Module buffer 

Experience replay buffers for reinforcement learning.

This module provides several buffer implementations, each suited to different training scenarios:

Buffer	Use case
ringbuf::ReplayBuffer	Standard uniform off-policy replay (DQN, SAC).
priority::PrioritizedReplayBuffer	Proportional PER with importance-sampling weights.
sequence::SequenceReplayBuffer	Contiguous subsequence sampling for recurrent policies.
her::HERBuffer	Hindsight Experience Replay for goal-conditioned RL.
mmap::MmapReplayBuffer	Disk-backed spill for large replay capacities.
columnar::ExperienceTable	Append-only on-policy table (PPO, A2C).
offline::OfflineDatasetBuffer	Read-only offline RL datasets.
concurrent::ConcurrentReplayBuffer	Lock-free multi-producer buffer.

All buffers store observations and actions as f32, rewards as f32, and boolean done flags. The ExperienceRecord struct is a convenience for the push() API; prefer push_slices() to avoid intermediate Vec allocations in hot paths.

Modules

columnar

concurrent

episode

Episode boundary tracking for ring buffers.

extra_columns

her

Hindsight Experience Replay (HER) buffer.

mixed

Mixed sampling from two replay buffers.

mmap

offline

Read-only offline dataset buffer for offline RL algorithms.

priority

Sum-tree backed prioritized experience replay.

provenance

ringbuf

sequence

Sequence replay buffer for recurrent/transformer-based RL algorithms.

varlen

Structs

ExperienceRecord

A single experience record to push into a buffer. Uses f32 throughout for numpy compatibility.