Source code for botorch.test_functions.base

#!/usr/bin/env python3
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

r"""
Base class for test functions for optimization benchmarks.
"""

from __future__ import annotations

from abc import ABC, abstractmethod
from typing import List, Optional, Tuple

import torch
from torch import Tensor
from torch.nn import Module


[docs]class BaseTestProblem(Module, ABC):
    r"""Base class for test functions."""

    dim: int
    _bounds: List[Tuple[float, float]]
    _check_grad_at_opt: bool = True

    def __init__(self, noise_std: Optional[float] = None, negate: bool = False) -> None:
        r"""Base constructor for test functions.

        Args:
            noise_std: Standard deviation of the observation noise.
            negate: If True, negate the function.
        """
        super().__init__()
        self.noise_std = noise_std
        self.negate = negate
        self.register_buffer(
            "bounds", torch.tensor(self._bounds, dtype=torch.float).transpose(-1, -2)
        )

[docs]    def forward(self, X: Tensor, noise: bool = True) -> Tensor:
        r"""Evaluate the function on a set of points.

        Args:
            X: A `batch_shape x d`-dim tensor of point(s) at which to evaluate the
                function.
            noise: If `True`, add observation noise as specified by `noise_std`.

        Returns:
            A `batch_shape`-dim tensor ouf function evaluations.
        """
        batch = X.ndimension() > 1
        X = X if batch else X.unsqueeze(0)
        f = self.evaluate_true(X=X)
        if noise and self.noise_std is not None:
            f += self.noise_std * torch.randn_like(f)
        if self.negate:
            f = -f
        return f if batch else f.squeeze(0)

[docs]    @abstractmethod
    def evaluate_true(self, X: Tensor) -> Tensor:
        r"""Evaluate the function (w/o observation noise) on a set of points."""
        pass  # pragma: no cover


[docs]class ConstrainedBaseTestProblem(BaseTestProblem, ABC):
    r"""Base class for test functions with constraints.

    In addition to one or more objectives, a problem may have a number of outcome
    constraints of the form `c_i(x) >= 0` for `i=1, ..., n_c`.

    This base class provides common functionality for such problems.
    """

    num_constraints: int
    _check_grad_at_opt: bool = False

[docs]    def evaluate_slack(self, X: Tensor, noise: bool = True) -> Tensor:
        r"""Evaluate the constraint slack on a set of points.

        Constraints `i` is assumed to be feasible at `x` if the associated slack
        `c_i(x)` is positive. Zero slack means that the constraint is active. Negative
        slack means that the constraint is violated.

        Args:
            X: A `batch_shape x d`-dim tensor of point(s) at which to evaluate the
                constraint slacks: `c_1(X), ...., c_{n_c}(X)`.
            noise: If `True`, add observation noise to the slack as specified by
                `noise_std`.

        Returns:
            A `batch_shape x n_c`-dim tensor of constraint slack (where positive slack
                corresponds to the constraint being feasible).
        """
        cons = self.evaluate_slack_true(X=X)
        if noise and self.noise_std is not None:
            # TODO: Allow different noise levels for objective and constraints (and
            # different noise levels between different constraints)
            cons += self.noise_std * torch.randn_like(cons)
        return cons

[docs]    def is_feasible(self, X: Tensor, noise: bool = True) -> Tensor:
        r"""Evaluate whether the constraints are feasible on a set of points.

        Args:
            X: A `batch_shape x d`-dim tensor of point(s) at which to evaluate the
                constraints.
            noise: If `True`, add observation noise as specified by `noise_std`.

        Returns:
            A `batch_shape`-dim boolean tensor that is `True` iff all constraint
                slacks (potentially including observation noise) are positive.
        """
        return (self.evaluate_slack(X=X, noise=noise) >= 0.0).all(dim=-1)

[docs]    @abstractmethod
    def evaluate_slack_true(self, X: Tensor) -> Tensor:
        r"""Evaluate the constraint slack (w/o observation noise) on a set of points.

        Args:
            X: A `batch_shape x d`-dim tensor of point(s) at which to evaluate the
                constraint slacks: `c_1(X), ...., c_{n_c}(X)`.

        Returns:
            A `batch_shape x n_c`-dim tensor of constraint slack (where positive slack
                corresponds to the constraint being feasible).
        """
        pass  # pragma: no cover


[docs]class MultiObjectiveTestProblem(BaseTestProblem):
    r"""Base class for test multi-objective test functions.

    TODO: add a pareto distance function that returns the distance
    between a provided point and the closest point on the true pareto front.
    """

    num_objectives: int
    _ref_point: List[float]
    _max_hv: float

    def __init__(self, noise_std: Optional[float] = None, negate: bool = False) -> None:
        r"""Base constructor for multi-objective test functions.

        Args:
            noise_std: Standard deviation of the observation noise.
            negate: If True, negate the objectives.
        """
        super().__init__(noise_std=noise_std, negate=negate)
        ref_point = torch.tensor(self._ref_point, dtype=torch.float)
        if negate:
            ref_point *= -1
        self.register_buffer("ref_point", ref_point)

    @property
    def max_hv(self) -> float:
        try:
            return self._max_hv
        except AttributeError:
            raise NotImplementedError(
                f"Problem {self.__class__.__name__} does not specify maximal "
                "hypervolume."
            )

[docs]    def gen_pareto_front(self, n: int) -> Tensor:
        r"""Generate `n` pareto optimal points."""
        raise NotImplementedError