trieste.acquisition.function.entropy#

This module contains entropy-based acquisition function builders.

Module Contents#

class MinValueEntropySearch(search_space: trieste.space.SearchSpace, num_samples: int = 5, grid_size: int = 1000, min_value_sampler: None = None)           MinValueEntropySearch(search_space: trieste.space.SearchSpace, num_samples: int = 5, grid_size: int = 1000, min_value_sampler: Optional[trieste.acquisition.sampler.ThompsonSampler[trieste.acquisition.interface.ProbabilisticModelType]] = None)[source]#

Bases: trieste.acquisition.interface.SingleModelAcquisitionBuilder[trieste.acquisition.interface.ProbabilisticModelType]

Builder for the max-value entropy search acquisition function modified for objective minimisation. MinValueEntropySearch estimates the information in the distribution of the objective minimum that would be gained by evaluating the objective at a given point.

This implementation largely follows [WJ17] and samples the objective’s minimum \(y^*\) across a large set of sampled locations via either a Gumbel sampler, an exact Thompson sampler or an approximate random Fourier feature-based Thompson sampler, with the Gumbel sampler being the cheapest but least accurate. Default behavior is to use the exact Thompson sampler.

Parameters

  • search_space – The global search space over which the optimisation is defined.

  • num_samples – Number of samples to draw from the distribution over the minimum of the objective function.

  • grid_size – Size of the grid from which to sample the min-values. We recommend scaling this with search space dimension.

  • min_value_sampler – Sampler which samples minimum values.

Raises

tf.errors.InvalidArgumentError –

If

  • num_samples or grid_size are negative.

prepare_acquisition_function(model: trieste.acquisition.interface.ProbabilisticModelType, dataset: Optional[trieste.data.Dataset] = None) → trieste.acquisition.interface.AcquisitionFunction[source]#
Parameters

  • model – The model.

  • dataset – The data from the observer.

Returns

The max-value entropy search acquisition function modified for objective minimisation. This function will raise ValueError or InvalidArgumentError if used with a batch size greater than one.

Raises

tf.errors.InvalidArgumentError – If dataset is empty.

update_acquisition_function(function: trieste.acquisition.interface.AcquisitionFunction, model: trieste.acquisition.interface.ProbabilisticModelType, dataset: Optional[trieste.data.Dataset] = None) → trieste.acquisition.interface.AcquisitionFunction[source]#
Parameters

  • function – The acquisition function to update.

  • model – The model.

  • dataset – The data from the observer.

Bases: trieste.acquisition.interface.AcquisitionFunctionClass

An AcquisitionFunctionClass is an acquisition function represented using a class rather than as a standalone function. Using a class to represent an acquisition function makes it easier to update it, to avoid having to retrace the function on every call.

Return the max-value entropy search acquisition function (adapted from [WJ17]), modified for objective minimisation. This function calculates the information gain (or change in entropy) in the distribution over the objective minimum \(y^*\), if we were to evaluate the objective at a given point.

Parameters

  • model – The model of the objective function.

  • samples – Samples from the distribution over \(y^*\).

Returns

The max-value entropy search acquisition function modified for objective minimisation. This function will raise ValueError or InvalidArgumentError if used with a batch size greater than one.

Raises

ValueError or tf.errors.InvalidArgumentError – If samples has rank less than two, or is empty.

update(samples: trieste.types.TensorType) → None[source]#

Update the acquisition function with new samples.

__call__(x: trieste.types.TensorType) → trieste.types.TensorType[source]#

Call acquisition function.

class SupportsCovarianceObservationNoise[source]#

Bases: trieste.models.gpflow.interface.SupportsCovarianceBetweenPoints, trieste.models.interfaces.SupportsGetObservationNoise, typing_extensions.Protocol

A model that supports both covariance_between_points and get_observation_noise.

class SupportsCovarianceObservationNoiseTrajectory[source]#

Bases: trieste.models.interfaces.HasTrajectorySampler, SupportsCovarianceObservationNoise, typing_extensions.Protocol

A model that supports covariance_between_points and get_observation_noise, and also has an associated trajectory sampler.

GIBBONModelType[source]#

Type variable bound to SupportsCovarianceObservationNoise.

class GIBBON(search_space: trieste.space.SearchSpace, num_samples: int = 5, grid_size: int = 1000, min_value_sampler: None = None, rescaled_repulsion: bool = True)           GIBBON(search_space: trieste.space.SearchSpace, num_samples: int = 5, grid_size: int = 1000, min_value_sampler: Optional[trieste.acquisition.sampler.ThompsonSampler[GIBBONModelType]] = None, rescaled_repulsion: bool = True)[source]#

Bases: trieste.acquisition.interface.SingleModelGreedyAcquisitionBuilder[GIBBONModelType]

The General-purpose Information-Based Bayesian Optimisation (GIBBON) acquisition function of [MLGR21]. GIBBON provides a computationally cheap approximation of the information gained about (i.e the change in entropy of) the objective function’s minimum by evaluating a batch of candidate points. Batches are built in a greedy manner.

This implementation follows [MLGR21] but is modified for function minimisation (rather than maximisation). We sample the objective’s minimum \(y^*\) across a large set of sampled locations via either a Gumbel sampler, an exact Thompson sampler or an approximate random Fourier feature-based Thompson sampler, with the Gumbel sampler being the cheapest but least accurate. Default behavior is to use the exact Thompson sampler.

Parameters

  • search_space – The global search space over which the optimisation is defined.

  • num_samples – Number of samples to draw from the distribution over the minimum of the objective function.

  • grid_size – Size of the grid from which to sample the min-values. We recommend scaling this with search space dimension.

