OakEmu-RL v0.1.0

Appearance

ResetParameterCurriculumConfiguration ​

Optional curriculum that varies reset parameters over training, one staggered stage at a time. Generic over whatever start parameters a game exposes — Manic Miner varies start_cavern; another game might vary a level / difficulty / world (possibly several at once). None on the experiment => a single stage with no reset parameters sent, that is unchanged single-configuration behaviour. A single stage with activate_at_step=0 is the degenerate case routed through the same path.

Each stage is fed its effective step (global step - activation step) to the existing step-based exploration schedule, so a freshly activated stage gets its own exploration window with no per-stage schedule duplication. Pair with store.archive_group_field (the field that identifies the stage in a recorded route, for example "cavern") so the best-route archive is per stage, and an OakEmu build with unsafe reset parameters.

Source

Values ​

focus_steps ​

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

While a stage is within focus_steps of its activation it is over-sampled by focus_weight so it gets a focused exploration burst before settling into the mix. 0 => uniform over active stages.

Source

focus_weight ​

python
focus_weight: float = Field(default=1.0, ge=1.0)

Relative sampling weight for a stage still inside its focus_steps window (1.0 = uniform).

Source

stages ​

python
stages: list[CurriculumStageConfiguration] = Field(min_length=1)

The curriculum stages and their activation steps.

Source