"""Contains the SlidingBoundariesArchive."""
from collections import deque
import numpy as np
from sortedcontainers import SortedList
from ribs._utils import (check_batch_shape, check_finite, validate_batch,
validate_single)
from ribs.archives._archive_base import ArchiveBase
from ribs.archives._grid_archive import GridArchive
class SolutionBuffer:
"""An internal class that stores relevant data to re-add after remapping.
Maintains two data structures:
- Queue storing the buffer_capacity last entries (solution + objective +
measures). When new items are inserted, the oldest ones are
popped.
- Sorted lists with the sorted measures in each dimension. Measures are
removed from these lists when they are removed from the queue.
"""
def __init__(self, buffer_capacity, measure_dim):
self._buffer_capacity = buffer_capacity
self._queue = deque()
self._measure_lists = [SortedList() for _ in range(measure_dim)]
self._iter_idx = 0
def __iter__(self):
"""Enables iterating over solutions stored in the buffer."""
return self
def __next__(self):
"""Returns the next entry in the buffer."""
if self._iter_idx >= self.size:
self._iter_idx = 0
raise StopIteration
result = self._queue[self._iter_idx]
self._iter_idx += 1
return result
def add(self, data):
"""Inserts a new entry.
Pops the oldest if it is full.
"""
if self.full():
# Remove item from the deque.
deleted_data = self._queue.popleft()
# Remove measures from sorted lists.
for i, m in enumerate(deleted_data["measures"]):
self._measure_lists[i].remove(m)
self._queue.append(data)
# Add measures to sorted lists.
for i, m in enumerate(data["measures"]):
self._measure_lists[i].add(m)
def full(self):
"""Whether buffer is full."""
return len(self._queue) >= self._buffer_capacity
@property
def sorted_measures(self):
"""(measure_dim, self.size) numpy.ndarray: Sorted measures of each
dimension."""
return np.array(self._measure_lists, dtype=np.float64)
@property
def size(self):
"""Number of solutions stored in the buffer."""
return len(self._queue)
@property
def capacity(self):
"""Capacity of the buffer."""
return self._buffer_capacity
[docs]class SlidingBoundariesArchive(ArchiveBase):
"""An archive with a fixed number of sliding boundaries on each dimension.
This archive is the container described in `Fontaine 2019
<https://arxiv.org/abs/1904.10656>`_. Just like the
:class:`~ribs.archives.GridArchive`, it can be visualized as an
n-dimensional grid in the measure space that is divided into a certain
number of cells in each dimension. Internally, this archive stores a buffer
with the ``buffer_capacity`` most recent solutions and uses them to
determine the boundaries of each dimension of the measure space. After every
``remap_frequency`` solutions are inserted, the archive remaps the
boundaries based on the solutions in the buffer.
Initially, the archive has no solutions, so it cannot automatically
calculate the boundaries. Thus, until the first remap, this archive divides
the measure space defined by ``ranges`` into equally sized cells.
Overall, this archive attempts to make the distribution of the space
illuminated by the archive more accurately match the true distribution of
the measures when they are not uniformly distributed.
.. note:: Unlike other archives, this archive does not currently support
`thresholds <../../tutorials/cma_mae.html>`_ or batched addition (see
:meth:`add`).
Args:
solution_dim (int): Dimension of the solution space.
dims (array-like): Number of cells in each dimension of the measure
space, e.g. ``[20, 30, 40]`` indicates there should be 3 dimensions
with 20, 30, and 40 cells. (The number of dimensions is implicitly
defined in the length of this argument).
ranges (array-like of (float, float)): `Initial` upper and lower bound
of each dimension of the measure space, e.g. ``[(-1, 1), (-2, 2)]``
indicates the first dimension should have bounds :math:`[-1,1]`
(inclusive), and the second dimension should have bounds
:math:`[-2,2]` (inclusive). ``ranges`` should be the same length as
``dims``.
epsilon (float): Due to floating point precision errors, we add a small
epsilon when computing the archive indices in the :meth:`index_of`
method -- refer to the implementation `here
<../_modules/ribs/archives/_sliding_boundaries_archive.html#SlidingBoundariesArchive.index_of>`_.
Pass this parameter to configure that epsilon.
qd_score_offset (float): Archives often contain negative objective
values, and if the QD score were to be computed with these negative
objectives, the algorithm would be penalized for adding new cells
with negative objectives. Thus, a standard practice is to normalize
all the objectives so that they are non-negative by introducing an
offset. This QD score offset will be *subtracted* from all
objectives in the archive, e.g., if your objectives go as low as
-300, pass in -300 so that each objective will be transformed as
``objective - (-300)``.
seed (int): Value to seed the random number generator. Set to None to
avoid a fixed seed.
dtype (str or data-type): Data type of the solutions, objectives,
and measures. We only support ``"f"`` / ``np.float32`` and ``"d"`` /
``np.float64``.
extra_fields (dict): Description of extra fields of data that is stored
next to elite data like solutions and objectives. The description is
a dict mapping from a field name (str) to a tuple of ``(shape,
dtype)``. For instance, ``{"foo": ((), np.float32), "bar": ((10,),
np.float32)}`` will create a "foo" field that contains scalar values
and a "bar" field that contains 10D values. Note that field names
must be valid Python identifiers, and names already used in the
archive are not allowed.
remap_frequency (int): Frequency of remapping. Archive will remap once
after ``remap_frequency`` number of solutions has been found.
buffer_capacity (int): Number of solutions to keep in the buffer.
