#!/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"""
A general implementation of multi-step look-ahead acquistion function with configurable
value functions. See [Jiang2020multistep]_.
.. [Jiang2020multistep]
S. Jiang, D. R. Jiang, M. Balandat, B. Karrer, J. Gardner, and R. Garnett.
Efficient Nonmyopic Bayesian Optimization via One-Shot Multi-Step Trees.
In Advances in Neural Information Processing Systems 33, 2020.
"""
from __future__ import annotations
import math
import warnings
from typing import Any, Callable, Dict, List, Optional, Tuple, Type
import numpy as np
import torch
from botorch.acquisition import AcquisitionFunction, OneShotAcquisitionFunction
from botorch.acquisition.analytic import AnalyticAcquisitionFunction, PosteriorMean
from botorch.acquisition.monte_carlo import MCAcquisitionFunction
from botorch.acquisition.objective import (
AcquisitionObjective,
MCAcquisitionObjective,
ScalarizedObjective,
)
from botorch.exceptions.errors import UnsupportedError
from botorch.exceptions.warnings import BotorchWarning
from botorch.models.model import Model
from botorch.optim.initializers import initialize_q_batch
from botorch.sampling.samplers import MCSampler, SobolQMCNormalSampler
from botorch.utils.transforms import (
match_batch_shape,
t_batch_mode_transform,
unnormalize,
)
from torch import Size, Tensor
from torch.distributions import Beta
from torch.nn import ModuleList
TAcqfArgConstructor = Callable[[Model, Tensor], Dict[str, Any]]
[docs]class qMultiStepLookahead(MCAcquisitionFunction, OneShotAcquisitionFunction):
r"""MC-based batch Multi-Step Look-Ahead (one-shot optimization)."""
def __init__(
self,
model: Model,
batch_sizes: List[int],
num_fantasies: Optional[List[int]] = None,
samplers: Optional[List[MCSampler]] = None,
valfunc_cls: Optional[List[Optional[Type[AcquisitionFunction]]]] = None,
valfunc_argfacs: Optional[List[Optional[TAcqfArgConstructor]]] = None,
objective: Optional[AcquisitionObjective] = None,
inner_mc_samples: Optional[List[int]] = None,
X_pending: Optional[Tensor] = None,
collapse_fantasy_base_samples: bool = True,
) -> None:
r"""q-Multi-Step Look-Ahead (one-shot optimization).
Performs a `k`-step lookahead by means of repeated fantasizing.
Allows to specify the stage value functions by passing the respective class
objects via the `valfunc_cls` list. Optionally, `valfunc_argfacs` takes a list
of callables that generate additional kwargs for these constructors. By default,
`valfunc_cls` will be chosen as `[None, ..., None, PosteriorMean]`, which
corresponds to the (parallel) multi-step KnowledgeGradient. If, in addition,
`k=1` and `q_1 = 1`, this reduces to the classic Knowledge Gradient.
WARNING: The complexity of evaluating this function is exponential in the number
of lookahead steps!
Args:
model: A fitted model.
batch_sizes: A list `[q_1, ..., q_k]` containing the batch sizes for the
`k` look-ahead steps.
num_fantasies: A list `[f_1, ..., f_k]` containing the number of fantasy
points to use for the `k` look-ahead steps.
samplers: A list of MCSampler objects to be used for sampling fantasies in
each stage.
valfunc_cls: A list of `k + 1` acquisition function classes to be used as
the (stage + terminal) value functions. Each element (except for the
last one) can be `None`, in which case a zero stage value is assumed for
the respective stage. If `None`, this defaults to
`[None, ..., None, PosteriorMean]`
valfunc_argfacs: A list of `k + 1` "argument factories", i.e. callables that
map a `Model` and input tensor `X` to a dictionary of kwargs for the
respective stage value function constructor (e.g. `best_f` for
`ExpectedImprovement`). If None, only the standard (`model`, `sampler`
and `objective`) kwargs will be used.
objective: The objective under which the output is evaluated. If `None`, use
the model output (requires a single-output model). If a
`ScalarizedObjective` and `value_function_cls` is a subclass of
`AnalyticAcquisitonFunction`, then the analytic posterior mean is used.
