Source code for botorch.acquisition.analytic

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

r"""
Analytic Acquisition Functions that evaluate the posterior without performing
Monte-Carlo sampling.
"""

from __future__ import annotations

import math

from abc import ABC

from contextlib import nullcontext
from copy import deepcopy

from typing import Dict, Optional, Tuple, Union

import torch
from botorch.acquisition.acquisition import AcquisitionFunction
from botorch.acquisition.objective import PosteriorTransform
from botorch.exceptions import UnsupportedError
from botorch.models.gp_regression import SingleTaskGP
from botorch.models.gpytorch import GPyTorchModel
from botorch.models.model import Model
from botorch.utils.constants import get_constants_like
from botorch.utils.probability import MVNXPB
from botorch.utils.probability.utils import (
    log_ndtr as log_Phi,
    log_phi,
    log_prob_normal_in,
    ndtr as Phi,
    phi,
)
from botorch.utils.safe_math import log1mexp, logmeanexp
from botorch.utils.transforms import convert_to_target_pre_hook, t_batch_mode_transform
from gpytorch.likelihoods.gaussian_likelihood import FixedNoiseGaussianLikelihood
from torch import Tensor
from torch.nn.functional import pad

# the following two numbers are needed for _log_ei_helper
_neg_inv_sqrt2 = -(2**-0.5)
_log_sqrt_pi_div_2 = math.log(math.pi / 2) / 2


[docs] class AnalyticAcquisitionFunction(AcquisitionFunction, ABC): r""" Base class for analytic acquisition functions. :meta private: """ def __init__( self, model: Model, posterior_transform: Optional[PosteriorTransform] = None, ) -> None: r"""Base constructor for analytic acquisition functions. Args: model: A fitted single-outcome model. posterior_transform: A PosteriorTransform. If using a multi-output model, a PosteriorTransform that transforms the multi-output posterior into a single-output posterior is required. """ super().__init__(model=model) if posterior_transform is None: if model.num_outputs != 1: raise UnsupportedError( "Must specify a posterior transform when using a " "multi-output model." ) else: if not isinstance(posterior_transform, PosteriorTransform): raise UnsupportedError( "AnalyticAcquisitionFunctions only support PosteriorTransforms." ) self.posterior_transform = posterior_transform
[docs] def set_X_pending(self, X_pending: Optional[Tensor] = None) -> None: raise UnsupportedError( "Analytic acquisition functions do not account for X_pending yet." )
def _mean_and_sigma( self, X: Tensor, compute_sigma: bool = True, min_var: float = 1e-12 ) -> Tuple[Tensor, Optional[Tensor]]: """Computes the first and second moments of the model posterior. Args: X: `batch_shape x q x d`-dim Tensor of model inputs. compute_sigma: Boolean indicating whether or not to compute the second moment (default: True). min_var: The minimum value the variance is clamped too. Should be positive. Returns: A tuple of tensors containing the first and second moments of the model posterior. Removes the last two dimensions if they have size one. Only returns a single tensor of means if compute_sigma is True. """ self.to(device=X.device) # ensures buffers / parameters are on the same device posterior = self.model.posterior( X=X, posterior_transform=self.posterior_transform ) mean = posterior.mean.squeeze(-2).squeeze(-1) # removing redundant dimensions if not compute_sigma: return mean, None sigma = posterior.variance.clamp_min(min_var).sqrt().view(mean.shape) return mean, sigma
[docs] class LogProbabilityOfImprovement(AnalyticAcquisitionFunction): r"""Single-outcome Log Probability of Improvement. Logarithm of the probability of improvement over the current best observed value, computed using the analytic formula under a Normal posterior distribution. Only supports the case of q=1. Requires the posterior to be Gaussian. The model must be single-outcome. The logarithm of the probability of improvement is numerically better behaved than the original function, which can lead to significantly improved optimization of the acquisition function. This is analogous to the common practice of optimizing the *log* likelihood of a probabilistic model - rather the likelihood - for the sake of maximium likelihood estimation. `logPI(x) = log(P(y >= best_f)), y ~ f(x)` Example: >>> model = SingleTaskGP(train_X, train_Y) >>> LogPI = LogProbabilityOfImprovement(model, best_f=0.2) >>> log_pi = LogPI(test_X) """ _log: bool = True def __init__( self, model: Model, best_f: Union[float, Tensor], posterior_transform: Optional[PosteriorTransform] = None, maximize: bool = True, ): r"""Single-outcome Probability of Improvement. Args: model: A fitted single-outcome model. best_f: Either a scalar or a `b`-dim Tensor (batch mode) representing the best function value observed so far (assumed noiseless). posterior_transform: A PosteriorTransform. If using a multi-output model, a PosteriorTransform that transforms the multi-output posterior into a single-output posterior is required. maximize: If True, consider the problem a maximization problem. """ super().__init__(model=model, posterior_transform=posterior_transform) self.register_buffer("best_f", torch.as_tensor(best_f)) self.maximize = maximize