Solutions in the buffer will be reinserted into the archive after
remapping.
"""
def __init__(self,
*,
solution_dim,
dims,
ranges,
epsilon=1e-6,
qd_score_offset=0.0,
seed=None,
dtype=np.float64,
extra_fields=None,
remap_frequency=100,
buffer_capacity=1000):
self._dims = np.array(dims)
if len(self._dims) != len(ranges):
raise ValueError(f"dims (length {len(self._dims)}) and ranges "
f"(length {len(ranges)}) must be the same length")
ArchiveBase.__init__(
self,
solution_dim=solution_dim,
cells=np.prod(self._dims),
measure_dim=len(self._dims),
qd_score_offset=qd_score_offset,
seed=seed,
dtype=dtype,
extra_fields=extra_fields,
)
ranges = list(zip(*ranges))
self._lower_bounds = np.array(ranges[0], dtype=self.dtype)
self._upper_bounds = np.array(ranges[1], dtype=self.dtype)
self._interval_size = self._upper_bounds - self._lower_bounds
self._epsilon = self.dtype(epsilon)
# Specifics for sliding boundaries.
self._remap_frequency = remap_frequency
# Allocate an extra entry in each row so we can put the upper bound at
# the end.
self._boundaries = np.full((self.measure_dim, np.max(self._dims) + 1),
np.nan,
dtype=self.dtype)
# Initialize the boundaries.
for i, (dim, lower_bound, upper_bound) in enumerate(
zip(self._dims, self._lower_bounds, self._upper_bounds)):
self._boundaries[i, :dim + 1] = np.linspace(lower_bound,
upper_bound, dim + 1)
# Create buffer.
self._buffer = SolutionBuffer(buffer_capacity, self.measure_dim)
# Total number of solutions encountered.
self._total_num_sol = 0
@property
def dims(self):
"""(measure_dim,) numpy.ndarray: Number of cells in each dimension."""
return self._dims
@property
def lower_bounds(self):
"""(measure_dim,) numpy.ndarray: Lower bound of each dimension."""
return self._lower_bounds
@property
def upper_bounds(self):
"""(measure_dim,) numpy.ndarray: Upper bound of each dimension."""
return self._upper_bounds
@property
def interval_size(self):
"""(measure_dim,) numpy.ndarray: The size of each dim (upper_bounds -
lower_bounds)."""
return self._interval_size
@property
def epsilon(self):
""":attr:`dtype`: Epsilon for computing archive indices. Refer to
the documentation for this class."""
return self._epsilon
@property
def remap_frequency(self):
"""int: Frequency of remapping. Archive will remap once after
``remap_frequency`` number of solutions has been found.
"""
return self._remap_frequency
@property
def buffer_capacity(self):
"""int: Maximum capacity of the buffer."""
return self._buffer.capacity
@property
def boundaries(self):
"""list of numpy.ndarray: The dynamic boundaries of the cells in each
dimension.
Entry ``i`` in this list is an array that contains the boundaries of the
cells in dimension ``i``. The array contains ``self.dims[i] + 1``
entries laid out like this::
Archive cells: | 0 | 1 | ... | self.dims[i] |
boundaries[i]: 0 1 2 self.dims[i] - 1 self.dims[i]
Thus, ``boundaries[i][j]`` and ``boundaries[i][j + 1]`` are the lower
and upper bounds of cell ``j`` in dimension ``i``. To access the lower
bounds of all the cells in dimension ``i``, use ``boundaries[i][:-1]``,
and to access all the upper bounds, use ``boundaries[i][1:]``.
"""
return [
bound[:dim + 1] for bound, dim in zip(self._boundaries, self._dims)
]
[docs] def index_of(self, measures):
"""Returns archive indices for the given batch of measures.
First, values are clipped to the bounds of the measure space. Then, the
values are mapped to cells via a binary search along the boundaries in
each dimension.
At this point, we have "grid indices" -- indices of each measure in each
dimension. Since indices returned by this method must be single integers
(as opposed to a tuple of grid indices), we convert these grid indices
into integer indices with :func:`numpy.ravel_multi_index` and return the
result.
It may be useful to have the original grid indices. Thus, we provide the
:meth:`grid_to_int_index` and :meth:`int_to_grid_index` methods for
converting between grid and integer indices.
As an example, the grid indices can be used to access boundaries of a
measure value's cell. For example, the following retrieves the lower
and upper bounds of the cell along dimension 0::
# Access only element 0 since this method operates in batch.
idx = archive.int_to_grid_index(archive.index_of(...))[0]
lower = archive.boundaries[0][idx[0]]
upper = archive.boundaries[0][idx[0] + 1]
See :attr:`boundaries` for more info.
