#!/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"""
Batch acquisition functions using the reparameterization trick in combination
with (quasi) Monte-Carlo sampling. See [Rezende2014reparam]_ and
[Wilson2017reparam]_
.. [Rezende2014reparam]
D. J. Rezende, S. Mohamed, and D. Wierstra. Stochastic backpropagation and
approximate inference in deep generative models. ICML 2014.
.. [Wilson2017reparam]
J. T. Wilson, R. Moriconi, F. Hutter, and M. P. Deisenroth.
The reparameterization trick for acquisition functions. ArXiv 2017.
"""
import math
from abc import ABC, abstractmethod
from typing import Optional, Union
import torch
from torch import Tensor
from ..exceptions.errors import UnsupportedError
from ..models.model import Model
from ..sampling.samplers import MCSampler, SobolQMCNormalSampler
from ..utils.transforms import (
concatenate_pending_points,
match_batch_shape,
t_batch_mode_transform,
)
from .acquisition import AcquisitionFunction
from .objective import IdentityMCObjective, MCAcquisitionObjective
from .utils import prune_inferior_points
[docs]class MCAcquisitionFunction(AcquisitionFunction, ABC):
r"""Abstract base class for Monte-Carlo based batch acquisition functions."""
def __init__(
self,
model: Model,
sampler: Optional[MCSampler] = None,
objective: Optional[MCAcquisitionObjective] = None,
X_pending: Optional[Tensor] = None,
) -> None:
r"""Constructor for the MCAcquisitionFunction base class.
Args:
model: A fitted model.
sampler: The sampler used to draw base samples. Defaults to
`SobolQMCNormalSampler(num_samples=512, collapse_batch_dims=True)`.
objective: The MCAcquisitionObjective under which the samples are
evaluated. Defaults to `IdentityMCObjective()`.
X_pending: A `m x d`-dim Tensor of `m` design points that have
points that have been submitted for function evaluation
but have not yet been evaluated.
"""
super().__init__(model=model)
if sampler is None:
sampler = SobolQMCNormalSampler(num_samples=512, collapse_batch_dims=True)
self.add_module("sampler", sampler)
if objective is None:
if model.num_outputs != 1:
raise UnsupportedError(
"Must specify an objective when using a multi-output model."
)
objective = IdentityMCObjective()
elif not isinstance(objective, MCAcquisitionObjective):
raise UnsupportedError(
"Only objectives of type MCAcquisitionObjective are supported for "
"MC acquisition functions."
)
self.add_module("objective", objective)
self.set_X_pending(X_pending)
[docs] @abstractmethod
def forward(self, X: Tensor) -> Tensor:
r"""Takes in a `(b) x q x d` X Tensor of `(b)` t-batches with `q` `d`-dim
design points each, and returns a one-dimensional Tensor with
`(b)` elements. Should utilize the result of set_X_pending as needed
to account for pending function evaluations.
"""
pass # pragma: no cover
[docs]class qExpectedImprovement(MCAcquisitionFunction):
r"""MC-based batch Expected Improvement.
This computes qEI by
(1) sampling the joint posterior over q points
(2) evaluating the improvement over the current best for each sample
(3) maximizing over q
(4) averaging over the samples
`qEI(X) = E(max(max Y - best_f, 0)), Y ~ f(X), where X = (x_1,...,x_q)`
Example:
>>> model = SingleTaskGP(train_X, train_Y)
>>> best_f = train_Y.max()[0]
>>> sampler = SobolQMCNormalSampler(1000)
>>> qEI = qExpectedImprovement(model, best_f, sampler)
>>> qei = qEI(test_X)
"""
def __init__(
self,
model: Model,
best_f: Union[float, Tensor],
sampler: Optional[MCSampler] = None,
objective: Optional[MCAcquisitionObjective] = None,
X_pending: Optional[Tensor] = None,
) -> None:
r"""q-Expected Improvement.
Args:
model: A fitted model.
best_f: The best objective value observed so far (assumed noiseless).
sampler: The sampler used to draw base samples. Defaults to
`SobolQMCNormalSampler(num_samples=500, collapse_batch_dims=True)`
objective: The MCAcquisitionObjective under which the samples are
evaluated. Defaults to `IdentityMCObjective()`.
X_pending: A `m x d`-dim Tensor of `m` design points that have
points that have been submitted for function evaluation
but have not yet been evaluated. Concatenated into X upon
forward call. Copied and set to have no gradient.
"""
super().__init__(
model=model, sampler=sampler, objective=objective, X_pending=X_pending
)
if not torch.is_tensor(best_f):
best_f = torch.tensor(float(best_f))
elif not (best_f.ndim < 2 and best_f.numel() == 1):
raise ValueError(
"best_f must be either a scalar or a single-elment "
"one-dimensional tensor."
)
self.register_buffer("best_f", best_f)
[docs] @concatenate_pending_points
@t_batch_mode_transform()
def forward(self, X: Tensor) -> Tensor:
r"""Evaluate qExpectedImprovement on the candidate set `X`.