[docs] @t_batch_mode_transform(expected_q=1) def forward(self, X: Tensor) -> Tensor: r"""Evaluate the Log Probability of Improvement on the candidate set X. Args: X: A `(b1 x ... bk) x 1 x d`-dim batched tensor of `d`-dim design points. Returns: A `(b1 x ... bk)`-dim tensor of Log Probability of Improvement values at the given design points `X`. """ mean, sigma = self._mean_and_sigma(X) u = _scaled_improvement(mean, sigma, self.best_f, self.maximize) return log_Phi(u)
[docs] class ProbabilityOfImprovement(AnalyticAcquisitionFunction): r"""Single-outcome Probability of Improvement. Probability of improvement over the current best observed value, computed using the analytic formula under a Normal posterior distribution. Only supports the case of q=1. Requires the posterior to be Gaussian. The model must be single-outcome. `PI(x) = P(y >= best_f), y ~ f(x)` Example: >>> model = SingleTaskGP(train_X, train_Y) >>> PI = ProbabilityOfImprovement(model, best_f=0.2) >>> pi = PI(test_X) """ def __init__( self, model: Model, best_f: Union[float, Tensor], posterior_transform: Optional[PosteriorTransform] = None, maximize: bool = True, ): r"""Single-outcome Probability of Improvement. Args: model: A fitted single-outcome model. best_f: Either a scalar or a `b`-dim Tensor (batch mode) representing the best function value observed so far (assumed noiseless). posterior_transform: A PosteriorTransform. If using a multi-output model, a PosteriorTransform that transforms the multi-output posterior into a single-output posterior is required. maximize: If True, consider the problem a maximization problem. """ super().__init__(model=model, posterior_transform=posterior_transform) self.register_buffer("best_f", torch.as_tensor(best_f)) self.maximize = maximize
[docs] @t_batch_mode_transform(expected_q=1) def forward(self, X: Tensor) -> Tensor: r"""Evaluate the Probability of Improvement on the candidate set X. Args: X: A `(b1 x ... bk) x 1 x d`-dim batched tensor of `d`-dim design points. Returns: A `(b1 x ... bk)`-dim tensor of Probability of Improvement values at the given design points `X`. """ mean, sigma = self._mean_and_sigma(X) u = _scaled_improvement(mean, sigma, self.best_f, self.maximize) return Phi(u)
[docs] class qAnalyticProbabilityOfImprovement(AnalyticAcquisitionFunction): r"""Approximate, single-outcome batch Probability of Improvement using MVNXPB. This implementation uses MVNXPB, a bivariate conditioning algorithm for approximating P(a <= Y <= b) for multivariate normal Y. See [Trinh2015bivariate]_. This (analytic) approximate q-PI is given by `approx-qPI(X) = P(max Y >= best_f) = 1 - P(Y < best_f), Y ~ f(X), X = (x_1,...,x_q)`, where `P(Y < best_f)` is estimated using MVNXPB. """ def __init__( self, model: Model, best_f: Union[float, Tensor], posterior_transform: Optional[PosteriorTransform] = None, maximize: bool = True, ) -> None: """qPI using an analytic approximation. Args: model: A fitted single-outcome model. best_f: Either a scalar or a `b`-dim Tensor (batch mode) representing the best function value observed so far (assumed noiseless). posterior_transform: A PosteriorTransform. If using a multi-output model, a PosteriorTransform that transforms the multi-output posterior into a single-output posterior is required. maximize: If True, consider the problem a maximization problem. """ super().__init__(model=model, posterior_transform=posterior_transform) self.maximize = maximize if not torch.is_tensor(best_f): best_f = torch.tensor(best_f) self.register_buffer("best_f", best_f)
[docs] @t_batch_mode_transform() def forward(self, X: Tensor) -> Tensor: """Evaluate approximate qPI on the candidate set X. Args: X: A `batch_shape x q x d`-dim Tensor of t-batches with `q` `d`-dim design points each Returns: A `batch_shape`-dim Tensor of approximate Probability of Improvement values at the given design points `X`, where `batch_shape'` is the broadcasted batch shape of model and input `X`. """ self.best_f = self.best_f.to(X) posterior = self.model.posterior( X=X, posterior_transform=self.posterior_transform ) covariance = posterior.distribution.covariance_matrix bounds = pad( (self.best_f.unsqueeze(-1) - posterior.distribution.mean).unsqueeze(-1), pad=(1, 0) if self.maximize else (0, 1), value=-float("inf") if self.maximize else float("inf"), ) # 1 - P(no improvement over best_f) solver = MVNXPB(covariance_matrix=covariance, bounds=bounds) return -solver.solve().expm1()
[docs] class ExpectedImprovement(AnalyticAcquisitionFunction): r"""Single-outcome Expected Improvement (analytic). Computes classic Expected Improvement over the current best observed value, using the analytic formula for a Normal posterior distribution. Unlike the MC-based acquisition functions, this relies on the posterior at single test point being Gaussian (and require the posterior to implement `mean` and `variance` properties). Only supports the case of `q=1`. The model must be single-outcome. `EI(x) = E(max(f(x) - best_f, 0)),` where the expectation is taken over the value of stochastic function `f` at `x`. Example: >>> model = SingleTaskGP(train_X, train_Y) >>> EI = ExpectedImprovement(model, best_f=0.2) >>> ei = EI(test_X) NOTE: It is *strongly* recommended to use LogExpectedImprovement instead of regular EI, because it solves the vanishing gradient problem by taking special care of numerical computations and can lead to substantially improved BO performance. """ def __init__( self, model: Model, best_f: Union[float, Tensor], posterior_transform: Optional[PosteriorTransform] = None, maximize: bool = True, ): r"""Single-outcome Expected Improvement (analytic). Args: model: A fitted single-outcome model. best_f: Either a scalar or a `b`-dim Tensor (batch mode) representing the best function value observed so far (assumed noiseless). posterior_transform: A PosteriorTransform. If using a multi-output model, a PosteriorTransform that transforms the multi-output posterior into a single-output posterior is required. maximize: If True, consider the problem a maximization problem. """ super().__init__(model=model, posterior_transform=posterior_transform) self.register_buffer("best_f", torch.as_tensor(best_f)) self.maximize = maximize
