Trait BackendOptimizer 

pub trait BackendOptimizer {
    type Module;

    // Required methods
    fn clip_grad_norm(&mut self, max: f64);
    fn step_module(&mut self, module: Self::Module) -> Self::Module;
    fn learning_rate(&self) -> f64;
}

Burn-side optimizer interface used by PPO and DQN trainers.

Burn’s Optimizer<M, B> is move-through: every gradient step consumes the module by value and returns the updated copy. This trait exposes that single fundamental verb plus the gradient-clipping configuration knob the trainer needs.

Required Associated Types

type Module

The module type the optimizer steps.

Required Methods

fn clip_grad_norm(&mut self, max: f64)

Stage the maximum global gradient L2-norm.

This only records the cap on the wrapper; the trainer bodies read it back via BurnOptimizer::grad_clip_norm and apply the clip to the gradient slice before their move-through inner_mut().step(...) call (see the joint trainer’s per-policy step, issue #239). The trait-level Self::step_module fallback does not itself clip.

fn step_module(&mut self, module: Self::Module) -> Self::Module

Burn-style move-through update.

Consumes module, applies the optimizer’s staged gradient (with any clipping configured by Self::clip_grad_norm), and returns the updated module.

fn learning_rate(&self) -> f64

Construction-time learning rate. Exposed for diagnostics.

Dyn Compatibility

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors

impl<B, M, O> BackendOptimizer for BurnOptimizer<B, M, O>

where B: AutodiffBackend, M: AutodiffModule<B>, O: Optimizer<M, B>,

type Module = M