Otherwise the objective is MC-evaluated (using `inner_sampler`).
inner_mc_samples: A list `[n_0, ..., n_k]` containing the number of MC
samples to be used for evaluating the stage value function. Ignored if
the objective is `None` or a `ScalarizedObjective`.
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.
collapse_fantasy_base_samples: If True, collapse_batch_dims of the Samplers
will be applied on fantasy batch dimensions as well, meaning that base
samples are the same in all subtrees starting from the same level.
"""
super(MCAcquisitionFunction, self).__init__(model=model)
self.batch_sizes = batch_sizes
if not ((num_fantasies is None) ^ (samplers is None)):
raise UnsupportedError(
"qMultiStepLookahead requires exactly one of `num_fantasies` or "
"`samplers` as arguments."
)
if samplers is None:
# The batch_range is not set here and left to sampler default of (0, -2),
# meaning that collapse_batch_dims will be applied on fantasy batch
# dimensions. If collapse_fantasy_base_samples is False, the batch_range
# is updated during the forward call.
samplers: List[MCSampler] = [
SobolQMCNormalSampler(
num_samples=nf, resample=False, collapse_batch_dims=True
)
for nf in num_fantasies
]
else:
num_fantasies = [sampler.sample_shape[0] for sampler in samplers]
self.num_fantasies = num_fantasies
# By default do not use stage values and use PosteriorMean as terminal value
# function (= multi-step KG)
if valfunc_cls is None:
valfunc_cls = [None for _ in num_fantasies] + [PosteriorMean]
if inner_mc_samples is None:
inner_mc_samples = [None] * (1 + len(num_fantasies))
# TODO: Allow passing in inner samplers directly
inner_samplers = _construct_inner_samplers(
batch_sizes=batch_sizes,
valfunc_cls=valfunc_cls,
objective=objective,
inner_mc_samples=inner_mc_samples,
)
if valfunc_argfacs is None:
valfunc_argfacs = [None] * (1 + len(batch_sizes))
self.objective = objective
self.set_X_pending(X_pending)
self.samplers = ModuleList(samplers)
self.inner_samplers = ModuleList(inner_samplers)
self._valfunc_cls = valfunc_cls
self._valfunc_argfacs = valfunc_argfacs
self._collapse_fantasy_base_samples = collapse_fantasy_base_samples
[docs] @t_batch_mode_transform()
def forward(self, X: Tensor) -> Tensor:
r"""Evaluate qMultiStepLookahead on the candidate set X.
Args:
X: A `batch_shape x q' x d`-dim Tensor with `q'` design points for each
batch, where `q' = q_0 + f_1 q_1 + f_2 f_1 q_2 + ...`. Here `q_i`
is the number of candidates jointly considered in look-ahead step
`i`, and `f_i` is respective number of fantasies.
Returns:
The acquisition value for each batch as a tensor of shape `batch_shape`.
"""
Xs = self.get_multi_step_tree_input_representation(X)
# set batch_range on samplers if not collapsing on fantasy dims
if not self._collapse_fantasy_base_samples:
self._set_samplers_batch_range(batch_shape=X.shape[:-2])
return _step(
model=self.model,
Xs=Xs,
samplers=self.samplers,
valfunc_cls=self._valfunc_cls,
valfunc_argfacs=self._valfunc_argfacs,
inner_samplers=self.inner_samplers,
objective=self.objective,
running_val=None,
)
@property
def _num_auxiliary(self) -> int:
r"""Number of auxiliary variables in the q-batch dimension.
Returns:
`q_aux` s.t. `q + q_aux = augmented_q_batch_size`
"""
return np.dot(self.batch_sizes, np.cumprod(self.num_fantasies)).item()
def _set_samplers_batch_range(self, batch_shape: Size) -> None:
r"""Set batch_range on samplers.
Args:
batch_shape:
"""
tbatch_dim_start = -2 - len(batch_shape)
for s in self.samplers:
s.batch_range = (tbatch_dim_start, -2)
[docs] def get_augmented_q_batch_size(self, q: int) -> int:
r"""Get augmented q batch size for one-shot optimzation.
Args:
q: The number of candidates to consider jointly.
