smac.optimizer.subspaces

Classes

ChallengerListLocal(cs_local, cs_global, ...)

LocalSubspace(config_space, bounds, hps_types)

A subspace that is designed for local Bayesian Optimization.
class smac.optimizer.subspaces.ChallengerListLocal(cs_local, cs_global, challengers, config_origin, incumbent_array=None)[source]

Bases: Iterator

class smac.optimizer.subspaces.LocalSubspace(config_space, bounds, hps_types, bounds_ss_cont=None, bounds_ss_cat=None, model_local=<class 'smac.epm.gaussian_process.augmented.GloballyAugmentedLocalGaussianProcess'>, model_local_kwargs={}, acq_func_local=<class 'smac.optimizer.acquisition.EI'>, acq_func_local_kwargs=None, rng=None, initial_data=None, activate_dims=None, incumbent_array=None)[source]

Bases: abc.ABC

A subspace that is designed for local Bayesian Optimization. If bounds_ss_cont and bounds_ss_cat are not given, this subspace is equivalent to the original configuration space. Additionally, this subspace supports local BO that only works with a subset of the dimensions, where the missing values are filled by the corresponding values from incumbent_array.

Parameters

  • config_space (ConfigurationSpace) – raw Configuration space

  • bounds (List[Tuple[float, float]]) – raw bounds of the Configuration space, notice that here bounds denotes the bounds of the entire space

  • hps_types (List[int],) – types of the hyperparameters

  • bounds_ss_cont (np.ndarray(D_cont, 2)) – subspaces bounds of continuous hyperparameters, its length is the number of continuous hyperparameters

  • bounds_ss_cat (List[Tuple]) – subspaces bounds of categorical hyperparameters, its length is the number of categorical hyperparameters

  • rng (np.random.RandomState) – random state

  • model_local (BaseEPM) – model in subspace

  • model_local_kwargs (Optional[Dict]) – argument for subspace model

  • acq_func_local (AbstractAcquisitionFunction) – local acquisition function

  • acq_func_local_kwargs (Optional[Dict]) – argument for acquisition function

  • activate_dims (Optional[np.ndarray]) – activate dimensions in the subspace, if it is None, we preserve all the dimensions

  • incumbent_array (Optional[np.ndarray]) – incumbent array, used when activate_dims has less dimension and this value is used to complementary the resulted configurations

add_new_observations(X, y)[source]

Add new observations to the subspace

Parameters

  • X (np.ndarray(N,D),) – new feature vector of the observations, constructed by the global configuration space

  • y (np.ndarray(N)) – new performances of the observations

Returns

indices_in_ss – indices of data that included in subspaces

Return type

np.ndarray(N)

static fit_forbidden_to_ss(cs_local, forbidden)[source]

Fit the forbidden to subspaces. If the target forbidden can be added to subspace, we return a new forbidden with exactly the same type of the input forbidden. Otherwise, None is returned.

Parameters

  • cs_local (ConfigurationSpace) – local configuration space of the subspace

  • forbidden (AbstractForbiddenComponent) – forbidden to check

Returns

forbidden_ss – forbidden in subspaces

Return type

Optional[AbstractForbiddenComponent]

generate_challengers(**optimizer_kwargs)[source]

Generate a list of challengers that will be transformed into the global configuration space

Parameters

optimizer_kwargs (Any) – additional configurations passed to ‘self._generate_challengers’

Return type

A list of challengers in the global configuration space

normalize_input(X)[source]

Normalize X to fit the local configuration space

Parameters

X (np.ndarray(N,D)) – input X, configurations arrays

Returns

X_normalized – normalized input X

Return type

np.ndarray(N,D)

update_incumbent_array(new_incumbent)[source]

Update a new incumbent array. The array is generated from the global configuration

Parameters

new_incumbent (np.ndarray(D)) – new incumbent, which correspondences to the global configuration

Return type

None

update_model(predict_x_best=True, update_incumbent_array=False)[source]

Update the model and acquisition function parameters

Parameters

  • predict_x_best (bool,) – if the incumbent is acquired by the predicted mean of a surrogate model

  • update_incumbent_array (bool) – if the incumbent_array of this subspace is replaced with the newly updated incumbent

Return type

None

Modules

smac.optimizer.subspaces.boing_subspace

smac.optimizer.subspaces.turbo_subspace

previous

smac.optimizer.smbo

next

smac.optimizer.subspaces.boing_subspace