  • min_value_sampler – Sampler which samples minimum values.

  • rescaled_repulsion – If True, then downweight GIBBON’s repulsion term to improve batch optimization performance.

Raises

tf.errors.InvalidArgumentError –

If

  • num_samples is not positive, or

  • grid_size is not positive.

prepare_acquisition_function(model: GIBBONModelType, dataset: Optional[trieste.data.Dataset] = None, pending_points: Optional[trieste.types.TensorType] = None) → trieste.acquisition.interface.AcquisitionFunction[source]#
Parameters

  • model – The model.

  • dataset – The data from the observer. Must be populated.

  • pending_points – The points we penalize with respect to.

Returns

The GIBBON acquisition function modified for objective minimisation.

Raises

tf.errors.InvalidArgumentError – If dataset is empty.

update_acquisition_function(function: trieste.acquisition.interface.AcquisitionFunction, model: GIBBONModelType, dataset: Optional[trieste.data.Dataset] = None, pending_points: Optional[trieste.types.TensorType] = None, new_optimization_step: bool = True) → trieste.acquisition.interface.AcquisitionFunction[source]#
Parameters

  • function – The acquisition function to update.

  • model – The model.

  • dataset – The data from the observer. Must be populated.

  • pending_points – Points already chosen to be in the current batch (of shape [M,D]), where M is the number of pending points and D is the search space dimension.

  • new_optimization_step – Indicates whether this call to update_acquisition_function is to start of a new optimization step, or to continue collecting batch of points for the current step. Defaults to True.

Returns

The updated acquisition function.

class GibbonAcquisition(quality_term: trieste.acquisition.interface.AcquisitionFunction, diversity_term: trieste.acquisition.interface.PenalizationFunction)[source]#

Class representing a GIBBON acquisition function.

Parameters

  • quality_term – Quality term.

  • diversity_term – Diversity term.

class gibbon_quality_term(model: SupportsCovarianceObservationNoise, samples: trieste.types.TensorType)[source]#

Bases: trieste.acquisition.interface.AcquisitionFunctionClass

An AcquisitionFunctionClass is an acquisition function represented using a class rather than as a standalone function. Using a class to represent an acquisition function makes it easier to update it, to avoid having to retrace the function on every call.

GIBBON’s quality term measures the amount of information that each individual batch element provides about the objective function’s minimal value \(y^*\) (ensuring that evaluations are targeted in promising areas of the space).

Parameters

  • model – The model of the objective function. GIBBON requires a model with a :method:covariance_between_points method and so GIBBON only supports GaussianProcessRegression models.

  • samples – Samples from the distribution over \(y^*\).

Returns

GIBBON’s quality term. This function will raise ValueError or InvalidArgumentError if used with a batch size greater than one.

Raises

  • ValueError or tf.errors.InvalidArgumentError – If samples does not have rank two, or is empty, or if model has no homoscedastic observation noise.

  • AttributeError – If model doesn’t implement covariance_between_points method.

update(samples: trieste.types.TensorType) → None[source]#

Update the acquisition function with new samples.

__call__(x: trieste.types.TensorType) → trieste.types.TensorType[source]#

Call acquisition function.

class gibbon_repulsion_term(model: SupportsCovarianceObservationNoise, pending_points: trieste.types.TensorType, rescaled_repulsion: bool = True)[source]#

Bases: trieste.acquisition.interface.UpdatablePenalizationFunction

An UpdatablePenalizationFunction builds and updates a penalization function. Defining a penalization function that can be updated avoids having to retrace on every call.

GIBBON’s repulsion term encourages diversity within the batch (achieving high values for points with low predictive correlation).

The repulsion term \(r=\log |C|\) is given by the log determinant of the predictive correlation matrix \(C\) between the m pending points and the current candidate. The predictive covariance \(V\) can be expressed as :math:V = [[v, A], [A, B]]` for a tensor \(B\) with shape [m,`m`] and so we can efficiently calculate \(|V|\) using the formula for the determinant of block matrices, i.e \(|V| = (v - A^T * B^{-1} * A) * |B|\). Note that when using GIBBON for purely sequential optimization, the repulsion term is not required.

As GIBBON’s batches are built in a greedy manner, i.e sequentially adding points to build a set of m pending points, we need only ever calculate the entropy reduction provided by adding the current candidate point to the current pending points, not the full information gain provided by evaluating all the pending points. This allows for a modest computational saving.

When performing batch BO, GIBBON’s approximation can sometimes become less accurate as its repulsion term dominates. Therefore, we follow the arguments of [MLGR21] and divide GIBBON’s repulsion term by \(B^{2}\). This behavior can be deactivated by setting rescaled_repulsion to False.

Parameters

  • model – The model of the objective function. GIBBON requires a model with a :method:covariance_between_points method and so GIBBON only supports GaussianProcessRegression models.

  • pending_points – The points already chosen in the current batch.

  • rescaled_repulsion – If True, then downweight GIBBON’s repulsion term to improve batch optimization performance.

Returns

GIBBON’s repulsion term. This function will raise ValueError or InvalidArgumentError if used with a batch size greater than one.

Raises

  • ValueError or tf.errors.InvalidArgumentError – If pending_points does not have rank two, or is empty, or if model has no homoscedastic observation noise.

  • AttributeError – If model doesn’t implement covariance_between_points method.

update(pending_points: trieste.types.TensorType, lipschitz_constant: trieste.types.TensorType = None, eta: trieste.types.TensorType = None) → None[source]#

Update the repulsion term with new variable values.

__call__(x: trieste.types.TensorType) → trieste.types.TensorType[source]#

Call penalization function..