Returns:
The augmented size for one-shot optimzation (including variables
parameterizing the fantasy solutions): `q_0 + f_1 q_1 + f_2 f_1 q_2 + ...`
"""
return q + self._num_auxiliary
[docs] def get_split_shapes(self, X: Tensor) -> Tuple[Size, List[Size], List[int]]:
r"""Get the split shapes from X.
Args:
X: A `batch_shape x q_aug x d`-dim tensor including fantasy points.
Returns:
A 3-tuple `(batch_shape, shapes, sizes)`, where
`shape[i] = f_i x .... x f_1 x batch_shape x q_i x d` and
`size[i] = f_i * ... f_1 * q_i`.
"""
batch_shape, (q_aug, d) = X.shape[:-2], X.shape[-2:]
q = q_aug - self._num_auxiliary
batch_sizes = [q] + self.batch_sizes
# X_i needs to have shape f_i x .... x f_1 x batch_shape x q_i x d
shapes = [
torch.Size(self.num_fantasies[:i][::-1] + [*batch_shape, q_i, d])
for i, q_i in enumerate(batch_sizes)
]
# Each X_i in the split X has shape batch_shape x qtilde x d with
# qtilde = f_i * ... * f_1 * q_i
sizes = [s[: (-2 - len(batch_shape))].numel() * s[-2] for s in shapes]
return batch_shape, shapes, sizes
[docs] def extract_candidates(self, X_full: Tensor) -> Tensor:
r"""We only return X as the set of candidates post-optimization.
Args:
X_full: A `batch_shape x q' x d`-dim Tensor with `q'` design points for
each batch, where `q' = q + f_1 q_1 + f_2 f_1 q_2 + ...`.
Returns:
A `batch_shape x q x d`-dim Tensor with `q` design points for each batch.
"""
return X_full[..., : -self._num_auxiliary, :]
[docs] def get_induced_fantasy_model(self, X: Tensor) -> Model:
r"""Fantasy model induced by X.
Args:
X: A `batch_shape x q' x d`-dim Tensor with `q'` design points for each
batch, where `q' = q_0 + f_1 q_1 + f_2 f_1 q_2 + ...`. Here `q_i`
is the number of candidates jointly considered in look-ahead step
`i`, and `f_i` is respective number of fantasies.
Returns:
The fantasy model induced by X.
"""
Xs = self.get_multi_step_tree_input_representation(X)
# set batch_range on samplers if not collapsing on fantasy dims
if not self._collapse_fantasy_base_samples:
self._set_samplers_batch_range(batch_shape=X.shape[:-2])
return _get_induced_fantasy_model(
model=self.model, Xs=Xs, samplers=self.samplers
)
def _step(
model: Model,
Xs: List[Tensor],
samplers: List[Optional[MCSampler]],
valfunc_cls: List[Optional[Type[AcquisitionFunction]]],
valfunc_argfacs: List[Optional[TAcqfArgConstructor]],
inner_samplers: List[Optional[MCSampler]],
objective: AcquisitionObjective,
running_val: Optional[Tensor] = None,
sample_weights: Optional[Tensor] = None,
step_index: int = 0,
) -> Tensor:
r"""Recursive multi-step look-ahead computation.
Helper function computing the "value-to-go" of a multi-step lookahead scheme.
Args:
model: A Model of appropriate batch size. Specifically, it must be possible to
evaluate the model's posterior at `Xs[0]`.
Xs: A list `[X_j, ..., X_k]` of tensors, where `X_i` has shape
`f_i x .... x f_1 x batch_shape x q_i x d`.
samplers: A list of `k - j` samplers, such that the number of samples of sampler
`i` is `f_i`. The last element of this list is considered the
"inner sampler", which is used for evaluating the objective in case it is an
MCAcquisitionObjective.
valfunc_cls: A list of acquisition function class to be used as the (stage +
terminal) value functions. Each element (except for the last one) can be
`None`, in which case a zero stage value is assumed for the respective
stage.
valfunc_argfacs: A list of callables that map a `Model` and input tensor `X` to
a dictionary of kwargs for the respective stage value function constructor.
