baxus.util.behaviors package¶

Submodules¶

baxus.util.behaviors.baxus_configuration module¶

class baxus.util.behaviors.baxus_configuration.BaxusBehavior(initial_base_length: float = 0.8, max_base_length: float = 1.6, min_base_length: float = 0.0078125, success_tolerance: int = 3, acquisition_function: AcquisitionFunctionType = AcquisitionFunctionType.THOMPSON_SAMPLING, noise: float = 0.0, embedding_type: EmbeddingType = EmbeddingType.BAXUS, success_decision_factor: float = 0.001, n_new_bins: int = 3, budget_until_input_dim: int = 0, adjust_initial_target_dim: bool = True)¶

Bases: EmbeddedTuRBOBehavior

The behavior of the AdaTheSBO1 algorithm.

adjust_initial_target_dim: bool = True¶: Whether to adjust the initial target dim such that the final split is as close to the ambient dim as possible.

budget_until_input_dim: int = 0¶: The budget after which we have reached the input dimension under the assumption that we always fail. If zero: use the entire evaluation budget.

property conf_dict: Dict[str, Any]¶

The configuration as a dictionary.

Returns: The configuration as a dictionary.

n_new_bins: int = 3¶

3

Type: Number of new bins after a splitting. Default

baxus.util.behaviors.embedded_turbo_configuration module¶

class baxus.util.behaviors.embedded_turbo_configuration.EmbeddedTuRBOBehavior(initial_base_length: float = 0.8, max_base_length: float = 1.6, min_base_length: float = 0.0078125, success_tolerance: int = 3, acquisition_function: AcquisitionFunctionType = AcquisitionFunctionType.THOMPSON_SAMPLING, noise: float = 0.0, embedding_type: EmbeddingType = EmbeddingType.BAXUS, success_decision_factor: float = 0.001)¶

Bases: object

The behavior of the embedded TuRBO algorithm

acquisition_function: AcquisitionFunctionType = 2¶

only Thompson sampling)

Type: The different acquisition functions to use in a multi-batch setting (default

property conf_dict: Dict[str, Any]¶

The configuration as a dictionary.

Returns: The configuration as a dictionary.

embedding_type: EmbeddingType = 0¶: Uniform bin sizing means that all target bins have approx. equally many contributing input dimensions. Random bin sizing means that a random target dimension is chosen for each input dimension (standard HeSBO behavior).

initial_base_length: float = 0.8¶: The initial base side length (see TuRBO paper)

max_base_length: float = 1.6¶: The maximum base side length (see TuRBO paper)

min_base_length: float = 0.0078125¶: The minimum base side length (see TuRBO paper). If you get lower than this, the trust region dies out.

noise: float = 0.0¶: The noise of the problem.

pretty_print() → str¶

A nice string of the configuration.

Returns: A nice string of the configuration.

success_decision_factor: float = 0.001¶: The difference wrt to the current incumbent solution required for a next point to be considered a success.

success_tolerance: int = 3¶: The number of times we consecutively have to find a better point in order to expand the trust region, initial value

baxus.util.behaviors.embedding_configuration module¶

class baxus.util.behaviors.embedding_configuration.EmbeddingType(value)¶

Bases: Enum

An enumeration.

BAXUS = 0¶: BAxUS embedding where each target bin has approx. the same number of contributing input dimensions.

HESBO = 1¶: HeSBO embedding where a target dimension is sampled for each input dimension.

baxus.util.behaviors.gp_configuration module¶

class baxus.util.behaviors.gp_configuration.GPBehaviour(mll_estimation: baxus.util.behaviors.gp_configuration.MLLEstimation = <MLLEstimation.LHS_PICK_BEST_START_GD: 2>, n_initial_samples: int = 50, n_best_on_lhs_selection: int = 5, n_mle_training_steps: int = 50)¶