Args:
measures (array-like): (batch_size, :attr:`measure_dim`) array of
coordinates in measure space.
Returns:
numpy.ndarray: (batch_size,) array of integer indices representing
the flattened grid coordinates.
Raises:
ValueError: ``measures`` is not of shape (batch_size,
:attr:`measure_dim`).
"""
measures = np.asarray(measures)
check_batch_shape(measures, "measures", self.measure_dim, "measure_dim")
check_finite(measures, "measures")
# Clip measures + epsilon to the range
# [lower_bounds, upper_bounds - epsilon].
measures = np.clip(measures + self._epsilon, self._lower_bounds,
self._upper_bounds - self._epsilon)
idx_cols = []
for boundary, dim, measures_col in zip(self._boundaries, self._dims,
measures.T):
idx_col = np.searchsorted(boundary[:dim], measures_col)
# The maximum index returned by searchsorted is `dim`, and since we
# subtract 1, the max will be dim - 1 which is within the range of
# the archive indices.
idx_cols.append(np.maximum(0, idx_col - 1))
# We cannot use `grid_to_int_index` since that takes in an array of
# indices, not index columns.
#
# pylint seems to think that ravel_multi_index returns a list and thus
# has no astype method.
# pylint: disable = no-member
return np.ravel_multi_index(idx_cols, self._dims).astype(np.int32)
# Copy these methods from GridArchive.
int_to_grid_index = GridArchive.int_to_grid_index
grid_to_int_index = GridArchive.grid_to_int_index
def _remap(self):
"""Remaps the archive.
The boundaries are relocated to the percentage marks of the distribution
of solutions stored in the archive.
Also re-adds all of the solutions in the buffer and the previous archive
to the archive.
Returns:
tuple: The result of calling :meth:`ArchiveBase.add` on the last
item in the buffer.
"""
# Sort each measure along its dimension.
sorted_measures = self._buffer.sorted_measures
# Calculate new boundaries.
for i in range(self.measure_dim):
for j in range(self.dims[i]):
sample_idx = int(j * self._buffer.size / self.dims[i])
self._boundaries[i][j] = sorted_measures[i][sample_idx]
# Set the upper bound to be the greatest BC.
self._boundaries[i][self.dims[i]] = sorted_measures[i][-1]
cur_data = self._store.data()
# These fields are only computed by the archive.
cur_data.pop("threshold")
cur_data.pop("index")
new_data_single = list(self._buffer) # List of dicts.
new_data = {name: None for name in new_data_single[0]}
for name in new_data:
new_data[name] = [d[name] for d in new_data_single]
# The last solution must be added on its own so that we get an accurate
# status and value to return to the user; hence we pop it from all the
# batches (note that pop removes the last value and returns it).
last_data = {name: arr.pop() for name, arr in new_data.items()}
self.clear()
final_data = {
name: np.concatenate((cur_data[name], new_data[name]))
for name in cur_data
}
ArchiveBase.add(self, **final_data)
add_info = ArchiveBase.add_single(self, **last_data)
return add_info
[docs] def add(self, solution, objective, measures, **fields):
"""Inserts a batch of solutions into the archive.
.. note:: Unlike in other archives, this method (currently) is not truly
batched; rather, it is implemented by calling :meth:`add_single` on
the solutions in the batch, in the order that they are passed in. As
such, this method is *not* invariant to the ordering of the
solutions in the batch.
See :meth:`ArchiveBase.add` for arguments and return values.
"""
new_data = validate_batch(
self,
{
"solution": solution,
"objective": objective,
"measures": measures,
**fields,
},
)
batch_size = new_data["solution"].shape[0]
add_info = {}
for i in range(batch_size):
single_info = self.add_single(**{
name: arr[i] for name, arr in new_data.items()
})
if i == 0:
# Initialize add_info.
for name, val in single_info.items():
add_info[name] = np.empty(batch_size, dtype=val.dtype)
for name, val in single_info.items():
add_info[name][i] = val
return add_info
[docs] def add_single(self, solution, objective, measures, **fields):
"""Inserts a single solution into the archive.
This method remaps the archive after every :attr:`remap_frequency`
solutions are added. Remapping involves changing the boundaries of the
archive to the percentage marks of the measures stored in the buffer and
re-adding all of the solutions stored in the buffer `and` the current
archive.
See :meth:`ArchiveBase.add_single` for arguments and return values.
"""
new_data = validate_single(
self,
{
"solution": solution,
"objective": objective,
"measures": measures,
**fields,
},
)
self._buffer.add(new_data)
self._total_num_sol += 1
if self._total_num_sol % self._remap_frequency == 0:
add_info = self._remap()
self._lower_bounds = np.array(
[bound[0] for bound in self._boundaries])
self._upper_bounds = np.array([
bound[dim] for bound, dim in zip(self._boundaries, self._dims)
])
self._interval_size = self._upper_bounds - self._lower_bounds
else:
add_info = ArchiveBase.add_single(self, **new_data)
return add_info