smac.optimizer.configuration_chooser.boing_chooser

Functions

subspace_extraction(X, challenger, model, ...)

Extract a subspace that contains at least num_min points but no more than num_max points

Classes

BOinGChooser(scenario, stats, runhistory, ...)

Interface to train the EPM and generate next configurations with both global and local models.
class smac.optimizer.configuration_chooser.boing_chooser.BOinGChooser(scenario, stats, runhistory, runhistory2epm, model, acq_optimizer, acquisition_func, rng, restore_incumbent=None, random_configuration_chooser=<smac.optimizer.configuration_chooser.random_chooser.ChooserNoCoolDown object>, predict_x_best=True, min_samples_model=1, model_local=<class 'smac.epm.gaussian_process.augmented.GloballyAugmentedLocalGaussianProcess'>, acquisition_func_local=<class 'smac.optimizer.acquisition.EI'>, model_local_kwargs=None, acquisition_func_local_kwargs=None, acq_optimizer_local=None, acq_optimizer_local_kwargs=None, max_configs_local_fracs=0.5, min_configs_local=None, do_switching=False, turbo_kwargs=None)[source]

Bases: smac.optimizer.configuration_chooser.epm_chooser.EPMChooser

Interface to train the EPM and generate next configurations with both global and local models.

Parameters

  • runhistory2epm (RunHistory2EPM4CostWithRaw,) – a transformer to transform rh to vectors, different from the rh2epm used in vanilla EPMChooser, this rh2epm object needs to provide the raw values for optimizer in different stages

  • model (smac.epm.rf_with_instances.RandomForestWithInstances) – empirical performance model (right now, we support only RandomForestWithInstances) as a global model

  • acq_optimizer (smac.optimizer.ei_optimization.AcquisitionFunctionMaximizer) – Optimizer of acquisition function of global models

  • model_local (BaseEPM,) – local empirical performance model, used in subspace

  • model_local_kwargs (Optional[Dict] = None,) – parameters for initializing a local model

  • acquisition_func_local (AbstractAcquisitionFunction,) – local acquisition function, used in subspace

  • acquisition_func_local_kwargs (Optional[Dict] = None,) – parameters for initializing a local acquisition function optimizer

  • acq_optimizer_local (Optional[AcquisitionFunctionMaximizer] = None,) – Optimizer of acquisition function of local models

  • acq_optimizer_local_kwargs (typing: Optional[Dict] = None,) – parameters for the optimizer of acquisition function of local models

  • max_configs_local_fracs (float) – The maximal number of fractions of samples to be included in the subspace. If the number of samples in the subspace is greater than this value and n_min_config_inner, the subspace will be cropped to fit the requirement

  • min_configs_local (int,) – Minimum number of samples included in the inner loop model

  • do_switching (bool) – if we want to switch between turbo and boing or do a pure BOinG search

  • turbo_kwargs (Optional[Dict] = None) – parameters for building a turbo optimizer

choose_next(incumbent_value=None)[source]
Choose next candidate solution with Bayesian optimization. We use TurBO optimizer or BOinG to suggest

the next configuration.

If we switch local model between TurBO and BOinG, we gradually increase the probability to switch to another optimizer if we cannot make further process. (Or if TurBO find a new incumbent, we will switch to BOinG to do further exploitation)

Parameters

incumbent_value (float) – Cost value of incumbent configuration (required for acquisition function); If not given, it will be inferred from runhistory or predicted; if not given and runhistory is empty, it will raise a ValueError.

Return type

Iterator

restart_TuRBOinG(X, Y, Y_raw, train_model=False)[source]

Restart a new TurBO Optimizer, the bounds of the TurBO Optimizer is determined by a RF, we randomly sample 20 points and extract subspaces that contain at least self.min_configs_local points, and we select the subspace with the largest volume to construct a turbo optimizer :type X: ndarray :param X: previous evaluated configurations :type X: np.ndarray (N, D) :type Y: ndarray :param Y: performances of previous evaluated configurations (transformed by rh2epm transformer) :type Y: np.ndarray (N,) :type Y_raw: ndarray :param Y_raw: performances of previous evaluated configurations (raw values, not transformed) :type Y_raw: np.ndarray (N,) :type train_model: bool :param train_model: if we retrain the model with the given X and Y :type train_model: bool

Return type

None

smac.optimizer.configuration_chooser.boing_chooser.subspace_extraction(X, challenger, model, num_min, num_max, bounds, cat_dims, cont_dims)[source]

Extract a subspace that contains at least num_min points but no more than num_max points

Parameters

  • X (np.ndarray (N, D)) – points used to train the model

  • challenger (np.ndarray (1, D)) – the challenger where the subspace would grow

  • model (RandomForestWithInstances) – a rf model

  • num_min (int) – minimal number of points to be included in the subspace

  • num_max (int) – maximal number of points to be included in the subspace

  • bounds (np.ndarray(D, 2)) – bounds of the entire space, D = D_cat + D_cont

  • cat_dims (np.ndarray (D_cat)) – categorical dimensions

  • cont_dims (np.ndarray(D_cont)) –

  • dimensions (continuous) –

Return type

Tuple[ndarray, List[Tuple], ndarray]

Returns

  • union_bounds_cont (np.ndarray(D_cont, 2),) – the continuous bounds of the subregion

  • union_bounds_cat, List[Tuple], – the categorical bounds of the subregion

  • in_ss_dims – indices of the points that lie inside the subregion

previous

smac.optimizer.configuration_chooser

next

smac.optimizer.configuration_chooser.epm_chooser