import typing
from collections.abc import Callable, Iterator, Sequence
import numpy as np
from stk._internal.ea.molecule_record import MoleculeRecord
from stk._internal.ea.selection.batch import Batch
from stk._internal.ea.selection.selectors.yielded_batches import YieldedBatches
from stk._internal.key_makers.inchi import Inchi
from stk._internal.key_makers.molecule import MoleculeKeyMaker
from .selector import Selector
T = typing.TypeVar("T", bound=MoleculeRecord)
[docs]
class StochasticUniversalSampling(Selector[T]):
"""
Yields batches of molecules through stochastic universal sampling.
Stochastic universal sampling lays out batches along a line, with
each batch taking up length proportional to its fitness. It
then creates a set of evenly spaced pointers to different points
on the line, each of which is occupied by a batch. Batches which
are pointed to are yielded.
This approach means weaker members of the population
are given a greater chance to be chosen than in
:class:`.Roulette` selection [#]_.
References:
.. [#] https://en.wikipedia.org/wiki/Stochastic_universal_sampling
Examples:
*Yielding Single Molecule Batches*
Yielding molecules one at a time. For example, if molecules need
to be selected for mutation or the next generation.
.. testcode:: yielding-single-molecule-batches
import stk
# Make the selector.
stochastic_sampling = stk.StochasticUniversalSampling(5)
population = {
stk.MoleculeRecord(
topology_graph=stk.polymer.Linear(
building_blocks=[
stk.BuildingBlock('BrCCBr', stk.BromoFactory()),
],
repeating_unit='A',
num_repeating_units=2,
),
): i
for i in range(100)
}
# Select the molecules.
for selected, in stochastic_sampling.select(population):
# Do stuff with each selected molecule.
pass
"""
def __init__(
self,
num_batches: int | None = None,
batch_size: int = 1,
duplicate_molecules: bool = True,
duplicate_batches: bool = True,
key_maker: MoleculeKeyMaker = Inchi(),
fitness_modifier: Callable[
[dict[T, float]], dict[T, float]
] = lambda x: x,
random_seed: int | np.random.Generator | None = None,
) -> None:
"""
Parameters:
num_batches:
The number of batches to yield. If ``None`` then yielding
will continue forever or until the generator is exhausted,
whichever comes first.
batch_size:
The number of molecules yielded at once.
duplicate_molecules:
If ``True`` the same molecule can be yielded in more than
one batch.
duplicate_batches:
If ``True`` the same batch can be yielded more than once.
key_maker:
Used to get the keys of molecules. If two molecules have
the same key, they are considered duplicates.
fitness_modifier:
Takes the `population` on which :meth:`~.Selector.select`
is called and returns a :class:`dict`, which maps records
in the `population` to the fitness values the
:class:`.Selector` should use.
random_seed:
The random seed to use.
"""
super().__init__(key_maker, fitness_modifier, batch_size)
if random_seed is None or isinstance(random_seed, int):
random_seed = np.random.default_rng(random_seed)
self._generator = random_seed
self._duplicate_molecules = duplicate_molecules
self._duplicate_batches = duplicate_batches
self._num_batches = (
float("inf") if num_batches is None else num_batches
)
def _select_from_batches(
self,
batches: Sequence[Batch[T]],
yielded_batches: YieldedBatches[T],
) -> Iterator[Batch[T]]:
batches = sorted(batches, reverse=True)
# SUS may need to run multiple rounds if duplicate_molecules or
# duplicate_batches is False. This is because in each round
# you can generate multiple pointers to the same batch or to
# batches sharing molecules. If this happens the lower fitness
# batch will not be yielded. Instead a second round of SUS will
# occur with any ineligible batches removed and a reduced
# number of pointers, to account for batches yielded in the
# previous rounds. This will repeat until the desired number
# of batches has been yielded, or there are no more valid
# batches.
while batches and yielded_batches.get_num() < self._num_batches:
yield from self._select_with_stochastic_universal_sampling(
batches=batches,
yielded_batches=yielded_batches,
)
if yielded_batches.get_num() < self._num_batches:
if not self._duplicate_molecules:
batches_ = filter(
yielded_batches.has_no_yielded_molecules,
batches,
)
if not self._duplicate_batches:
batches_ = filter(
yielded_batches.is_unyielded_batch,
batches,
)
if (
not self._duplicate_molecules
or not self._duplicate_batches
):
batches = tuple(batches_)
def _select_with_stochastic_universal_sampling(
self,
batches: Sequence[Batch[T]],
yielded_batches: YieldedBatches[T],
) -> Iterator[Batch[T]]:
total = sum(batch.get_fitness_value() for batch in batches)
batch_positions = []
batch_position = 0.0
for batch in batches:
batch_position += batch.get_fitness_value() / total
batch_positions.append(batch_position)
num_batches = typing.cast(
int,
min(self._num_batches - yielded_batches.get_num(), len(batches)),
)
pointer_distance = 1 / num_batches
pointers = []
pointer = self._generator.uniform(0, pointer_distance)
for _ in range(num_batches):
pointers.append(pointer)
pointer += pointer_distance
batch_index = 0
for pointer in pointers:
while pointer > batch_positions[batch_index]:
batch_index += 1
batch = batches[batch_index]
if (
not self._duplicate_molecules
and yielded_batches.has_yielded_molecules(batch)
):
continue
if (
not self._duplicate_batches
and yielded_batches.is_yielded_batch(batch)
):
continue
yield batch