gpjax 0.9.4__py3-none-any.whl → 0.9.5__py3-none-any.whl

gpjax/decision_making/utility_functions/expected_improvement.py

@@ -1,112 +0,0 @@
-# Copyright 2024 The JaxGaussianProcesses 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 dataclasses import dataclass
-from functools import partial
-
-from beartype.typing import Mapping
-import jax.numpy as jnp
-import tensorflow_probability.substrates.jax as tfp
-
-from gpjax.dataset import Dataset
-from gpjax.decision_making.utility_functions.base import (
-    AbstractSinglePointUtilityFunctionBuilder,
-    SinglePointUtilityFunction,
-)
-from gpjax.decision_making.utils import (
-    OBJECTIVE,
-    get_best_latent_observation_val,
-)
-from gpjax.gps import ConjugatePosterior
-from gpjax.typing import (
-    Array,
-    Float,
-    KeyArray,
-)
-
-
-@dataclass
-class ExpectedImprovement(AbstractSinglePointUtilityFunctionBuilder):
-    """
-    Expected Improvement acquisition function as introduced by [Močkus,
-    1974](https://link.springer.com/chapter/10.1007/3-540-07165-2_55). The "best"
-    incumbent value is defined as the lowest posterior mean value evaluated at the the
-    previously observed points. This enables the acquisition function to be utilised with noisy observations.
-    """
-
-    def build_utility_function(
-        self,
-        posteriors: Mapping[str, ConjugatePosterior],
-        datasets: Mapping[str, Dataset],
-        key: KeyArray,
-    ) -> SinglePointUtilityFunction:
-        r"""
-        Build the Expected Improvement acquisition function. This computes the expected
-        improvement over the "best" of the previously observed points, utilising the
-        posterior distribution of the surrogate model. For posterior distribution
-        $`f(\cdot)`$, and best incumbent value $`\eta`$, this is defined
-        as:
-        ```math
-        \alpha_{\text{EI}}(\mathbf{x}) = \mathbb{E}\left[\max(0, \eta - f(\mathbf{x}))\right]
-        ```
-
-        Args:
-            posteriors (Mapping[str, ConjugatePosterior]): Dictionary of posteriors to
-            used to form the utility function. One posteriors must correspond to the
-            `OBJECTIVE` key, as we utilise the objective posterior to form the utility
-            function.
-            datasets (Mapping[str, Dataset]): Dictionary of datasets used to form the
-            utility function. Keys in `datasets` should correspond to keys in
-            `posteriors`. One of the datasets must correspond to the `OBJECTIVE` key.
-            key (KeyArray): JAX PRNG key used for random number generation.
-
-        Returns:
-            SinglePointUtilityFunction: The Expected Improvement acquisition function to
-            to be *maximised* in order to decide which point to query next.
-        """
-        self.check_objective_present(posteriors, datasets)
-        objective_posterior = posteriors[OBJECTIVE]
-        objective_dataset = datasets[OBJECTIVE]
-
-        if not isinstance(objective_posterior, ConjugatePosterior):
-            raise ValueError(
-                "Objective posterior must be a ConjugatePosterior to compute the Expected Improvement."
-            )
-
-        if (
-            objective_dataset.X is None
-            or objective_dataset.n == 0
-            or objective_dataset.y is None
-        ):
-            raise ValueError("Objective dataset must contain at least one item")
-
-        eta = get_best_latent_observation_val(objective_posterior, objective_dataset)
-        return partial(
-            _expected_improvement, objective_posterior, objective_dataset, eta
-        )
-
-
-def _expected_improvement(
-    objective_posterior: ConjugatePosterior,
-    objective_dataset: Dataset,
-    eta: Float[Array, ""],
-    x: Float[Array, "N D"],
-) -> Float[Array, "N 1"]:
-    latent_dist = objective_posterior(x, objective_dataset)
-    mean = latent_dist.mean()
-    var = latent_dist.variance()
-    normal = tfp.distributions.Normal(mean, jnp.sqrt(var))
-    return jnp.expand_dims(
-        ((eta - mean) * normal.cdf(eta) + var * normal.prob(eta)), -1
-    )