If `None`, only the standard `model`, `sampler` and `objective` kwargs will
be used.
inner_samplers: A list of `MCSampler` objects, each to be used in the stage
value function at the corresponding index.
objective: The AcquisitionObjective under which the model output is evaluated.
running_val: As `batch_shape`-dim tensor containing the current running value.
sample_weights: A tensor of shape `f_i x .... x f_1 x batch_shape` when called
in the `i`-th step by which to weight the stage value samples. Used in
conjunction with Gauss-Hermite integration or importance sampling. Assumed
to be `None` in the initial step (when `step_index=0`).
step_index: The index of the look-ahead step. `step_index=0` indicates the
initial step.
Returns:
A `b`-dim tensor containing the multi-step value of the design `X`.
"""
X = Xs[0]
if sample_weights is None: # only happens in the initial step
sample_weights = torch.ones(*X.shape[:-2], device=X.device, dtype=X.dtype)
# compute stage value
stage_val = _compute_stage_value(
model=model,
valfunc_cls=valfunc_cls[0],
X=X,
objective=objective,
inner_sampler=inner_samplers[0],
arg_fac=valfunc_argfacs[0],
)
if stage_val is not None: # update running value
# if not None, running_val has shape f_{i-1} x ... x f_1 x batch_shape
# stage_val has shape f_i x ... x f_1 x batch_shape
# this sum will add a dimension to running_val so that
# updated running_val has shape f_i x ... x f_1 x batch_shape
running_val = stage_val if running_val is None else running_val + stage_val
# base case: no more fantasizing, return value
if len(Xs) == 1:
# compute weighted average over all leaf nodes of the tree
batch_shape = running_val.shape[step_index:]
# expand sample weights to make sure it is the same shape as running_val,
# because we need to take a sum over sample weights for computing the
# weighted average
sample_weights = sample_weights.expand(running_val.shape)
return (running_val * sample_weights).view(-1, *batch_shape).sum(dim=0)
# construct fantasy model (with batch shape f_{j+1} x ... x f_1 x batch_shape)
prop_grads = step_index > 0 # need to propagate gradients for steps > 0
fantasy_model = model.fantasize(
X=X, sampler=samplers[0], observation_noise=True, propagate_grads=prop_grads
)
# augment sample weights appropriately
sample_weights = _construct_sample_weights(
prev_weights=sample_weights, sampler=samplers[0]
)
return _step(
model=fantasy_model,
Xs=Xs[1:],
samplers=samplers[1:],
valfunc_cls=valfunc_cls[1:],
valfunc_argfacs=valfunc_argfacs[1:],
inner_samplers=inner_samplers[1:],
objective=objective,
sample_weights=sample_weights,
running_val=running_val,
step_index=step_index + 1,
)
def _compute_stage_value(
model: Model,
valfunc_cls: Optional[Type[AcquisitionFunction]],
X: Tensor,
objective: AcquisitionObjective,
inner_sampler: Optional[MCSampler] = None,
arg_fac: Optional[TAcqfArgConstructor] = None,
) -> Optional[Tensor]:
r"""Compute the stage value of a multi-step look-ahead policy.
Args:
model: A Model of appropriate batch size. Specifically, it must be possible to
evaluate the model's posterior at `Xs[0]`.
valfunc_cls: The acquisition function class to be used as the stage value
functions. If `None`, a zero stage value is assumed (returns `None`)
X: A tensor with shape `f_i x .... x f_1 x batch_shape x q_i x d` when called in
the `i`-th step.
objective: The AcquisitionObjective under which the model output is evaluated.
inner_sampler: An `MCSampler` object to be used in the stage value function. Can
be `None` for analytic acquisition functions or when using the default
sampler of the acquisition function class.
arg_fac: A callable mapping a `Model` and the input tensor `X` to a dictionary
of kwargs for the stage value function constructor. If `None`, only the
standard `model`, `sampler` and `objective` kwargs will be used.
Returns:
A `f_i x ... x f_1 x batch_shape`-dim tensor of stage values, or `None`
(= zero stage value).
