Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add logic for running a decision making loop #365

Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion gpjax/citation.py
Original file line number Diff line number Diff line change
Expand Up @@ -10,11 +10,11 @@
from jaxlib.xla_extension import PjitFunction
from plum import dispatch

from gpjax.decision_making.acquisition_functions import ThompsonSampling
from gpjax.decision_making.test_functions import (
Forrester,
LogarithmicGoldsteinPrice,
)
from gpjax.decision_making.utility_functions import ThompsonSampling
from gpjax.kernels import (
RFF,
ArcCosine,
Expand Down
32 changes: 20 additions & 12 deletions gpjax/decision_making/__init__.py
Original file line number Diff line number Diff line change
Expand Up @@ -12,14 +12,9 @@
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
from gpjax.decision_making.acquisition_functions import (
AbstractAcquisitionFunctionBuilder,
AcquisitionFunction,
ThompsonSampling,
)
from gpjax.decision_making.acquisition_maximizer import (
AbstractAcquisitionMaximizer,
ContinuousAcquisitionMaximizer,
from gpjax.decision_making.decision_maker import (
AbstractDecisionMaker,
UtilityDrivenDecisionMaker,
)
from gpjax.decision_making.posterior_handler import PosteriorHandler
from gpjax.decision_making.search_space import (
Expand All @@ -32,14 +27,27 @@
LogarithmicGoldsteinPrice,
Quadratic,
)
from gpjax.decision_making.utility_functions import (
AbstractUtilityFunctionBuilder,
ThompsonSampling,
UtilityFunction,
)
from gpjax.decision_making.utility_maximizer import (
AbstractUtilityMaximizer,
ContinuousUtilityMaximizer,
)
from gpjax.decision_making.utils import build_function_evaluator

__all__ = [
"AbstractAcquisitionFunctionBuilder",
"AbstractAcquisitionMaximizer",
"AbstractUtilityFunctionBuilder",
"AbstractUtilityMaximizer",
"AbstractDecisionMaker",
"AbstractSearchSpace",
"AcquisitionFunction",
"ContinuousAcquisitionMaximizer",
"UtilityFunction",
"build_function_evaluator",
"ContinuousUtilityMaximizer",
"ContinuousSearchSpace",
"UtilityDrivenDecisionMaker",
"AbstractContinuousTestFunction",
"Forrester",
"LogarithmicGoldsteinPrice",
Expand Down
249 changes: 249 additions & 0 deletions gpjax/decision_making/decision_maker.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,249 @@
# Copyright 2023 The GPJax Contributors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
from abc import (
ABC,
abstractmethod,
)
from dataclasses import dataclass

from beartype.typing import (
Callable,
Dict,
List,
Mapping,
)
import jax.random as jr

from gpjax.dataset import Dataset
from gpjax.decision_making.posterior_handler import PosteriorHandler
from gpjax.decision_making.search_space import AbstractSearchSpace
from gpjax.decision_making.utility_functions import AbstractUtilityFunctionBuilder
from gpjax.decision_making.utility_maximizer import AbstractUtilityMaximizer
from gpjax.decision_making.utils import FunctionEvaluator
from gpjax.gps import AbstractPosterior
from gpjax.typing import (
Array,
Float,
KeyArray,
)


@dataclass
class AbstractDecisionMaker(ABC):
"""
AbstractDecisionMaker abstract base class which handles the core decision making
loop, where we sequentially decide on points to query our function of interest at.
The decision making loop is split into two key steps, `ask` and `tell`. The `ask`
step is typically used to decide which point to query next. The `tell` step is
typically used to update models and datasets with newly queried points. These steps
can be combined in a 'run' loop which alternates between asking which point to query
next and telling the decision maker about the newly queried point having evaluated
the black-box function of interest at this point.
Attributes:
search_space (AbstractSearchSpace): Search space over which we can evaluate the
function(s) of interest.
posterior_handlers (Dict[str, PosteriorHandler]): Dictionary of posterior
handlers, which are used to update posteriors throughout the decision making
Thomas-Christie marked this conversation as resolved.
Show resolved Hide resolved
loop. Note that the word `posteriors` is used for consistency with GPJax, but these
objects are typically referred to as `models` in the model-based decision
making literature. Tags are used to distinguish between posteriors. In a typical
Bayesian optimisation setup one of the tags will be `OBJECTIVE`, defined in
decision_making.utils.
datasets (Dict[str, Dataset]): Dictionary of datasets, which are augmented with
observations throughout the decision making loop. In a typical setup they are
also used to update the posteriors, using the `posterior_handlers`. Tags are used
to distinguish datasets, and correspond to tags in `posterior_handlers`.
key (KeyArray): JAX random key, used to generate random numbers.
post_ask (List[Callable]): List of functions to be executed after each ask step.
post_tell (List[Callable]): List of functions to be executed after each tell
step.
"""