gpjax/decision_making/utility_functions/probability_of_improvement.py

@@ -1,125 +0,0 @@
-# Copyright 2024 The JaxGaussianProcesses 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 dataclasses import dataclass
-
-from beartype.typing import Mapping
-from jaxtyping import Num
-import tensorflow_probability.substrates.jax as tfp
-
-from gpjax.dataset import Dataset
-from gpjax.decision_making.utility_functions.base import (
-    AbstractSinglePointUtilityFunctionBuilder,
-    SinglePointUtilityFunction,
-)
-from gpjax.decision_making.utils import (
-    OBJECTIVE,
-    get_best_latent_observation_val,
-)
-from gpjax.gps import ConjugatePosterior
-from gpjax.typing import (
-    Array,
-    KeyArray,
-)
-
-
-@dataclass
-class ProbabilityOfImprovement(AbstractSinglePointUtilityFunctionBuilder):
-    r"""
-    An acquisition function which returns the probability of improvement
-    of the objective function over the best observed value.
-
-    More precisely, given a predictive posterior distribution of the objective
-    function $`f`$, the probability of improvement at a test point $`x`$ is defined as:
-    $$`\text{PI}(x) = \text{Prob}[f(x) < f(x_{\text{best}})]`$$
-    where $`x_{\text{best}}`$ is the minimiser of the posterior mean
-    at previously observed values (to handle noisy observations).
-
-    The probability of improvement can be easily computed using the
-    cumulative distribution function of the standard normal distribution $`\Phi`$:
-    $$`\text{PI}(x) = \Phi\left(\frac{f(x_{\text{best}}) - \mu}{\sigma}\right)`$$
-    where $`\mu`$ and $`\sigma`$ are the mean and standard deviation of the
-    predictive distribution of the objective function at $`x`$.
-
-    References
-    ----------
-    [1] Kushner, H. J. (1964).
-    A new method of locating the maximum point of an arbitrary multipeak curve in the presence of noise.
-    Journal of Basic Engineering, 86(1), 97-106.
-
-    [2] Shahriari, B., Swersky, K., Wang, Z., Adams, R. P., & de Freitas, N. (2016).
-    Taking the human out of the loop: A review of Bayesian optimization.
-    Proceedings of the IEEE, 104(1), 148-175. doi: 10.1109/JPROC.2015.2494218
-    """
-
-    def build_utility_function(
-        self,
-        posteriors: Mapping[str, ConjugatePosterior],
-        datasets: Mapping[str, Dataset],
-        key: KeyArray,
-    ) -> SinglePointUtilityFunction:
-        """
-        Constructs the probability of improvement utility function
-        using the predictive posterior of the objective function.
-
-        Args:
-            posteriors (Mapping[str, AbstractPosterior]): Dictionary of posteriors to be
-            used to form the utility function. One of the posteriors must correspond
-            to the `OBJECTIVE` key, as we sample from the objective posterior to form
-            the utility function.
-            datasets (Mapping[str, Dataset]): Dictionary of datasets which may be used
-            to form the utility function. Keys in `datasets` should correspond to
-            keys in `posteriors`. One of the datasets must correspond
-            to the `OBJECTIVE` key.
-            key (KeyArray): JAX PRNG key used for random number generation. Since
-            the probability of improvement is computed deterministically
-            from the predictive posterior, the key is not used.
-
-        Returns:
-            SinglePointUtilityFunction: the probability of improvement utility function.
-        """
-        self.check_objective_present(posteriors, datasets)
-
-        objective_posterior = posteriors[OBJECTIVE]
-        if not isinstance(objective_posterior, ConjugatePosterior):
-            raise ValueError(
-                "Objective posterior must be a ConjugatePosterior to compute the Probability of Improvement using a Gaussian CDF."
-            )
-
-        objective_dataset = datasets[OBJECTIVE]
-        if (
-            objective_dataset.X is None
-            or objective_dataset.n == 0
-            or objective_dataset.y is None
-        ):
-            raise ValueError(
-                "Objective dataset must be non-empty to compute the "
-                "Probability of Improvement (since we need a "
-                "`best_y` value)."
-            )
-
-        def probability_of_improvement(x_test: Num[Array, "N D"]):
-            best_y = get_best_latent_observation_val(
-                objective_posterior, objective_dataset
-            )
-            predictive_dist = objective_posterior.predict(x_test, objective_dataset)
-
-            normal_dist = tfp.distributions.Normal(
-                loc=predictive_dist.mean(),
-                scale=predictive_dist.stddev(),
-            )
-
-            return normal_dist.cdf(best_y).reshape(-1, 1)
-
-        return probability_of_improvement

