trait Implementations

impl Clone for WolfeLineSearch

fn clone(&self) -> WolfeLineSearch

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for WolfeLineSearch

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

Formats the value using the given formatter.

impl Default for WolfeLineSearch

fn default() -> WolfeLineSearch

Creates a Wolfe line search with default parameters.

Defaults: c1=1e-4, c2=0.9, max_iter=50

impl LineSearch for WolfeLineSearch

fn search<F, G>(&self, f: &F, grad: &G, x: &Vector<f32>, d: &Vector<f32>) -> f32

where F: Fn(&Vector<f32>) -> f32, G: Fn(&Vector<f32>) -> Vector<f32>,

Finds a suitable step size along the search direction.