Utility Maximizer
gpjax.decision_making.utility_maximizer
AbstractUtilityMaximizer = AbstractSinglePointUtilityMaximizer
module-attribute
Type alias for a utility maximizer. Currently we only support single point utility functions, but in future may support batched utility functions.
AbstractSinglePointUtilityMaximizer
dataclass
Bases: ABC
Abstract base class for single point utility function maximizers.
__init__() -> None
maximize(utility_function: SinglePointUtilityFunction, search_space: AbstractSearchSpace, key: KeyArray) -> Float[Array, '1 D']
abstractmethod
Maximize the given utility function over the search space provided.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
utility_function |
UtilityFunction
|
Utility function to be |
required |
search_space |
AbstractSearchSpace
|
Search space over which to maximize |
required |
key |
KeyArray
|
JAX PRNG key. |
required |
Returns:
| Type | Description |
|---|---|
Float[Array, '1 D']
|
Float[Array, "1 D"]: Point at which the utility function is maximized. |
ContinuousSinglePointUtilityMaximizer
dataclass
Bases: AbstractSinglePointUtilityMaximizer
The ContinuousUtilityMaximizer class is used to maximize utility
functions over the continuous domain with L-BFGS-B. First we sample the utility
function at num_initial_samples points from the search space, and then we run
L-BFGS-B from the best of these initial points. We run this process num_restarts
number of times, each time sampling a different random set of
num_initial_samplesinitial points.