Args:
X: A `(b) x q x d`-dim Tensor of `(b)` t-batches with `q` `d`-dim
design points each.
Returns:
A `(b)`-dim Tensor of Expected Improvement values at the given
design points `X`.
"""
posterior = self.model.posterior(X)
samples = self.sampler(posterior)
obj = self.objective(samples)
obj = (obj - self.best_f).clamp_min(0)
q_ei = obj.max(dim=-1)[0].mean(dim=0)
return q_ei
[docs]class qNoisyExpectedImprovement(MCAcquisitionFunction):
r"""MC-based batch Noisy Expected Improvement.
This function does not assume a `best_f` is known (which would require
noiseless observations). Instead, it uses samples from the joint posterior
over the `q` test points and previously observed points. The improvement
over previously observed points is computed for each sample and averaged.
`qNEI(X) = E(max(max Y - max Y_baseline, 0))`, where
`(Y, Y_baseline) ~ f((X, X_baseline)), X = (x_1,...,x_q)`
Example:
>>> model = SingleTaskGP(train_X, train_Y)
>>> sampler = SobolQMCNormalSampler(1000)
>>> qNEI = qNoisyExpectedImprovement(model, train_X, sampler)
>>> qnei = qNEI(test_X)
"""
def __init__(
self,
model: Model,
X_baseline: Tensor,
sampler: Optional[MCSampler] = None,
objective: Optional[MCAcquisitionObjective] = None,
X_pending: Optional[Tensor] = None,
prune_baseline: bool = False,
) -> None:
r"""q-Noisy Expected Improvement.
Args:
model: A fitted model.
X_baseline: A `r x d`-dim Tensor of `r` design points that have
already been observed. These points are considered as the
potential best design point.
sampler: The sampler used to draw base samples. Defaults to
`SobolQMCNormalSampler(num_samples=500, collapse_batch_dims=True)`.
objective: The MCAcquisitionObjective under which the samples are
evaluated. Defaults to `IdentityMCObjective()`.
X_pending: A `m x d`-dim Tensor of `m` design points that have
points that have been submitted for function evaluation
but have not yet been evaluated. Concatenated into X upon
forward call. Copied and set to have no gradient.
prune_baseline: If True, remove points in `X_baseline` that are
highly unlikely to be the best point. This can significantly
improve performance and is generally recommended. In order to
customize pruning parameters, instead manually call
`botorch.acquisition.utils.prune_inferior_points` on `X_baseline`
before instantiating the acquisition function.
"""
super().__init__(
model=model, sampler=sampler, objective=objective, X_pending=X_pending
)
if prune_baseline:
X_baseline = prune_inferior_points(
model=model, X=X_baseline, objective=objective
)
self.register_buffer("X_baseline", X_baseline)
[docs] @concatenate_pending_points
@t_batch_mode_transform()
def forward(self, X: Tensor) -> Tensor:
r"""Evaluate qNoisyExpectedImprovement on the candidate set `X`.
Args:
X: A `(b) x q x d`-dim Tensor of `(b)` t-batches with `q` `d`-dim
design points each.
Returns:
A `(b)`-dim Tensor of Noisy Expected Improvement values at the given
design points `X`.
"""
q = X.shape[-2]
X_full = torch.cat([X, match_batch_shape(self.X_baseline, X)], dim=-2)
# TODO (T41248036): Implement more efficient way to compute posterior
# over both training and test points in GPyTorch
posterior = self.model.posterior(X_full)
samples = self.sampler(posterior)
obj = self.objective(samples)
diffs = obj[:, :, :q].max(dim=-1)[0] - obj[:, :, q:].max(dim=-1)[0]
return diffs.clamp_min(0).mean(dim=0)
[docs]class qProbabilityOfImprovement(MCAcquisitionFunction):
r"""MC-based batch Probability of Improvement.
Estimates the probability of improvement over the current best observed
value by sampling from the joint posterior distribution of the q-batch.
MC-based estimates of a probability involves taking expectation of an
indicator function; to support auto-differntiation, the indicator is
replaced with a sigmoid function with temperature parameter `tau`.
`qPI(X) = P(max Y >= best_f), Y ~ f(X), X = (x_1,...,x_q)`
Example:
>>> model = SingleTaskGP(train_X, train_Y)
>>> best_f = train_Y.max()[0]
>>> sampler = SobolQMCNormalSampler(1000)
>>> qPI = qProbabilityOfImprovement(model, best_f, sampler)
>>> qpi = qPI(test_X)
"""
def __init__(
self,
model: Model,
best_f: Union[float, Tensor],
sampler: Optional[MCSampler] = None,
objective: Optional[MCAcquisitionObjective] = None,
X_pending: Optional[Tensor] = None,
tau: float = 1e-3,
) -> None:
r"""q-Probability of Improvement.
Args:
model: A fitted model.
best_f: The best objective value observed so far (assumed noiseless).
sampler: The sampler used to draw base samples. Defaults to
`SobolQMCNormalSampler(num_samples=500, collapse_batch_dims=True)`
objective: The MCAcquisitionObjective under which the samples are
evaluated. Defaults to `IdentityMCObjective()`.