[docs] @t_batch_mode_transform(expected_q=1) def forward(self, X: Tensor) -> Tensor: r"""Evaluate Expected Improvement on the candidate set X. Args: X: A `(b1 x ... bk) x 1 x d`-dim batched tensor of `d`-dim design points. Expected Improvement is computed for each point individually, i.e., what is considered are the marginal posteriors, not the joint. Returns: A `(b1 x ... bk)`-dim tensor of Expected Improvement values at the given design points `X`. """ mean, sigma = self._mean_and_sigma(X) u = _scaled_improvement(mean, sigma, self.best_f, self.maximize) return sigma * _ei_helper(u)
[docs] class LogExpectedImprovement(AnalyticAcquisitionFunction): r"""Logarithm of single-outcome Expected Improvement (analytic). Computes the logarithm of the classic Expected Improvement acquisition function, in a numerically robust manner. In particular, the implementation takes special care to avoid numerical issues in the computation of the acquisition value and its gradient in regions where improvement is predicted to be virtually impossible. See [Ament2023logei]_ for details. Formally, `LogEI(x) = log(E(max(f(x) - best_f, 0))),` where the expectation is taken over the value of stochastic function `f` at `x`. Example: >>> model = SingleTaskGP(train_X, train_Y) >>> LogEI = LogExpectedImprovement(model, best_f=0.2) >>> ei = LogEI(test_X) """ _log: bool = True def __init__( self, model: Model, best_f: Union[float, Tensor], posterior_transform: Optional[PosteriorTransform] = None, maximize: bool = True, ): r"""Logarithm of single-outcome Expected Improvement (analytic). Args: model: A fitted single-outcome model. best_f: Either a scalar or a `b`-dim Tensor (batch mode) representing the best function value observed so far (assumed noiseless). posterior_transform: A PosteriorTransform. If using a multi-output model, a PosteriorTransform that transforms the multi-output posterior into a single-output posterior is required. maximize: If True, consider the problem a maximization problem. """ super().__init__(model=model, posterior_transform=posterior_transform) self.register_buffer("best_f", torch.as_tensor(best_f)) self.maximize = maximize
[docs] @t_batch_mode_transform(expected_q=1) def forward(self, X: Tensor) -> Tensor: r"""Evaluate logarithm of Expected Improvement on the candidate set X. Args: X: A `(b1 x ... bk) x 1 x d`-dim batched tensor of `d`-dim design points. Expected Improvement is computed for each point individually, i.e., what is considered are the marginal posteriors, not the joint. Returns: A `(b1 x ... bk)`-dim tensor of the logarithm of the Expected Improvement values at the given design points `X`. """ mean, sigma = self._mean_and_sigma(X) u = _scaled_improvement(mean, sigma, self.best_f, self.maximize) return _log_ei_helper(u) + sigma.log()
[docs] class LogConstrainedExpectedImprovement(AnalyticAcquisitionFunction): r"""Log Constrained Expected Improvement (feasibility-weighted). Computes the logarithm of the analytic expected improvement for a Normal posterior distribution weighted by a probability of feasibility. The objective and constraints are assumed to be independent and have Gaussian posterior distributions. Only supports non-batch mode (i.e. `q=1`). The model should be multi-outcome, with the index of the objective and constraints passed to the constructor. See [Ament2023logei]_ for details. Formally, `LogConstrainedEI(x) = log(EI(x)) + Sum_i log(P(y_i \in [lower_i, upper_i]))`, where `y_i ~ constraint_i(x)` and `lower_i`, `upper_i` are the lower and upper bounds for the i-th constraint, respectively. Example: # example where the 0th output has a non-negativity constraint and # the 1st output is the objective >>> model = SingleTaskGP(train_X, train_Y) >>> constraints = {0: (0.0, None)} >>> LogCEI = LogConstrainedExpectedImprovement(model, 0.2, 1, constraints) >>> cei = LogCEI(test_X) """ _log: bool = True def __init__( self, model: Model, best_f: Union[float, Tensor], objective_index: int, constraints: Dict[int, Tuple[Optional[float], Optional[float]]], maximize: bool = True, ) -> None: r"""Analytic Log Constrained Expected Improvement. Args: model: A fitted multi-output model. best_f: Either a scalar or a `b`-dim Tensor (batch mode) representing the best feasible function value observed so far (assumed noiseless). objective_index: The index of the objective. constraints: A dictionary of the form `{i: [lower, upper]}`, where `i` is the output index, and `lower` and `upper` are lower and upper bounds on that output (resp. interpreted as -Inf / Inf if None) maximize: If True, consider the problem a maximization problem. """ # Use AcquisitionFunction constructor to avoid check for posterior transform. super(AnalyticAcquisitionFunction, self).__init__(model=model) self.posterior_transform = None self.maximize = maximize self.objective_index = objective_index self.constraints = constraints self.register_buffer("best_f", torch.as_tensor(best_f)) _preprocess_constraint_bounds(self, constraints=constraints) self.register_forward_pre_hook(convert_to_target_pre_hook)