gpjax/decision_making/utility_functions/thompson_sampling.py

@@ -1,101 +0,0 @@
-# Copyright 2023 The JaxGaussianProcesses 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 dataclasses import dataclass
-
-from beartype.typing import Mapping
-
-from gpjax.dataset import Dataset
-from gpjax.decision_making.utility_functions.base import (
-    AbstractSinglePointUtilityFunctionBuilder,
-    SinglePointUtilityFunction,
-)
-from gpjax.decision_making.utils import OBJECTIVE
-from gpjax.gps import ConjugatePosterior
-from gpjax.typing import KeyArray
-
-
-@dataclass
-class ThompsonSampling(AbstractSinglePointUtilityFunctionBuilder):
-    """
-    Form a utility function by drawing an approximate sample from the posterior,
-    using decoupled sampling as introduced in [Wilson et. al.
-    (2020)](https://arxiv.org/abs/2002.09309). Note that we return the *negative* of the
-    sample as the utility function, as utility functions are *maximised*.
-
-    Note that this is a single batch utility function, as it doesn't support classical
-    batching. However, Thompson sampling can be used in a batched setting by drawing a
-    batch of different samples from the GP posterior. This can be done by calling
-    `build_utility_function` with different keys, an example of which can be seen in the
-    `ask` method of the `UtilityDrivenDecisionMaker` class. The samples can then be
-    optimised sequentially.
-
-    Attributes:
-        num_features (int): The number of random Fourier features to use when drawing
-            the approximate sample from the posterior. Defaults to 100.
-    """
-
-    num_features: int = 100
-
-    def __post_init__(self):
-        if self.num_features <= 0:
-            raise ValueError(
-                "The number of random Fourier features must be a positive integer."
-            )
-
-    def build_utility_function(
-        self,
-        posteriors: Mapping[str, ConjugatePosterior],
-        datasets: Mapping[str, Dataset],
-        key: KeyArray,
-    ) -> SinglePointUtilityFunction:
-        """
-        Draw an approximate sample from the posterior of the objective model and return
-        the *negative* of this sample as a utility function, as utility functions
-        are *maximised*.
-
-        Args:
-            posteriors (Mapping[str, ConjugatePosterior]): Dictionary of posteriors to
-            be used to form the utility function. One of the posteriors must correspond
-            to the `OBJECTIVE` key, as we sample from the objective posterior to form
-            the utility function.
-            datasets (Mapping[str, Dataset]): Dictionary of datasets which may be used
-            to form the utility function. Keys in `datasets` should correspond to
-            keys in `posteriors`. One of the datasets must correspond
-            to the `OBJECTIVE` key.
-            key (KeyArray): JAX PRNG key used for random number generation. This can be
-            changed to draw different samples.
-
-        Returns:
-            SinglePointUtilityFunction: An appproximate sample from the objective model
-                posterior to to be *maximised* in order to decide which point to query
-                next.
-        """
-        self.check_objective_present(posteriors, datasets)
-
-        objective_posterior = posteriors[OBJECTIVE]
-        if not isinstance(objective_posterior, ConjugatePosterior):
-            raise ValueError(
-                "Objective posterior must be a ConjugatePosterior to draw an approximate sample."
-            )
-
-        objective_dataset = datasets[OBJECTIVE]
-        thompson_sample = objective_posterior.sample_approx(
-            num_samples=1,
-            train_data=objective_dataset,
-            key=key,
-            num_features=self.num_features,
-        )
-
-        return lambda x: -1.0 * thompson_sample(x)  # Utility functions are *maximised*

