Struct AdaptiveLeafModel 

pub struct AdaptiveLeafModel { /* private fields */ }

Leaf model that starts as closed-form and promotes to a more complex model when the Hoeffding bound confirms it is statistically superior.

Runs a shadow model alongside the active (closed-form) model. On each update, both models are trained and their per-sample losses compared using the second-order Taylor approximation:

loss_i = gradient * prediction + 0.5 * hessian * prediction^2
advantage_i = loss_active_i - loss_shadow_i

When mean(advantage) > epsilon where epsilon is the Hoeffding bound (using the tree’s delta parameter), the shadow model is promoted to active and the overhead drops to zero.

This uses the same statistical guarantee as the tree’s split decisions – no arbitrary thresholds.

Implementations

impl AdaptiveLeafModel

pub fn new( shadow: Box<dyn LeafModel>, promote_to: LeafModelType, delta: f64, seed: u64, ) -> Self

Create a new adaptive leaf model.

The active model starts as ClosedFormLeaf. The shadow model is the candidate that will be promoted if it proves statistically superior.

Trait Implementations

impl LeafModel for AdaptiveLeafModel

fn predict(&self, features: &[f64]) -> f64

Produce a prediction given input features.

fn update(&mut self, features: &[f64], gradient: f64, hessian: f64, lambda: f64)

Update model parameters given a gradient, hessian, and regularization lambda.

fn clone_fresh(&self) -> Box<dyn LeafModel>

Create a fresh (zeroed / re-initialized) clone of this model’s architecture.

fn clone_warm(&self) -> Box<dyn LeafModel>

Create a warm clone preserving learned weights but resetting optimizer state.

impl Send for AdaptiveLeafModel

impl Sync for AdaptiveLeafModel