OakEmu-RL v0.1.0

Appearance

TrajectoryCostAwareUtilityConfiguration ​

Cell score for cost-aware (optimality-gap) exploration: the cost of the best known winning route through the cell.

Coverage utilities stop at "a win is reachable here"; this one asks "…but how expensive is the cheapest win through here?" The cost of each winning episode (for example air consumed, or route length) is recorded against every cell on its route under a min aggregator, so each cell remembers the cheapest win that has ever passed through it. Cells whose cheapest known win is still costly score high — the policy keeps probing them for a shorter route — and the score falls automatically as a cheaper win is found, so exploration self-extinguishes once a cell is on an efficient route. This is the efficiency lever (the air/length signal) moved from retention into exploration; cells with no recorded win score 0 (coverage utilities cover those).

Source

Properties ​

is_cell_scoring ​

python
is_cell_scoring: bool

Source

uses_cell_key_grouping ​

python
uses_cell_key_grouping: bool

Source

wants_route_cost_table ​

python
wants_route_cost_table: bool

Source

Values ​

cost ​

python
cost: TrajectoryUtilityMetricConfiguration

Per-winning-episode cost metric (higher = costlier), recorded against every cell on the route.

Source

kind ​

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

Source