[docs] @t_batch_mode_transform(expected_q=1) def forward(self, X: Tensor) -> Tensor: r"""Evaluate Constrained Log Expected Improvement on the candidate set X. Args: X: A `(b) x 1 x d`-dim Tensor of `(b)` t-batches of `d`-dim design points each. Returns: A `(b)`-dim Tensor of Log Expected Improvement values at the given design points `X`. """ means, sigmas = self._mean_and_sigma(X) # (b) x 1 + (m = num constraints) ind = self.objective_index mean_obj, sigma_obj = means[..., ind], sigmas[..., ind] u = _scaled_improvement(mean_obj, sigma_obj, self.best_f, self.maximize) log_ei = _log_ei_helper(u) + sigma_obj.log() log_prob_feas = _compute_log_prob_feas(self, means=means, sigmas=sigmas) return log_ei + log_prob_feas
[docs] class ConstrainedExpectedImprovement(AnalyticAcquisitionFunction): r"""Constrained Expected Improvement (feasibility-weighted). Computes the analytic expected improvement for a Normal posterior distribution, weighted by a probability of feasibility. The objective and constraints are assumed to be independent and have Gaussian posterior distributions. Only supports non-batch mode (i.e. `q=1`). The model should be multi-outcome, with the index of the objective and constraints passed to the constructor. `Constrained_EI(x) = EI(x) * Product_i P(y_i \in [lower_i, upper_i])`, where `y_i ~ constraint_i(x)` and `lower_i`, `upper_i` are the lower and upper bounds for the i-th constraint, respectively. Example: # example where the 0th output has a non-negativity constraint and # 1st output is the objective >>> model = SingleTaskGP(train_X, train_Y) >>> constraints = {0: (0.0, None)} >>> cEI = ConstrainedExpectedImprovement(model, 0.2, 1, constraints) >>> cei = cEI(test_X) """ def __init__( self, model: Model, best_f: Union[float, Tensor], objective_index: int, constraints: Dict[int, Tuple[Optional[float], Optional[float]]], maximize: bool = True, ) -> None: r"""Analytic Constrained Expected Improvement. Args: model: A fitted multi-output model. best_f: Either a scalar or a `b`-dim Tensor (batch mode) representing the best feasible function value observed so far (assumed noiseless). objective_index: The index of the objective. constraints: A dictionary of the form `{i: [lower, upper]}`, where `i` is the output index, and `lower` and `upper` are lower and upper bounds on that output (resp. interpreted as -Inf / Inf if None) maximize: If True, consider the problem a maximization problem. """ # Use AcquisitionFunction constructor to avoid check for posterior transform. super(AnalyticAcquisitionFunction, self).__init__(model=model) self.posterior_transform = None self.maximize = maximize self.objective_index = objective_index self.constraints = constraints self.register_buffer("best_f", torch.as_tensor(best_f)) _preprocess_constraint_bounds(self, constraints=constraints) self.register_forward_pre_hook(convert_to_target_pre_hook)
[docs] @t_batch_mode_transform(expected_q=1) def forward(self, X: Tensor) -> Tensor: r"""Evaluate Constrained Expected Improvement on the candidate set X. Args: X: A `(b) x 1 x d`-dim Tensor of `(b)` t-batches of `d`-dim design points each. Returns: A `(b)`-dim Tensor of Expected Improvement values at the given design points `X`. """ means, sigmas = self._mean_and_sigma(X) # (b) x 1 + (m = num constraints) ind = self.objective_index mean_obj, sigma_obj = means[..., ind], sigmas[..., ind] u = _scaled_improvement(mean_obj, sigma_obj, self.best_f, self.maximize) ei = sigma_obj * _ei_helper(u) log_prob_feas = _compute_log_prob_feas(self, means=means, sigmas=sigmas) return ei.mul(log_prob_feas.exp())
[docs] class LogNoisyExpectedImprovement(AnalyticAcquisitionFunction): r"""Single-outcome Log Noisy Expected Improvement (via fantasies). This computes Log Noisy Expected Improvement by averaging over the Expected Improvement values of a number of fantasy models. Only supports the case `q=1`. Assumes that the posterior distribution of the model is Gaussian. The model must be single-outcome. See [Ament2023logei]_ for details. Formally, `LogNEI(x) = log(E(max(y - max Y_base), 0))), (y, Y_base) ~ f((x, X_base))`, where `X_base` are previously observed points. Note: This acquisition function currently relies on using a SingleTaskGP with known observation noise. In other words, `train_Yvar` must be passed to the model. (required for noiseless fantasies). Example: >>> model = SingleTaskGP(train_X, train_Y, train_Yvar=train_Yvar) >>> LogNEI = LogNoisyExpectedImprovement(model, train_X) >>> nei = LogNEI(test_X) """ _log: bool = True def __init__( self, model: GPyTorchModel, X_observed: Tensor, num_fantasies: int = 20, maximize: bool = True, posterior_transform: Optional[PosteriorTransform] = None, ) -> None: r"""Single-outcome Noisy Log Expected Improvement (via fantasies). Args: model: A