Struct ReplayBuffer 

pub struct ReplayBuffer { /* private fields */ }

Fixed-capacity ring buffer with uniform random sampling.

Pre-allocates all arrays at construction for zero-allocation push. Oldest transitions are overwritten when capacity is reached.

Supports optional extra f32 columns (e.g. log-probs, value estimates) via ColumnHandle. When no extra columns are registered, there is zero overhead — no allocations and no branches in the hot push/sample path.

Implementations

impl ReplayBuffer

pub fn new(capacity: usize, obs_dim: usize, act_dim: usize) -> Self

Create a ring buffer with fixed capacity. All arrays are pre-allocated.

pub fn register_column(&mut self, name: &str, dim: usize) -> ColumnHandle

Register an extra f32 column with the given name and dimensionality.

Returns a ColumnHandle for O(1) push/sample access. Must be called before any push() — the column is pre-allocated to match the buffer’s capacity.

pub fn push_extra( &mut self, handle: ColumnHandle, values: &[f32], ) -> Result<(), RloxError>

Push extra column data for the most recently pushed transition.

Must be called after push() and before the next push(). The values slice length must match the column’s registered dim.

pub fn obs_dim(&self) -> usize

Observation dimensionality.

pub fn act_dim(&self) -> usize

Action dimensionality.

pub fn len(&self) -> usize

Number of valid transitions currently stored.

pub fn is_empty(&self) -> bool

Whether the buffer is empty.

pub fn push_slices( &mut self, obs: &[f32], next_obs: &[f32], action: &[f32], reward: f32, terminated: bool, truncated: bool, ) -> Result<(), RloxError>

Push a transition from borrowed slices, avoiding intermediate allocation.

pub fn push_batch( &mut self, obs_batch: &[f32], next_obs_batch: &[f32], actions_batch: &[f32], rewards: &[f32], terminated: &[bool], truncated: &[bool], ) -> Result<(), RloxError>

Push multiple transitions at once from flat arrays.

obs_batch shape: [n * obs_dim], next_obs_batch: same, actions_batch: [n * act_dim], others: [n].

pub fn push(&mut self, record: ExperienceRecord) -> Result<(), RloxError>

Push a transition, overwriting the oldest if at capacity.

Prefer push_slices to avoid the intermediate Vec<f32> allocations inside ExperienceRecord.

pub fn sample( &self, batch_size: usize, seed: u64, ) -> Result<SampledBatch, RloxError>

Sample a batch of transitions uniformly at random.

Uses ChaCha8Rng seeded with seed for deterministic cross-platform reproducibility. Returns owned SampledBatch.

If extra columns have been registered, their data is included in SampledBatch::extra.

pub fn sample_into( &self, batch: &mut SampledBatch, batch_size: usize, seed: u64, ) -> Result<(), RloxError>

Sample into a pre-allocated batch, reusing its capacity.

Same as sample() but avoids allocation by reusing batch.

pub fn extra_columns(&self) -> &ExtraColumns

Access the extra columns storage (for advanced use / testing).

Trait Implementations

impl Debug for ReplayBuffer

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

Formats the value using the given formatter.