"""
if valfunc_cls is None:
return None
common_kwargs: Dict[str, Any] = {"model": model, "objective": objective}
if issubclass(valfunc_cls, MCAcquisitionFunction):
common_kwargs["sampler"] = inner_sampler
kwargs = arg_fac(model=model, X=X) if arg_fac is not None else {}
stage_val_func = valfunc_cls(**common_kwargs, **kwargs)
# shape of stage_val is f_i x ... x f_1 x batch_shape
stage_val = stage_val_func(X=X)
return stage_val
def _construct_sample_weights(
prev_weights: Tensor, sampler: MCSampler
) -> Optional[Tensor]:
r"""Iteratively construct tensor of sample weights for multi-step look-ahead.
Args:
prev_weights: A `f_i x .... x f_1 x batch_shape` tensor of previous sample
weights.
sampler: A `MCSampler` that may have sample weights as the `base_weights`
attribute. If the sampler does not have a `base_weights` attribute,
samples are weighted uniformly.
Returns:
A `f_{i+1} x .... x f_1 x batch_shape` tensor of sample weights for the next
step.
"""
new_weights = getattr(sampler, "base_weights", None) # TODO: generalize this
if new_weights is None:
# uniform weights
nf = sampler.sample_shape[0]
new_weights = torch.ones(
nf, device=prev_weights.device, dtype=prev_weights.dtype
)
# reshape new_weights to be f_{i+1} x 1 x ... x 1
new_weights = new_weights.view(-1, *(1 for _ in prev_weights.shape))
# normalize new_weights to sum to 1.0
new_weights = new_weights / new_weights.sum()
return new_weights * prev_weights
def _construct_inner_samplers(
batch_sizes: List[int],
valfunc_cls: List[Optional[Type[AcquisitionFunction]]],
inner_mc_samples: List[Optional[int]],
objective: Optional[AcquisitionObjective] = None,
) -> List[Optional[MCSampler]]:
r"""Check validity of inputs and construct inner samplers.
Helper function to be used internally for constructing inner samplers.
Args:
batch_sizes: A list `[q_1, ..., q_k]` containing the batch sizes for the
`k` look-ahead steps.
valfunc_cls: A list of `k + 1` acquisition function classes to be used as the
(stage + terminal) value functions. Each element (except for the last one)
can be `None`, in which case a zero stage value is assumed for the
respective stage.
inner_mc_samples: A list `[n_0, ..., n_k]` containing the number of MC
samples to be used for evaluating the stage value function. Ignored if
the objective is `None` or a `ScalarizedObjective`.
objective: The objective under which the output is evaluated. If `None`, use
the model output (requires a single-output model). If a
`ScalarizedObjective` and `value_function_cls` is a subclass of
`AnalyticAcquisitonFunction`, then the analytic posterior mean is used.
Otherwise the objective is MC-evaluated (using `inner_sampler`).
Returns:
A list with `k + 1` elements that are either `MCSampler`s or `None.
"""
inner_samplers = []
for q, vfc, mcs in zip([None] + batch_sizes, valfunc_cls, inner_mc_samples):
if vfc is None:
inner_samplers.append(None)
elif vfc == qMultiStepLookahead:
raise UnsupportedError(
"qMultiStepLookahead not supported as a value function "
"(I see what you did there, nice try...)."
)
elif issubclass(vfc, AnalyticAcquisitionFunction):
if objective is not None and not isinstance(objective, ScalarizedObjective):
raise UnsupportedError(
"Only objectives of type ScalarizedObjective are supported "
"for analytic value functions."
)
# At this point, we don't know the initial q-batch size here
if q is not None and q > 1:
raise UnsupportedError(
"Only batch sizes of q=1 are supported for analytic value "
"functions."
)
if q is not None and mcs is not None:
warnings.warn(
"inner_mc_samples is ignored for analytic acquistion functions",
BotorchWarning,
)
inner_samplers.append(None)
else:
if objective is not None and not isinstance(
objective, MCAcquisitionObjective
):
raise UnsupportedError(
"Only objectives of type MCAcquisitionObjective are supported "
"for MC value functions."