fitted single-outcome model. X_observed: A `n x d` Tensor of observed points that are likely to be the best observed points so far. num_fantasies: The number of fantasies to generate. The higher this number the more accurate the model (at the expense of model complexity and performance). maximize: If True, consider the problem a maximization problem. """ # sample fantasies from botorch.sampling.normal import SobolQMCNormalSampler # Drop gradients from model.posterior if X_observed does not require gradients # as otherwise, gradients of the GP's kernel's hyper-parameters are tracked # through the rsample_from_base_sample method of GPyTorchPosterior. These # gradients are usually only required w.r.t. the marginal likelihood. with nullcontext() if X_observed.requires_grad else torch.no_grad(): posterior = model.posterior(X=X_observed) sampler = SobolQMCNormalSampler(sample_shape=torch.Size([num_fantasies])) Y_fantasized = sampler(posterior).squeeze(-1) batch_X_observed = X_observed.expand(num_fantasies, *X_observed.shape) # The fantasy model will operate in batch mode fantasy_model = _get_noiseless_fantasy_model( model=model, batch_X_observed=batch_X_observed, Y_fantasized=Y_fantasized ) super().__init__(model=fantasy_model, posterior_transform=posterior_transform) best_f, _ = Y_fantasized.max(dim=-1) if maximize else Y_fantasized.min(dim=-1) self.best_f, self.maximize = best_f, maximize
[docs] @t_batch_mode_transform(expected_q=1) def forward(self, X: Tensor) -> Tensor: r"""Evaluate logarithm of the mean Expected Improvement on the candidate set X. Args: X: A `b1 x ... bk x 1 x d`-dim batched tensor of `d`-dim design points. Returns: A `b1 x ... bk`-dim tensor of Log Noisy Expected Improvement values at the given design points `X`. """ # add batch dimension for broadcasting to fantasy models mean, sigma = self._mean_and_sigma(X.unsqueeze(-3)) u = _scaled_improvement(mean, sigma, self.best_f, self.maximize) log_ei = _log_ei_helper(u) + sigma.log() # this is mathematically - though not numerically - equivalent to log(mean(ei)) return logmeanexp(log_ei, dim=-1)
[docs] class NoisyExpectedImprovement(ExpectedImprovement): r"""Single-outcome Noisy Expected Improvement (via fantasies). This computes Noisy Expected Improvement by averaging over the Expected Improvement values of a number of fantasy models. Only supports the case `q=1`. Assumes that the posterior distribution of the model is Gaussian. The model must be single-outcome. `NEI(x) = E(max(y - max Y_baseline), 0)), (y, Y_baseline) ~ f((x, X_baseline))`, where `X_baseline` are previously observed points. Note: This acquisition function currently relies on using a SingleTaskGP with known observation noise. In other words, `train_Yvar` must be passed to the model. (required for noiseless fantasies). Example: >>> model = SingleTaskGP(train_X, train_Y, train_Yvar=train_Yvar) >>> NEI = NoisyExpectedImprovement(model, train_X) >>> nei = NEI(test_X) """ def __init__( self, model: GPyTorchModel, X_observed: Tensor, num_fantasies: int = 20, maximize: bool = True, ) -> None: r"""Single-outcome Noisy Expected Improvement (via fantasies). Args: model: A fitted single-outcome model. X_observed: A `n x d` Tensor of observed points that are likely to be the best observed points so far. num_fantasies: The number of fantasies to generate. The higher this number the more accurate the model (at the expense of model complexity and performance). maximize: If True, consider the problem a maximization problem. """ # sample fantasies from botorch.sampling.normal import SobolQMCNormalSampler # Drop gradients from model.posterior if X_observed does not require gradients # as otherwise, gradients of the GP's kernel's hyper-parameters are tracked # through the rsample_from_base_sample method of GPyTorchPosterior. These # gradients are usually only required w.r.t. the marginal likelihood. with nullcontext() if X_observed.requires_grad else torch.no_grad(): posterior = model.posterior(X=X_observed) sampler = SobolQMCNormalSampler(sample_shape=torch.Size([num_fantasies])) Y_fantasized = sampler(posterior).squeeze(-1) batch_X_observed = X_observed.expand(num_fantasies, *X_observed.shape) # The fantasy model will operate in batch mode fantasy_model = _get_noiseless_fantasy_model( model=model, batch_X_observed=batch_X_observed, Y_fantasized=Y_fantasized ) best_f, _ = Y_fantasized.max(dim=-1) if maximize else Y_fantasized.min(dim=-1) super().__init__(model=fantasy_model, best_f=best_f, maximize=maximize)
[docs] @t_batch_mode_transform(expected_q=1) def forward(self, X: Tensor) -> Tensor: r"""Evaluate Expected Improvement on the candidate set X. Args: X: A `b1 x ... bk x 1 x d`-dim batched tensor of `d`-dim design points. Returns: A `b1 x ... bk`-dim tensor of Noisy Expected Improvement values at the given design points `X`. """ # add batch dimension for broadcasting to fantasy models mean, sigma = self._mean_and_sigma(X.unsqueeze(-3)) u = _scaled_improvement(mean, sigma, self.best_f, self.maximize) return (sigma * _ei_helper(u)).mean(dim=-1)
