Struct RewardNormalizer 

pub struct RewardNormalizer { /* private fields */ }

Reward normalizer with optional return-based normalisation and clipping.

Implementations

impl RewardNormalizer

pub fn new(n_envs: usize, gamma: f32, clip: f32, mode: RewardNormMode) -> Self

Create a reward normalizer.

• n_envs — number of parallel environments.
• gamma — discount factor for return accumulation.
• clip — symmetric clip range.
• mode — normalisation mode.

pub fn process(&mut self, rewards: &[f32], dones: &[f32]) -> Vec<f32>

Process rewards for a step across n_envs parallel environments.

Updates the running return estimates and normalizes.

Returns the normalised / clipped rewards.

Panics

Panics if rewards.len() != n_envs or dones.len() != n_envs.

pub fn normalise_eval(&self, rewards: &[f32]) -> Vec<f32>

Normalise a batch of rewards without updating running statistics (evaluation).

pub fn reset_returns(&mut self)

Reset running returns (call at episode start).

Trait Implementations

impl Clone for RewardNormalizer

fn clone(&self) -> RewardNormalizer

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for RewardNormalizer

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

Formats the value using the given formatter.