gpjax/decision_making/utility_maximizer.py

@@ -1,157 +0,0 @@
-# Copyright 2023 The JaxGaussianProcesses 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
-
-import jax.numpy as jnp
-import jax.random as jr
-from jaxopt import ScipyBoundedMinimize
-
-from gpjax.decision_making.search_space import (
-    AbstractSearchSpace,
-    ContinuousSearchSpace,
-)
-from gpjax.decision_making.utility_functions import SinglePointUtilityFunction
-from gpjax.typing import (
-    Array,
-    Float,
-    KeyArray,
-    ScalarFloat,
-)
-
-
-def _get_discrete_maximizer(
-    query_points: Float[Array, "N D"], utility_function: SinglePointUtilityFunction
-) -> Float[Array, "1 D"]:
-    """Get the point which maximises the utility function evaluated at a given set of points.
-
-    Args:
-        query_points: set of points at which to evaluate the utility function, as an array
-            of shape `[n_points, n_dims]`.
-        utility_function: the single point utility function to be evaluated at `query_points`.
-
-    Returns:
-        Array of shape `[1, n_dims]` representing the point which maximises the utility function.
-    """
-    utility_function_values = utility_function(query_points)
-    max_utility_function_value_idx = jnp.argmax(
-        utility_function_values, axis=0, keepdims=True
-    )
-    best_sample_point = jnp.take_along_axis(
-        query_points, max_utility_function_value_idx, axis=0
-    )
-    return best_sample_point
-
-
-@dataclass
-class AbstractSinglePointUtilityMaximizer(ABC):
-    """Abstract base class for single point utility function maximizers."""
-
-    @abstractmethod
-    def maximize(
-        self,
-        utility_function: SinglePointUtilityFunction,
-        search_space: AbstractSearchSpace,
-        key: KeyArray,
-    ) -> Float[Array, "1 D"]:
-        """Maximize the given utility function over the search space provided.
-
-        Args:
-            utility_function: utility function to be maximized.
-            search_space: search space over which to maximize the utility function.
-            key: JAX PRNG key.
-
-        Returns:
-            Float[Array, "1 D"]: Point at which the utility function is maximized.
-        """
-        raise NotImplementedError
-
-
-@dataclass
-class ContinuousSinglePointUtilityMaximizer(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_samples`initial points.
-    """
-
-    num_initial_samples: int
-    num_restarts: int
-
-    def __post_init__(self):
-        if self.num_initial_samples < 1:
-            raise ValueError(
-                f"num_initial_samples must be greater than 0, got {self.num_initial_samples}."
-            )
-        elif self.num_restarts < 1:
-            raise ValueError(
-                f"num_restarts must be greater than 0, got {self.num_restarts}."
-            )
-
-    def maximize(
-        self,
-        utility_function: SinglePointUtilityFunction,
-        search_space: ContinuousSearchSpace,
-        key: KeyArray,
-    ) -> Float[Array, "1 D"]:
-        max_observed_utility_function_value = None
-        maximizer = None
-
-        for _ in range(self.num_restarts):
-            key, _ = jr.split(key)
-            initial_sample_points = search_space.sample(
-                self.num_initial_samples, key=key
-            )
-            best_initial_sample_point = _get_discrete_maximizer(
-                initial_sample_points, utility_function
-            )
-
-            def _scalar_utility_function(x: Float[Array, "1 D"]) -> ScalarFloat:
-                """
-                The Jaxopt minimizer requires a function which returns a scalar. It calls the
-                utility function with one point at a time, so the utility function
-                returns an array of shape [1, 1], so  we index to return a scalar. Note that
-                we also return the negative of the utility function - this is because
-                utility functions should be *maximimized* but the Jaxopt minimizer
-                minimizes functions.
-                """
-                return -utility_function(x)[0][0]
-
-            lbfgsb = ScipyBoundedMinimize(
-                fun=_scalar_utility_function, method="l-bfgs-b"
-            )
-            bounds = (search_space.lower_bounds, search_space.upper_bounds)
-            optimized_point = lbfgsb.run(
-                best_initial_sample_point, bounds=bounds
-            ).params
-            optimized_utility_function_value = _scalar_utility_function(optimized_point)
-            if (max_observed_utility_function_value is None) or (
-                optimized_utility_function_value > max_observed_utility_function_value
-            ):
-                max_observed_utility_function_value = optimized_utility_function_value
-                maximizer = optimized_point
-        return maximizer
-
-
-AbstractUtilityMaximizer = AbstractSinglePointUtilityMaximizer
-"""
-Type alias for a utility maximizer. Currently we only support single point utility
-functions, but in future may support batched utility functions.
-"""

gpjax/decision_making/utils.py

@@ -1,64 +0,0 @@
-# Copyright 2023 The JaxGaussianProcesses 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 beartype.typing import (
-    Callable,
-    Dict,
-    Final,
-)
-import jax.numpy as jnp
-
-from gpjax.dataset import Dataset
-from gpjax.gps import AbstractPosterior
-from gpjax.typing import (
-    Array,
-    Float,
-)
-
-OBJECTIVE: Final[str] = "OBJECTIVE"
-"""
-Tag for the objective dataset/function in standard utility functions.
-"""
-
-
-FunctionEvaluator = Callable[[Float[Array, "N D"]], Dict[str, Dataset]]
-"""
-Type alias for function evaluators, which take an array of points of shape $[N, D]$
-and evaluate a set of functions at each point, returning a mapping from function tags
-to datasets of the evaluated points. This is the same as the `Observer` in Trieste:
-https://github.com/secondmind-labs/trieste/blob/develop/trieste/observer.py
-"""
-
-
-def build_function_evaluator(
-    functions: Dict[str, Callable[[Float[Array, "N D"]], Float[Array, "N 1"]]],
-) -> FunctionEvaluator:
-    """
-    Takes a dictionary of functions and returns a `FunctionEvaluator` which can be
-    used to evaluate each of the functions at a supplied set of points and return a
-    dictionary of datasets storing the evaluated points.
-    """
-    return lambda x: {tag: Dataset(x, f(x)) for tag, f in functions.items()}
-
-
-def get_best_latent_observation_val(
-    posterior: AbstractPosterior, dataset: Dataset
-) -> Float[Array, ""]:
-    """
-    Takes a posterior and dataset and returns the best (latent) function value in the
-    dataset, corresponding to the minimum of the posterior mean value evaluated at
-    locations in the dataset. In the noiseless case, this corresponds to the minimum
-    value in the dataset.
-    """
-    return jnp.min(posterior(dataset.X, dataset).mean())