OakEmu-RL v0.1.0

Appearance

TrajectoryReplayConfiguration ​

Replay/BC configuration: selection over the shared store and the imitation gate.

Source

Methods ​

validate_bc_actions ​

python
validate_bc_actions() -> TrajectoryReplayConfiguration

Source

Values ​

bc_competence_gate ​

python
bc_competence_gate: str | None = None

Optional competence utility used to down-weight BC samples from mastered start cells, so imitation pressure fades from sections the policy already executes reliably.

Source

behaviour ​

python
behaviour: TrajectoryReplayBehaviourConfiguration | None = None

Source

cache_replays ​

python
cache_replays: bool = True

Cache replayed observations per (initial state, actions) slice and skip the emulator for repeat replays; replays are deterministic so cached entries never go stale. Trades memory (roughly 34 KB per cached step) for speed; disable to re-run every replay on the emulator.

Source

max_bc_actions ​

python
max_bc_actions: int = Field(default=48, ge=1)

Maximum number of actions in a cloned sub-segment slice.

Source

min_bc_actions ​

python
min_bc_actions: int = Field(default=16, ge=1)

Minimum number of actions in a cloned sub-segment slice. Cloning whole routes caused win/loss flip-flops, while one-action fragments imitate isolated frames rather than a coherent manoeuvre, so BC clones contiguous sub-segments of at least this many actions.

Source

selection ​

python
selection: TrajectorySelectionConfiguration = Field(default_factory=TrajectoryDirectSelectionConfiguration)

Source