X_pending: A `m x d`-dim Tensor of `m` design points that have
points that have been submitted for function evaluation
but have not yet been evaluated. Concatenated into X upon
forward call. Copied and set to have no gradient.
tau: The temperature parameter used in the sigmoid approximation
of the step function. Smaller values yield more accurate
approximations of the function, but result in gradients
estimates with higher variance.
"""
super().__init__(
model=model, sampler=sampler, objective=objective, X_pending=X_pending
)
if not torch.is_tensor(best_f):
best_f = torch.tensor(float(best_f))
self.register_buffer("best_f", best_f)
if not torch.is_tensor(tau):
tau = torch.tensor(float(tau))
self.register_buffer("tau", tau)
[docs] @concatenate_pending_points
@t_batch_mode_transform()
def forward(self, X: Tensor) -> Tensor:
r"""Evaluate qProbabilityOfImprovement on the candidate set `X`.
Args:
X: A `(b) x q x d`-dim Tensor of `(b)` t-batches with `q` `d`-dim
design points each.
Returns:
A `(b)`-dim Tensor of Probability of Improvement values at the given
design points `X`.
"""
posterior = self.model.posterior(X)
samples = self.sampler(posterior)
obj = self.objective(samples)
max_obj = obj.max(dim=-1)[0]
val = torch.sigmoid((max_obj - self.best_f) / self.tau).mean(dim=0)
return val
[docs]class qSimpleRegret(MCAcquisitionFunction):
r"""MC-based batch Simple Regret.
Samples from the joint posterior over the q-batch and computes the simple
regret.
`qSR(X) = E(max Y), Y ~ f(X), X = (x_1,...,x_q)`
Example:
>>> model = SingleTaskGP(train_X, train_Y)
>>> sampler = SobolQMCNormalSampler(1000)
>>> qSR = qSimpleRegret(model, sampler)
>>> qsr = qSR(test_X)
"""
[docs] @concatenate_pending_points
@t_batch_mode_transform()
def forward(self, X: Tensor) -> Tensor:
r"""Evaluate qSimpleRegret on the candidate set `X`.
Args:
X: A `(b) x q x d`-dim Tensor of `(b)` t-batches with `q` `d`-dim
design points each.
Returns:
A `(b)`-dim Tensor of Simple Regret values at the given design
points `X`.
"""
posterior = self.model.posterior(X)
samples = self.sampler(posterior)
obj = self.objective(samples)
val = obj.max(dim=-1)[0].mean(dim=0)
return val
[docs]class qUpperConfidenceBound(MCAcquisitionFunction):
r"""MC-based batch Upper Confidence Bound.
Uses a reparameterization to extend UCB to qUCB for q > 1 (See Appendix A
of [Wilson2017reparam].)
`qUCB = E(max(mu + |Y_tilde - mu|))`, where `Y_tilde ~ N(mu, beta pi/2 Sigma)`
and `f(X)` has distribution `N(mu, Sigma)`.
Example:
>>> model = SingleTaskGP(train_X, train_Y)
>>> sampler = SobolQMCNormalSampler(1000)
>>> qUCB = qUpperConfidenceBound(model, 0.1, sampler)
>>> qucb = qUCB(test_X)
"""
def __init__(
self,
model: Model,
beta: float,
sampler: Optional[MCSampler] = None,
objective: Optional[MCAcquisitionObjective] = None,
X_pending: Optional[Tensor] = None,
) -> None:
r"""q-Upper Confidence Bound.
Args:
model: A fitted model.
beta: Controls tradeoff between mean and standard deviation in UCB.
sampler: The sampler used to draw base samples. Defaults to
`SobolQMCNormalSampler(num_samples=500, collapse_batch_dims=True)`
objective: The MCAcquisitionObjective under which the samples are
evaluated. Defaults to `IdentityMCObjective()`.
X_pending: A `m x d`-dim Tensor of `m` design points that have
points that have been submitted for function evaluation
but have not yet been evaluated. Concatenated into X upon
forward call. Copied and set to have no gradient.
"""
super().__init__(
model=model, sampler=sampler, objective=objective, X_pending=X_pending
)
self.beta_prime = math.sqrt(beta * math.pi / 2)
[docs] @concatenate_pending_points
@t_batch_mode_transform()
def forward(self, X: Tensor) -> Tensor:
r"""Evaluate qUpperConfidenceBound on the candidate set `X`.
Args:
X: A `(b) x q x d`-dim Tensor of `(b)` t-batches with `q` `d`-dim
design points each.
Returns:
A `(b)`-dim Tensor of Upper Confidence Bound values at the given
design points `X`.
"""
posterior = self.model.posterior(X)
samples = self.sampler(posterior)
obj = self.objective(samples)
mean = obj.mean(dim=0)
ucb_samples = mean + self.beta_prime * (obj - mean).abs()
return ucb_samples.max(dim=-1)[0].mean(dim=0)
