Struct OnlineGradientDescent 

pub struct OnlineGradientDescent {
    pub weights: Vec<f64>,
    pub eta: f64,
    pub d: f64,
    pub g: f64,
    pub t: usize,
}

Online Gradient Descent with regret bound O(√T).

Fields

weights: Vec<f64>

Current parameter vector w_t.

eta: f64

Learning rate η.

d: f64

Constraint set radius D (‖w‖ ≤ D).

g: f64

Gradient norm bound G (‖∇_t‖ ≤ G).

t: usize

Round count.

Implementations

impl OnlineGradientDescent

pub fn new(dim: usize, eta: f64, d: f64, g: f64) -> Self

Create a new OGD instance.

pub fn update(&mut self, gradient: &[f64])

Update: w_{t+1} = project(w_t - η ∇_t) onto ‖w‖ ≤ D.

pub fn regret_bound(&self) -> f64

Regret bound after T rounds: R_T ≤ D * G * √(2T).