ribs.emitters.rankers.TwoStageRandomDirectionRanker¶

class ribs.emitters.rankers.TwoStageRandomDirectionRanker(seed: int | integer | None = None)[source]¶

Ranks solutions based on status and on projection onto a direction in measure space.

This ranker differs from ribs.emitters.rankers.RandomDirectionRanker in that solutions are ranked in two stages: first by whether they are added, then by their projection onto a random direction in the archive space.

This ranker originates from the random direction emitter in Fontaine 2020.

Note that archives used with this ranker must have lower_bounds and upper_bounds attributes.

Methods

`rank`(emitter, archive, data, add_info)	Ranks solutions first by whether they are added, then by their projection onto a random direction in the archive.
`reset`(emitter, archive)	Generates a random direction in the archive's measure space.

Attributes

`target_measure_dir`	`(measure_dim,)` array with the target measure direction vector.

rank(emitter: EmitterBase, archive: ArchiveBase, data: dict[str, ndarray], add_info: dict[str, ndarray]) → tuple[ndarray, ndarray][source]¶

Ranks solutions first by whether they are added, then by their projection onto a random direction in the archive.

Parameters:¶

emitter: EmitterBase¶: Emitter that this ranker object belongs to.
archive: ArchiveBase¶: Archive used by emitter when creating and inserting solutions.
data: dict[str, ndarray]¶: Dict mapping from field names like solution and objective to arrays with data for the solutions.
add_info: dict[str, ndarray]¶: Information returned by an archive’s add() method.

Returns:¶

The first array (shape (batch_size,)) is an array of indices representing a ranking of the solutions and the second array (shape (batch_size,) or (batch_size, n_values)) is an array of values that this ranker used to rank the solutions. batch_size is the number of solutions and n_values is the number of values that the rank function used.

reset(emitter: EmitterBase, archive: ArchiveBase) → None[source]¶

Generates a random direction in the archive’s measure space.

A random direction is sampled from a standard Gaussian – since the standard Gaussian is isotropic, there is equal probability for any direction. The direction is then scaled to the archive bounds so that it is a random archive direction.

Parameters:¶

emitter: EmitterBase¶: Emitter that this ranker object belongs to.
archive: ArchiveBase¶: Archive used by emitter when creating and inserting solutions.

property target_measure_dir : ndarray¶: (measure_dim,) array with the target measure direction vector.