[docs] class UpperConfidenceBound(AnalyticAcquisitionFunction): r"""Single-outcome Upper Confidence Bound (UCB). Analytic upper confidence bound that comprises of the posterior mean plus an additional term: the posterior standard deviation weighted by a trade-off parameter, `beta`. Only supports the case of `q=1` (i.e. greedy, non-batch selection of design points). The model must be single-outcome. `UCB(x) = mu(x) + sqrt(beta) * sigma(x)`, where `mu` and `sigma` are the posterior mean and standard deviation, respectively. Example: >>> model = SingleTaskGP(train_X, train_Y) >>> UCB = UpperConfidenceBound(model, beta=0.2) >>> ucb = UCB(test_X) """ def __init__( self, model: Model, beta: Union[float, Tensor], posterior_transform: Optional[PosteriorTransform] = None, maximize: bool = True, ) -> None: r"""Single-outcome Upper Confidence Bound. Args: model: A fitted single-outcome GP model (must be in batch mode if candidate sets X will be) beta: Either a scalar or a one-dim tensor with `b` elements (batch mode) representing the trade-off parameter between mean and covariance posterior_transform: A PosteriorTransform. If using a multi-output model, a PosteriorTransform that transforms the multi-output posterior into a single-output posterior is required. maximize: If True, consider the problem a maximization problem. """ super().__init__(model=model, posterior_transform=posterior_transform) self.register_buffer("beta", torch.as_tensor(beta)) self.maximize = maximize
[docs] @t_batch_mode_transform(expected_q=1) def forward(self, X: Tensor) -> Tensor: r"""Evaluate the Upper Confidence Bound on the candidate set X. Args: X: A `(b1 x ... bk) x 1 x d`-dim batched tensor of `d`-dim design points. Returns: A `(b1 x ... bk)`-dim tensor of Upper Confidence Bound values at the given design points `X`. """ mean, sigma = self._mean_and_sigma(X) return (mean if self.maximize else -mean) + self.beta.sqrt() * sigma
[docs] class PosteriorMean(AnalyticAcquisitionFunction): r"""Single-outcome Posterior Mean. Only supports the case of q=1. Requires the model's posterior to have a `mean` property. The model must be single-outcome. Example: >>> model = SingleTaskGP(train_X, train_Y) >>> PM = PosteriorMean(model) >>> pm = PM(test_X) """ def __init__( self, model: Model, posterior_transform: Optional[PosteriorTransform] = None, maximize: bool = True, ) -> None: r"""Single-outcome Posterior Mean. Args: model: A fitted single-outcome GP model (must be in batch mode if candidate sets X will be) posterior_transform: A PosteriorTransform. If using a multi-output model, a PosteriorTransform that transforms the multi-output posterior into a single-output posterior is required. maximize: If True, consider the problem a maximization problem. Note that if `maximize=False`, the posterior mean is negated. As a consequence `optimize_acqf(PosteriorMean(gp, maximize=False))` actually returns -1 * minimum of the posterior mean. """ super().__init__(model=model, posterior_transform=posterior_transform) self.maximize = maximize
[docs] @t_batch_mode_transform(expected_q=1) def forward(self, X: Tensor) -> Tensor: r"""Evaluate the posterior mean on the candidate set X. Args: X: A `(b1 x ... bk) x 1 x d`-dim batched tensor of `d`-dim design points. Returns: A `(b1 x ... bk)`-dim tensor of Posterior Mean values at the given design points `X`. """ mean, _ = self._mean_and_sigma(X, compute_sigma=False) return mean if self.maximize else -mean
[docs] class ScalarizedPosteriorMean(AnalyticAcquisitionFunction): r"""Scalarized Posterior Mean. This acquisition function returns a scalarized (across the q-batch) posterior mean given a vector of weights. """ def __init__( self, model: Model, weights: Tensor, posterior_transform: Optional[PosteriorTransform] = None, ) -> None: r"""Scalarized Posterior Mean. Args: model: A fitted single-outcome model. weights: A tensor of shape `q` for scalarization. In order to minimize the scalarized posterior mean, pass -weights. posterior_transform: A PosteriorTransform. If using a multi-output model, a PosteriorTransform that transforms the multi-output posterior into a single-output posterior is required. """ super().__init__(model=model, posterior_transform=posterior_transform) self.register_buffer("weights", weights)
[docs] @t_batch_mode_transform() def forward(self, X: Tensor) -> Tensor: r"""Evaluate the scalarized posterior mean on the candidate set X. Args: X: A `(b) x q x d`-dim Tensor of `(b)` t-batches of `d`-dim design points each. Returns: A `(b)`-dim Tensor of Posterior Mean values at the given design points `X`. """ return self._mean_and_sigma(X, compute_sigma=False)[0] @ self.weights
[docs] class PosteriorStandardDeviation(AnalyticAcquisitionFunction): r"""Single-outcome Posterior Standard Deviation. An acquisition function for pure exploration. Only supports the case of q=1. Requires the model's posterior to have `mean` and `variance` properties. The model must be either single-outcome or combined with a `posterior_transform` to produce a single-output posterior. Example: >>> import torch >>> from botorch.models.gp_regression import SingleTaskGP >>> from botorch.models.transforms.input import Normalize >>> from botorch.models.transforms.outcome import Standardize >>> >>> # Set up a model >>> train_X = torch.rand(20, 2, dtype=torch.float64) >>> train_Y = torch.sin(train_X).sum(dim=1, keepdim=True) >>> model = SingleTaskGP( ... train_X, train_Y, outcome_transform=Standardize(m=1), ... input_transform=Normalize(d=2), ... ) >>> # Now set up the acquisition function >>> PSTD = PosteriorStandardDeviation(model) >>> test_X = torch.zeros((1, 2), dtype=torch.float64) >>> std = PSTD(test_X) >>> std.item() 0.16341639895667773 """ def __init__( self, model: Model, posterior_transform: Optional[PosteriorTransform] = None, maximize: bool = True, ) -> None: r"""Single-outcome Posterior Mean. Args: model: A fitted single-outcome GP model (must be in batch mode if candidate sets X will be) posterior_transform: A PosteriorTransform. If using a multi-output model, a PosteriorTransform that transforms the multi-output posterior into a single-output posterior is required. maximize: If True, consider the problem a maximization problem. Note that if `maximize=False`, the posterior standard deviation is negated. As a consequence, `optimize_acqf(PosteriorStandardDeviation(gp, maximize=False))` actually returns -1 * minimum of the posterior standard deviation. """ super().__init__(model=model, posterior_transform=posterior_transform) self.maximize = maximize
[docs] @t_batch_mode_transform(expected_q=1) def forward(self, X: Tensor) -> Tensor: r"""Evaluate the posterior standard deviation on the candidate set X. Args: X: A `(b1 x ... bk) x 1 x d`-dim batched tensor of `d`-dim design points. Returns: A `(b1 x ... bk)`-dim tensor of Posterior Mean values at the given design points `X`. """ _, std = self._mean_and_sigma(X) return std if self.maximize else -std
# --------------- Helper functions for analytic acquisition functions. --------------- def _scaled_improvement( mean: Tensor, sigma: Tensor, best_f: Tensor, maximize: bool ) -> Tensor: """Returns `u = (mean - best_f) / sigma`, -u if maximize == True.""" u = (mean - best_f) / sigma return u if maximize else -u def _ei_helper(u: Tensor) -> Tensor: """Computes phi(u) + u * Phi(u), where phi and Phi are the standard normal pdf and cdf, respectively. This is used to compute Expected Improvement. """ return phi(u) + u * Phi(u) def _log_ei_helper(u: Tensor) -> Tensor: """Accurately computes log(phi(u) + u * Phi(u)) in a differentiable manner for u in [-10^100, 10^100] in double precision, and [-10^20, 10^20] in single precision. Beyond these intervals, a basic squaring of u can lead to floating point overflow. In contrast, the implementation in _ei_helper only yields usable gradients down to u ~ -10. As a consequence, _log_ei_helper improves the range of inputs for which a backward pass yields usable gradients by many orders of magnitude. """ if not (u.dtype == torch.float32 or u.dtype == torch.float64): raise TypeError( f"LogExpectedImprovement only supports torch.float32 and torch.float64 " f"dtypes, but received {u.dtype = }." ) # The function has two branching decisions. The first is u < bound, and in this # case, just taking the logarithm of the naive _ei_helper implementation works. bound = -1 u_upper = u.masked_fill(u < bound, bound) # mask u to avoid NaNs in gradients log_ei_upper = _ei_helper(u_upper).log() # When u <= bound, we need to be more careful and rearrange the EI formula as # log(phi(u)) + log(1 - exp(w)), where w = log(abs(u) * Phi(u) / phi(u)). # To this end, a second branch is necessary, depending on whether or not u is # smaller than approximately the negative inverse square root of the machine # precision. Below this point, numerical issues in computing log(1 - exp(w)) occur # as w approaches zero from below, even though the relative contribution to log_ei # vanishes in machine precision at that point. neg_inv_sqrt_eps = -1e6 if u.dtype == torch.float64 else -1e3 # mask u for to avoid NaNs in gradients in first and second branch u_lower = u.masked_fill(u > bound, bound) u_eps = u_lower.masked_fill(u < neg_inv_sqrt_eps, neg_inv_sqrt_eps) # compute the logarithm of abs(u) * Phi(u) / phi(u) for moderately large negative u w = _log_abs_u_Phi_div_phi(u_eps) # 1) Now, we use a special implementation of log(1 - exp(w)) for moderately # large negative numbers, and # 2) capture the leading order of log(1 - exp(w)) for very large negative numbers. # The second special case is technically only required for single precision numbers # but does "the right thing" regardless. log_ei_lower = log_phi(u) + ( torch.where( u > neg_inv_sqrt_eps, log1mexp(w), # The contribution of the next term relative to log_phi vanishes when # w_lower << eps but captures the leading order of the log1mexp term. -2 * u_lower.abs().log(), ) ) return torch.where(u > bound, log_ei_upper, log_ei_lower) def _log_abs_u_Phi_div_phi(u: Tensor) -> Tensor: """Computes log(abs(u) * Phi(u) / phi(u)), where phi and Phi are the normal pdf and cdf, respectively. The function is valid for u < 0. NOTE: In single precision arithmetic, the function becomes numerically unstable for u < -1e3. For this reason, a second branch in _log_ei_helper is necessary to handle this regime, where this function approaches -abs(u)^-2 asymptotically. The implementation is based on the following implementation of the logarithm of the scaled complementary error function (i.e. erfcx). Since we only require the positive branch for _log_ei_helper, _log_abs_u_Phi_div_phi does not have a branch, but is only valid for u < 0 (so that _neg_inv_sqrt2 * u > 0). def logerfcx(x: Tensor) -> Tensor: return torch.where( x < 0, torch.erfc(x.masked_fill(x > 0, 0)).log() + x**2, torch.special.erfcx(x.masked_fill(x < 0, 0)).log(), ) Further, it is