search_space: AbstractSearchSpace
posterior_handlers: Dict[str, PosteriorHandler]
datasets: Dict[str, Dataset]
key: KeyArray
post_ask: List[
Callable
] # Specific type is List[Callable[[AbstractDecisionMaker, Float[Array, ["1 D"]]], None]] but causes Beartype issues
post_tell: List[
Callable
] # Specific type is List[Callable[[AbstractDecisionMaker], None]] but causes Beartype issues

def __post_init__(self):
Thomas-Christie marked this conversation as resolved.
Show resolved Hide resolved
"""
At initialisation we check that the posterior handlers and datasets are
consistent (i.e. have the same tags), and then initialise the posteriors, optimizing them using the
corresponding datasets.
"""
# Check that posterior handlers and datasets are consistent
if self.posterior_handlers.keys() != self.datasets.keys():
raise ValueError(
"Posterior handlers and datasets must have the same keys. "
f"Got posterior handlers keys {self.posterior_handlers.keys()} and "
f"datasets keys {self.datasets.keys()}."
)

# Initialize posteriors
self.posteriors: Dict[str, AbstractPosterior] = {}
for tag, posterior_handler in self.posterior_handlers.items():
self.posteriors[tag] = posterior_handler.get_posterior(
self.datasets[tag], optimize=True, key=self.key
)

@abstractmethod
def ask(self, key: KeyArray) -> Float[Array, "1 D"]:
"""
Get the point to be queried next.
Args:
key (KeyArray): JAX PRNG key for controlling random state.
Returns:
Float[Array, "1 D"]: Point to be queried next
"""
raise NotImplementedError

def tell(self, observation_datasets: Mapping[str, Dataset], key: KeyArray):
"""
Add newly observed data to datasets and update the corresponding posteriors.
Args:
observation_datasets (Mapping[str, Dataset]): Dictionary of datasets
containing new observations. Tags are used to distinguish datasets, and
correspond to tags in `posterior_handlers` and `self.datasets`.
key (KeyArray): JAX PRNG key for controlling random state.
"""
if observation_datasets.keys() != self.datasets.keys():
raise ValueError(
"Observation datasets and existing datasets must have the same keys. "
f"Got observation datasets keys {observation_datasets.keys()} and "
f"existing datasets keys {self.datasets.keys()}."
)

for tag, observation_dataset in observation_datasets.items():
self.datasets[tag] += observation_dataset

for tag, posterior_handler in self.posterior_handlers.items():
key, _ = jr.split(key)
self.posteriors[tag] = posterior_handler.update_posterior(
self.datasets[tag], self.posteriors[tag], optimize=True, key=key
)

