Source code for botorch.utils.multi_objective.box_decompositions.box_decomposition

#!/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"""Box decomposition algorithms.

References

.. [Lacour17]
    R. Lacour, K. Klamroth, C. Fonseca. A box decomposition algorithm to
    compute the hypervolume indicator. Computers & Operations Research,
    Volume 79, 2017.

"""

from __future__ import annotations

from abc import ABC, abstractmethod
from typing import Optional

import torch
from botorch.exceptions.errors import BotorchError
from botorch.utils.multi_objective.box_decompositions.utils import (
    _expand_ref_point,
    _pad_batch_pareto_frontier,
    update_local_upper_bounds_incremental,
)
from botorch.utils.multi_objective.pareto import is_non_dominated
from torch import Tensor
from torch.nn import Module


[docs]class BoxDecomposition(Module, ABC): r"""An abstract class for box decompositions. Note: Internally, we store the negative reference point (minimization). """ def __init__( self, ref_point: Tensor, sort: bool, Y: Optional[Tensor] = None ) -> None: """Initialize BoxDecomposition. Args: ref_point: A `m`-dim tensor containing the reference point. sort: A boolean indicating whether to sort the Pareto frontier. Y: A `(batch_shape) x n x m`-dim tensor of outcomes. """ super().__init__() self.register_buffer("_neg_ref_point", -ref_point) self.register_buffer("sort", torch.tensor(sort, dtype=torch.bool)) self.num_outcomes = ref_point.shape[-1] if Y is not None: self._update_neg_Y(Y=Y) self.reset() @property def pareto_Y(self) -> Tensor: r"""This returns the non-dominated set. Returns: A `n_pareto x m`-dim tensor of outcomes. """ try: return -self._neg_pareto_Y except AttributeError: raise BotorchError("pareto_Y has not been initialized") @property def ref_point(self) -> Tensor: r"""Get the reference point. Returns: A `m`-dim tensor of outcomes. """ return -self._neg_ref_point @property def Y(self) -> Tensor: r"""Get the raw outcomes. Returns: A `n x m`-dim tensor of outcomes. """ return -self._neg_Y def _reset_pareto_Y(self) -> bool: r"""Update the non-dominated front. Returns: A boolean indicating whether the Pareto frontier has changed. """ # is_non_dominated assumes maximization if self._neg_Y.shape[-2] == 0: pareto_Y = self._neg_Y else: # assumes maximization pareto_Y = -_pad_batch_pareto_frontier( Y=self.Y, ref_point=_expand_ref_point( ref_point=self.ref_point, batch_shape=self.batch_shape ), ) if self.sort: # sort by first objective if len(self.batch_shape) > 0: pareto_Y = pareto_Y.gather( index=torch.argsort(pareto_Y[..., :1], dim=-2).expand( pareto_Y.shape ), dim=-2, ) else: pareto_Y = pareto_Y[torch.argsort(pareto_Y[:, 0])] if not hasattr(self, "_neg_pareto_Y") or not torch.equal( pareto_Y, self._neg_pareto_Y ): self.register_buffer("_neg_pareto_Y", pareto_Y) return True return False
[docs] def partition_space(self) -> None: r"""Compute box decomposition.""" if self.num_outcomes == 2: try: self._partition_space_2d() except NotImplementedError: self._partition_space() else: self._partition_space()
def _partition_space_2d(self) -> None: r"""Compute box decomposition for 2 objectives.""" raise NotImplementedError @abstractmethod def _partition_space(self): r"""Partition the non-dominated space into disjoint hypercells. This method supports an arbitrary number of outcomes, but is less efficient than `partition_space_2d` for the 2-outcome case. """ pass # pragma: no cover
[docs] @abstractmethod def get_hypercell_bounds(self) -> Tensor: r"""Get the bounds of each hypercell in the decomposition. Returns: A `2 x num_cells x num_outcomes`-dim tensor containing the lower and upper vertices bounding each hypercell. """ pass # pragma: no cover
def _update_neg_Y(self, Y: Tensor) -> bool: r"""Update the set of outcomes. Returns: A boolean indicating if _neg_Y was initialized. """ # multiply by -1, since internally we minimize. try: self._neg_Y = torch.cat([self._neg_Y, -Y], dim=-2) return False except AttributeError: self.register_buffer("_neg_Y", -Y) return True
[docs] def update(self, Y: Tensor) -> None: r"""Update non-dominated front and decomposition. By default, the partitioning is recomputed. Subclasses can override this functionality. Args: Y: A `(batch_shape) x n x m`-dim tensor of new, incremental outcomes. """ self._update_neg_Y(Y=Y) self.reset()
[docs] def reset(self) -> None: r"""Reset non-dominated front and decomposition.""" self.batch_shape = self.Y.shape[:-2] self.num_outcomes = self.Y.shape[-1] if len(self.batch_shape) > 1: raise NotImplementedError( f"{type(self).__name__} only supports a single " f"batch dimension, but got {len(self.batch_shape)} " "batch dimensions." ) elif len(self.batch_shape) > 0 and self.num_outcomes > 2: raise NotImplementedError( f"{type(self).__name__} only supports a batched box " f"decompositions in the 2-objective setting." ) is_new_pareto = self._reset_pareto_Y() # Update decomposition if the Pareto front changed if is_new_pareto: self.partition_space()
