ribs.emitters.opt.OpenAIEvolutionStrategy¶

class ribs.emitters.opt.OpenAIEvolutionStrategy(sigma0: float | floating, solution_dim: int | integer, batch_size: int | integer | None = None, seed: int | integer | None = None, dtype: numpy.typing.DTypeLike = numpy.float64, lower_bounds: float | floating | ndarray = -inf, upper_bounds: float | floating | ndarray = inf, mirror_sampling: bool = True, **adam_kwargs: __SPHINX_IMMATERIAL_TYPE_VAR__P_P)[source]¶

OpenAI-ES optimizer for use with emitters.

Refer to EvolutionStrategyBase for usage instruction.

Parameters:¶

sigma0: float | floating¶

Initial step size.

batch_size: int | integer | None = None¶

Number of solutions to evaluate at a time. If None, we calculate a default batch size based on solution_dim.

solution_dim: int | integer¶

Size of the solution space.

seed: int | integer | None = None¶

Seed for the random number generator.

dtype: numpy.typing.DTypeLike = numpy.float64¶

Data type of solutions.

lower_bounds: float | floating | ndarray = -inf¶

scalar or (solution_dim,) array indicating lower bounds of the solution space. Scalars specify the same bound for the entire space, while arrays specify a bound for each dimension. Pass -np.inf in the array or scalar to indicated unbounded space.

upper_bounds: float | floating | ndarray = inf¶

Same as above, but for upper bounds (and pass np.inf instead of -np.inf).

mirror_sampling: bool = True¶

Whether to use mirror sampling when gathering solutions. Defaults to True.

**adam_kwargs: __SPHINX_IMMATERIAL_TYPE_VAR__P_P¶

Keyword arguments passed to AdamOpt.

Methods

ask([batch_size])	Samples new solutions.
check_stop(ranking_values)	Checks if the ES should stop and be reset.
reset(x0)	Resets the ES to start at x0.
tell(ranking_indices, ranking_values, ...)	Passes the solutions back to the ES.

ask(batch_size: int | integer | None = None) → ndarray[source]¶

Samples new solutions.

Parameters:¶

batch_size: int | integer | None = None¶

batch size of the sample. Defaults to self.batch_size.

Returns:¶

Array of new solutions.

check_stop(ranking_values: ndarray) → bool[source]¶

Checks if the ES should stop and be reset.

Parameters:¶

ranking_values: ndarray¶

Array of values that were used to rank the solutions. Shape can be either (batch_size,) or (batch_size, n_values)``, where batch_size is the number of solutions and n_values is the number of values that the ranker used. Note that unlike in tell(), these values must be sorted according to the ranking_indices passed to tell().

Returns:¶

True if any of the stopping conditions are satisfied.

reset(x0: ndarray) → None[source]¶

Resets the ES to start at x0.

Parameters:¶

x0: ndarray¶

Initial mean.

tell(ranking_indices: ndarray, ranking_values: ndarray, num_parents: int | integer) → None[source]¶

Passes the solutions back to the ES.

Parameters:¶

ranking_indices: ndarray¶

Integer indices that are used to rank the solutions returned in ask(). Note that these are NOT the ranks of the solutions. Rather, they are indices such that solutions[ranking_indices] will correctly rank the solutions (think of an argsort).

ranking_values: ndarray¶

Array of values that were used to rank the solutions. Shape can be either (batch_size,) or (batch_size, n_values)``, where batch_size is the number of solutions and n_values is the number of values that the ranker used. Note that we assume a descending sort, i.e., higher values should come first.

num_parents: int | integer¶

Number of top solutions to select from the ranked solutions.