def run(
Thomas-Christie marked this conversation as resolved.
Show resolved Hide resolved
self, n_steps: int, black_box_function_evaluator: FunctionEvaluator
) -> Mapping[str, Dataset]:
"""
Run the decision making loop continuously for for `n_steps`. This is broken down
into three main steps:
1. Call the `ask` method to get the point to be queried next.
2. Call the `black_box_function_evaluator` to evaluate the black box functions
of interest at the point chosen to be queried.
3. Call the `tell` method to update the datasets and posteriors with the newly
observed data.
In addition to this, after the `ask` step, the functions in the `post_ask` list
are executed, taking as arguments the decision maker and the point chosen to be
queried next. Similarly, after the `tell` step, the functions in the `post_tell`
list are executed, taking the decision maker as the sole argument.
Args:
n_steps (int): Number of steps to run the decision making loop for.
black_box_function_evaluator (FunctionEvaluator): Function evaluator which
evaluates the black box functions of interest at supplied points.
Returns:
Mapping[str, Dataset]: Dictionary of datasets containing the observations
made throughout the decision making loop, as well as the initial data
supplied when initialising the `DecisionMaker`.
"""
for _ in range(n_steps):
query_point = self.ask(self.key)

for post_ask_method in self.post_ask:
post_ask_method(self, query_point)

self.key, _ = jr.split(self.key)
observation_datasets = black_box_function_evaluator(query_point)
self.tell(observation_datasets, self.key)

for post_tell_method in self.post_tell:
post_tell_method(self)

return self.datasets


@dataclass
class UtilityDrivenDecisionMaker(AbstractDecisionMaker):
"""
UtilityDrivenDecisionMaker class which handles the core decision making loop in a
typical model-based decision making setup. In this setup we use surrogate model(s)
for the function(s) of interest, and define a utility function (often called the
'acquisition function' in the context of Bayesian optimisation) which characterises
how useful it would be to query a given point within the search space given the data
we have observed so far. This can then be used to decide which point(s) to query
next.
The decision making loop is split into two key steps, `ask` and `tell`. The `ask`
step forms a `UtilityFunction` from the current `posteriors` and `datasets` and
returns the point which maximises it. It also stores the formed utility function
under the attribute `self.current_utility_function` so that it can be called,
for instance for plotting, after the `ask` function has been called. The `tell` step
adds a newly queried point to the `datasets` and updates the `posteriors`.
This can be run as a typical ask-tell loop, or the `run` method can be used to run
the decision making loop for a fixed number of steps. Moreover, the `run` method executes
the functions in `post_ask` and `post_tell` after each ask and tell step
respectively. This enables the user to add custom functionality, such as the ability
to plot values of interest during the optimization process.
Attributes:
utility_function_builder (AbstractUtilityFunctionBuilder): Object which
builds utility functions from posteriors and datasets, to decide where
to query next. In a typical Bayesian optimisation setup the point chosen to
be queried next is the point which maximizes the utility function.
utility_maximizer (AbstractUtilityMaximizer): Object which maximizes
utility functions over the search space.
"""

utility_function_builder: AbstractUtilityFunctionBuilder
utility_maximizer: AbstractUtilityMaximizer

def ask(self, key: KeyArray) -> Float[Array, "1 D"]:
"""
Get updated utility function and return the point which maximises it. This
method also stores the utility function in
`self.current_utility_function` so that it can be accessed after the ask
function has been called. This is useful for non-deterministic utility
functions, which will differ between calls to `ask` due to the splitting of
`self.key`.
Args:
key (KeyArray): JAX PRNG key for controlling random state.
Returns:
Float[Array, "1 D"]: Point to be queried next.
"""
self.current_utility_function = (
self.utility_function_builder.build_utility_function(
self.posteriors, self.datasets, key
)
)

key, _ = jr.split(key)
return self.utility_maximizer.maximize(
self.current_utility_function, self.search_space, key
)
Original file line number Diff line number Diff line change
Expand Up @@ -12,16 +12,14 @@
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
from gpjax.decision_making.acquisition_functions.base import (
AbstractAcquisitionFunctionBuilder,
AcquisitionFunction,
)
from gpjax.decision_making.acquisition_functions.thompson_sampling import (
ThompsonSampling,
from gpjax.decision_making.utility_functions.base import (
AbstractUtilityFunctionBuilder,
UtilityFunction,
)
from gpjax.decision_making.utility_functions.thompson_sampling import ThompsonSampling

__all__ = [
"AcquisitionFunction",
"AbstractAcquisitionFunctionBuilder",
"UtilityFunction",
"AbstractUtilityFunctionBuilder",
"ThompsonSampling",
]
Loading