important for numerical accuracy to move u.abs() into the logarithm, rather than adding u.abs().log() to logerfcx. This is the reason for the rather complex name of this function: _log_abs_u_Phi_div_phi. """ # get_constants_like allocates tensors with the appropriate dtype and device and # caches the result, which improves efficiency. a, b = get_constants_like(values=(_neg_inv_sqrt2, _log_sqrt_pi_div_2), ref=u) return torch.log(torch.special.erfcx(a * u) * u.abs()) + b def _get_noiseless_fantasy_model( model: SingleTaskGP, batch_X_observed: Tensor, Y_fantasized: Tensor ) -> SingleTaskGP: r"""Construct a fantasy model from a fitted model and provided fantasies. The fantasy model uses the hyperparameters from the original fitted model and assumes the fantasies are noiseless. Args: model: A fitted SingleTaskGP with known observation noise. batch_X_observed: A `b x n x d` tensor of inputs where `b` is the number of fantasies. Y_fantasized: A `b x n` tensor of fantasized targets where `b` is the number of fantasies. Returns: The fantasy model. """ if not isinstance(model, SingleTaskGP) or not isinstance( model.likelihood, FixedNoiseGaussianLikelihood ): raise UnsupportedError( "Only SingleTaskGP models with known observation noise " "are currently supported for fantasy-based NEI & LogNEI." ) # initialize a copy of SingleTaskGP on the original training inputs # this makes SingleTaskGP a non-batch GP, so that the same hyperparameters # are used across all batches (by default, a GP with batched training data # uses independent hyperparameters for each batch). fantasy_model = SingleTaskGP( train_X=model.train_inputs[0], train_Y=model.train_targets.unsqueeze(-1), train_Yvar=model.likelihood.noise_covar.noise.unsqueeze(-1), ) # update training inputs/targets to be batch mode fantasies fantasy_model.set_train_data( inputs=batch_X_observed, targets=Y_fantasized, strict=False ) # use noiseless fantasies fantasy_model.likelihood.noise_covar.noise = torch.full_like(Y_fantasized, 1e-7) # load hyperparameters from original model state_dict = deepcopy(model.state_dict()) fantasy_model.load_state_dict(state_dict) return fantasy_model def _preprocess_constraint_bounds( acqf: Union[LogConstrainedExpectedImprovement, ConstrainedExpectedImprovement], constraints: Dict[int, Tuple[Optional[float], Optional[float]]], ) -> None: r"""Set up constraint bounds. Args: constraints: A dictionary of the form `{i: [lower, upper]}`, where `i` is the output index, and `lower` and `upper` are lower and upper bounds on that output (resp. interpreted as -Inf / Inf if None) """ con_lower, con_lower_inds = [], [] con_upper, con_upper_inds = [], [] con_both, con_both_inds = [], [] con_indices = list(constraints.keys()) if len(con_indices) == 0: raise ValueError("There must be at least one constraint.") if acqf.objective_index in con_indices: raise ValueError( "Output corresponding to objective should not be a constraint." ) for k in con_indices: if constraints[k][0] is not None and constraints[k][1] is not None: if constraints[k][1] <= constraints[k][0]: raise ValueError("Upper bound is less than the lower bound.") con_both_inds.append(k) con_both.append([constraints[k][0], constraints[k][1]]) elif constraints[k][0] is not None: con_lower_inds.append(k) con_lower.append(constraints[k][0]) elif constraints[k][1] is not None: con_upper_inds.append(k) con_upper.append(constraints[k][1]) # tensor-based indexing is much faster than list-based advanced indexing for name, indices in [ ("con_lower_inds", con_lower_inds), ("con_upper_inds", con_upper_inds), ("con_both_inds", con_both_inds), ("con_both", con_both), ("con_lower", con_lower), ("con_upper", con_upper), ]: acqf.register_buffer(name, tensor=torch.as_tensor(indices)) def _compute_log_prob_feas( acqf: Union[LogConstrainedExpectedImprovement, ConstrainedExpectedImprovement], means: Tensor, sigmas: Tensor, ) -> Tensor: r"""Compute logarithm of the feasibility probability for each batch of X. Args: X: A `(b) x 1 x d`-dim Tensor of `(b)` t-batches of `d`-dim design points each. means: A `(b) x m`-dim Tensor of means. sigmas: A `(b) x m`-dim Tensor of standard deviations. Returns: A `b`-dim tensor of log feasibility probabilities Note: This function does case-work for upper bound, lower bound, and both-sided bounds. Another way to do it would be to use 'inf' and -'inf' for the one-sided bounds and use the logic for the both-sided case. But this causes an issue with autograd since we get 0 * inf. TODO: Investigate further. """ acqf.to(device=means.device) log_prob = torch.zeros_like(means[..., 0]) if len(acqf.con_lower_inds) > 0: i = acqf.con_lower_inds dist_l = (acqf.con_lower - means[..., i]) / sigmas[..., i] log_prob = log_prob + log_Phi(-dist_l).sum(dim=-1) # 1 - Phi(x) = Phi(-x) if len(acqf.con_upper_inds) > 0: i = acqf.con_upper_inds dist_u = (acqf.con_upper - means[..., i]) / sigmas[..., i] log_prob = log_prob + log_Phi(dist_u).sum(dim=-1) if len(acqf.con_both_inds) > 0: i = acqf.con_both_inds con_lower, con_upper = acqf.con_both[:, 0], acqf.con_both[:, 1] # scaled distance to lower and upper constraint boundary: dist_l = (con_lower - means[..., i]) / sigmas[..., i] dist_u = (con_upper - means[..., i]) / sigmas[..., i] log_prob = log_prob + log_prob_normal_in(a=dist_l, b=dist_u).sum(dim=-1) return log_prob