)
inner_sampler = SobolQMCNormalSampler(
num_samples=32 if mcs is None else mcs,
resample=False,
collapse_batch_dims=True,
)
inner_samplers.append(inner_sampler)
return inner_samplers
def _get_induced_fantasy_model(
model: Model, Xs: List[Tensor], samplers: List[Optional[MCSampler]]
) -> Model:
r"""Recursive computation of the fantasy model induced by an input tree.
Args:
model: A Model of appropriate batch size. Specifically, it must be possible to
evaluate the model's posterior at `Xs[0]`.
Xs: A list `[X_j, ..., X_k]` of tensors, where `X_i` has shape
`f_i x .... x f_1 x batch_shape x q_i x d`.
samplers: A list of `k - j` samplers, such that the number of samples of sampler
`i` is `f_i`. The last element of this list is considered the
"inner sampler", which is used for evaluating the objective in case it is an
MCAcquisitionObjective.
Returns:
A Model obtained by iteratively fantasizing over the input tree `Xs`.
"""
if len(Xs) == 1:
return model
else:
fantasy_model = model.fantasize(
X=Xs[0],
sampler=samplers[0],
observation_noise=True,
)
return _get_induced_fantasy_model(
model=fantasy_model, Xs=Xs[1:], samplers=samplers[1:]
)
[docs]def warmstart_multistep(
acq_function: qMultiStepLookahead,
bounds: Tensor,
num_restarts: int,
raw_samples: int,
full_optimizer: Tensor,
**kwargs: Any,
) -> Tensor:
r"""Warm-start initialization for multi-step look-ahead acquisition functions.
For now uses the same q' as in `full_optimizer`. TODO: allow different `q`.
Args:
acq_function: A qMultiStepLookahead acquisition function.
bounds: A `2 x d` tensor of lower and upper bounds for each column of features.
num_restarts: The number of starting points for multistart acquisition
function optimization.
raw_samples: The number of raw samples to consider in the initialization
heuristic.
full_optimizer: The full tree of optimizers of the previous iteration of shape
`batch_shape x q' x d`. Typically obtained by passing
`return_best_only=False` and `return_full_tree=True` into `optimize_acqf`.
kwargs: Optimization kwargs.
Returns:
A `num_restarts x q' x d` tensor for initial points for optimization.
This is a very simple initialization heuristic.
TODO: Use the observed values to identify the fantasy sub-tree that is closest to
the observed value.
"""
batch_shape, shapes, sizes = acq_function.get_split_shapes(full_optimizer)
Xopts = torch.split(full_optimizer, sizes, dim=-2)
tkwargs = {"device": Xopts[0].device, "dtype": Xopts[0].dtype}
B = Beta(torch.ones(1, **tkwargs), 3 * torch.ones(1, **tkwargs))
def mixin_layer(X: Tensor, bounds: Tensor, eta: float) -> Tensor:
perturbations = unnormalize(B.sample(X.shape).squeeze(-1), bounds)
return (1 - eta) * X + eta * perturbations
def make_init_tree(Xopts: List[Tensor], bounds: Tensor, etas: Tensor) -> Tensor:
Xtrs = [mixin_layer(X=X, bounds=bounds, eta=eta) for eta, X in zip(etas, Xopts)]
return torch.cat(Xtrs, dim=-2)
def mixin_tree(T: Tensor, bounds: Tensor, alpha: float) -> Tensor:
return (1 - alpha) * T + alpha * unnormalize(torch.rand_like(T), bounds)
n_repeat = math.ceil(raw_samples / batch_shape[0])
alphas = torch.linspace(0, 0.75, n_repeat, **tkwargs)
etas = torch.linspace(0.1, 1.0, len(Xopts), **tkwargs)
X_full = torch.cat(
[
mixin_tree(
T=make_init_tree(Xopts=Xopts, bounds=bounds, etas=etas),
bounds=bounds,
alpha=alpha,
)
for alpha in alphas
],
dim=0,
)
with torch.no_grad():
Y_full = acq_function(X_full)
X_init = initialize_q_batch(X=X_full, Y=Y_full, n=num_restarts, eta=1.0)
return X_init[:raw_samples]
[docs]def make_best_f(model: Model, X: Tensor) -> Dict[str, Any]:
r"""Extract the best observed training input from the model."""
return {"best_f": model.train_targets.max(dim=-1).values}