OakEmu-RL v0.1.0

Appearance

TrajectoryMidpointConfiguration ​

Midpoint-start phase configuration: selection and the start-probability policy.

Source

Values ​

backward_curriculum ​

python
backward_curriculum: BackwardCurriculumConfiguration = Field(default_factory=BackwardCurriculumConfiguration)

Optional strict receding-front curriculum over the archived route (replaces the utility blend when on).

Source

frame_stack_replays_per_update ​

python
frame_stack_replays_per_update: int = Field(default=0, ge=0)

Stored paths replayed per rollout to mint restart frame stacks for bucket-sampled midpoints.

A snapshot is an emulator state, not a rendered-frame history, so a restart without a minted stack sees four copies of one frame -- a world in which nothing is moving. When this is above zero the actor replays up to this many un-minted paths per rollout (newest first) AND midpoint sampling only offers segments whose stack exists, so a restart is never blind; watch Trajectories/Frame_Stack_Coverage for lag. 0 keeps the old behaviour: no minting, no gate, bucket restarts restore blind. Curriculum restarts are unaffected either way -- the archived route has its own minting and coverage verification.

Source

partition_reset_parameter ​

python
partition_reset_parameter: str | None = None

Name of a reset parameter (set by a curriculum stage, for example start_cavern) whose value selects which partition a midpoint restart may come from. When set, midpoint restarts are filtered to recordings whose store.archive_group_field cell value equals this reset parameter's current value, so a focused curriculum stage actually restarts within its own partition instead of the dominant one. A midpoint restart loads a full-state snapshot that fixes the level, so without this filter the snapshot's partition wins and a freshly-activated level (for example a new cavern) is starved of midpoint exploration. None disables partition filtering (single-partition training).

Source

selection ​

python
selection: TrajectorySelectionConfiguration = Field(default_factory=TrajectoryDirectSelectionConfiguration)

Source

start ​

python
start: MidpointStartPolicyConfiguration = Field(default_factory=ScheduledMidpointStartConfiguration)

How often a policy episode starts from a stored midpoint instead of the true start.

Source