[docs] @abstractmethod def compute_hypervolume(self) -> Tensor: r"""Compute hypervolume that is dominated by the Pareto Froniter. Returns: A `(batch_shape)`-dim tensor containing the hypervolume dominated by each Pareto frontier. """ pass # pragma: no cover
[docs]class FastPartitioning(BoxDecomposition, ABC): r"""A class for partitioning the (non-)dominated space into hyper-cells. Note: this assumes maximization. Internally, it multiplies outcomes by -1 and performs the decomposition under minimization. This class is abstract to support to two applications of Alg 1 from [Lacour17]_: 1) partitioning the space that is dominated by the Pareto frontier and 2) partitioning the space that is not dominated by the Pareto frontier. """ def __init__( self, ref_point: Tensor, Y: Optional[Tensor] = None, ) -> None: """Initialize FastPartitioning. Args: ref_point: A `m`-dim tensor containing the reference point. Y: A `(batch_shape) x n x m`-dim tensor """ super().__init__(ref_point=ref_point, Y=Y, sort=ref_point.shape[-1] == 2)
[docs] def update(self, Y: Tensor) -> None: r"""Update non-dominated front and decomposition. Args: Y: A `(batch_shape) x n x m`-dim tensor of new, incremental outcomes. """ if self._update_neg_Y(Y=Y): self.reset() else: if self.num_outcomes == 2 or self._neg_pareto_Y.shape[-2] == 0: # If there are two objective, recompute the box decomposition # because the partitions can be computed analytically. # If the current pareto set has no points, recompute the box # decomposition. self.reset() else: # only include points that are better than the reference point better_than_ref = (Y > self.ref_point).all(dim=-1) Y = Y[better_than_ref] Y_all = torch.cat([self._neg_pareto_Y, -Y], dim=-2) pareto_mask = is_non_dominated(-Y_all) # determine the number of points in Y that are Pareto optimal num_new_pareto = pareto_mask[-Y.shape[-2] :].sum() self._neg_pareto_Y = Y_all[pareto_mask] if num_new_pareto > 0: # update local upper bounds for the minimization problem self._U, self._Z = update_local_upper_bounds_incremental( # this assumes minimization new_pareto_Y=self._neg_pareto_Y[-num_new_pareto:], U=self._U, Z=self._Z, ) # use the negative local upper bounds as the new pareto # frontier for the minimization problem and perform # box decomposition on dominated space. self._get_partitioning()
@abstractmethod def _get_single_cell(self) -> None: r"""Set the partitioning to be a single cell in the case of no Pareto points. This method should set self.hypercell_bounds """ pass # pragma: no cover
[docs] def partition_space(self) -> None: if self._neg_pareto_Y.shape[-2] == 0: self._get_single_cell() else: super().partition_space()
def _partition_space(self): r"""Partition the non-dominated space into disjoint hypercells. This method supports an arbitrary number of outcomes, but is less efficient than `partition_space_2d` for the 2-outcome case. """ if len(self.batch_shape) > 0: # this could be triggered when m=2 outcomes and # BoxDecomposition._partition_space_2d is not overridden. raise NotImplementedError( "_partition_space does not support batch dimensions." ) # this assumes minimization # initialize local upper bounds self.register_buffer("_U", self._neg_ref_point.unsqueeze(-2).clone()) # initialize defining points to be the dummy points \hat{z} that are # defined in Sec 2.1 in [Lacour17]_. Note that in [Lacour17]_, outcomes # are assumed to be between [0,1], so they used 0 rather than -inf. self._Z = torch.zeros( 1, self.num_outcomes, self.num_outcomes, dtype=self.Y.dtype, device=self.Y.device, ) for j in range(self.ref_point.shape[-1]): # use ref point for maximization as the ideal point for minimization. self._Z[0, j] = float("-inf") self._Z[0, j, j] = self._U[0, j] # incrementally update local upper bounds and defining points # for each new Pareto point self._U, self._Z = update_local_upper_bounds_incremental( new_pareto_Y=self._neg_pareto_Y, U=self._U, Z=self._Z, ) self._get_partitioning() @abstractmethod def _get_partitioning(self) -> None: r"""Compute partitioning given local upper bounds for the minimization problem. This method should set self.hypercell_bounds """ pass # pragma: no cover
[docs] def get_hypercell_bounds(self) -> Tensor: r"""Get the bounds of each hypercell in the decomposition. Returns: A `2 x (batch_shape) x num_cells x m`-dim tensor containing the lower and upper vertices bounding each hypercell. """ return self.hypercell_bounds