Posterior Handler

gpjax.decision_making.posterior_handler

LikelihoodBuilder = Callable[[int], AbstractLikelihood] module-attribute

Type alias for likelihood builders, which take the number of datapoints as input and return a likelihood object initialised with the given number of datapoints.

PosteriorHandler dataclass

Class for handling the creation and updating of a GP posterior as new data is observed.

Attributes:

Name	Type	Description
prior	AbstractPrior	Prior to use when forming the posterior.
likelihood_builder	LikelihoodBuilder	Function which takes the number of
optimization_objective	AbstractObjective	Objective to use for optimizing the
optimizer	GradientTransformation	Optax optimizer to use for optimizing the
num_optimization_iterations	int	Number of iterations to optimize

prior: AbstractPrior instance-attribute

likelihood_builder: LikelihoodBuilder instance-attribute

optimization_objective: AbstractObjective instance-attribute

optimizer: ox.GradientTransformation instance-attribute

num_optimization_iters: int instance-attribute

__init__(prior: AbstractPrior, likelihood_builder: LikelihoodBuilder, optimization_objective: AbstractObjective, optimizer: ox.GradientTransformation, num_optimization_iters: int) -> None

__post_init__()

get_posterior(dataset: Dataset, optimize: bool, key: Optional[KeyArray] = None) -> AbstractPosterior

Initialise (and optionally optimize) a posterior using the given dataset.

Parameters:

Name	Type	Description	Default
dataset	Dataset	Dataset to get posterior for.	required
optimize	bool	Whether to optimize the posterior hyperparameters.	required
key	Optional[KeyArray]	A JAX PRNG key which is used for optimizing the posterior	None

Returns:

Type	Description
AbstractPosterior	Posterior for the given dataset.

update_posterior(dataset: Dataset, previous_posterior: AbstractPosterior, optimize: bool, key: Optional[KeyArray] = None) -> AbstractPosterior

Update the given posterior with the given dataset. This needs to be done when the number of datapoints in the (training) dataset of the posterior changes, as the AbstractLikelihood class requires the number of datapoints to be specified. Hyperparameters may or may not be optimized, depending on the value of the optimize parameter. Note that the updated poterior will be initialised with the same prior hyperparameters as the previous posterior, but the likelihood will be re-initialised with the new number of datapoints, and hyperparameters set as in the likelihood_builder function.

Parameters:

Name	Type	Description	Default
dataset	Dataset	Dataset to get posterior for.	required
previous_posterior	AbstractPosterior	Posterior being updated. This is supplied as one may	required
optimize	bool	Whether to optimize the posterior hyperparameters.	required
key	Optional[KeyArray]	A JAX PRNG key which is used for optimizing the posterior	None