OakEmu-RL v0.1.0

Appearance

ReverseCurriculumMidpointStartConfiguration ​

Competence-driven reverse curriculum: keep starting episodes from midpoints near the exit and recede the start toward the true start as the policy learns to win from progressively earlier cells of the best route. Replaces a hand-tuned probability schedule with a measured-competence signal.

Source

Methods ​

validate_bounds ​

python
validate_bounds() -> ReverseCurriculumMidpointStartConfiguration

Source

Values ​

bootstrap_probability ​

python
bootstrap_probability: float = Field(default=1.0, ge=0.0, le=1.0)

Probability used before any archived route / competence exists (the random-prefill phase).

Source

competence_utility ​

python
competence_utility: str = 'win_competence'

The win-from-cell competence utility whose per-cell success rate drives the recede.

Source

graduation_success_rate ​

python
graduation_success_rate: float = Field(default=0.5, ge=0.0, le=1.0)

A route cell counts as solved once its success rate reaches this; the solved region growing back toward the start is what lowers the midpoint probability.

Source

kind ​

python
kind: Literal['reverse_curriculum'] = 'reverse_curriculum'

Source

max_midpoint_probability ​

python
max_midpoint_probability: float = Field(default=0.85, ge=0.0, le=1.0)

Cap below 1 so the true-start task is never fully starved while the route is unsolved.

Source

min_midpoint_probability ​

python
min_midpoint_probability: float = Field(default=0.0, ge=0.0, le=1.0)

Floor applied until the route is fully solved from the start.

Source