"""Provides the Bandit Scheduler."""
import warnings
from collections import defaultdict
import numpy as np
from ribs.schedulers._scheduler import Scheduler
[docs]class BanditScheduler:
"""Schedules emitters with a bandit algorithm.
This implementation is based on `Cully 2021
<https://arxiv.org/abs/2007.05352>`_.
.. note::
This class follows the similar ask-tell framework as
:class:`Scheduler`, and enforces similar constraints in the arguments
and methods. Please refer to the documentation of :class:`Scheduler`
for more details.
.. note::
The main difference between :class:`BanditScheduler` and
:class:`Scheduler` is that, unlike :class:`Scheduler`, DQD emitters are
not supported by :class:`BanditScheduler`.
To initialize this class, first create an archive and a list of emitters
for the QD algorithm. The BanditScheduler will schedule the emitters using
the Upper Confidence Bound - 1 algorithm (UCB1). Everytime :meth:`ask` is
called, the emitters are sorted based on the potential reward function from
UCB1. Then, the top `num_active` emitters are used for ask-tell.
Args:
archive (ribs.archives.ArchiveBase): An archive object, e.g. one
selected from :mod:`ribs.archives`.
emitter_pool (list of ribs.archives.EmitterBase): A pool of emitters to
select from, e.g. :class:`ribs.emitters.GaussianEmitter`. On the
first iteration, the first `num_active` emitters from the
emitter_pool will be activated.
num_active (int): The number of active emitters at a time. Active
emitters are used when calling ask-tell.
zeta (float): Hyperparamter of UCB1 that balances the trade-off between
the accuracy and the uncertainty of the emitters. Increasing this
parameter will emphasize the uncertainty of the emitters. Refer to
the original paper for more information.
reselect (str): Indicates how emitters are reselected from the pool.
The default is "terminated", where only terminated/restarted
emitters are deactivated and reselected (but they might be selected
again). Alternatively, use "all" to reselect all active emitters
every iteration.
add_mode (str): Indicates how solutions should be added to the archive.
The default is "batch", which adds all solutions with one call to
:meth:`~ribs.archives.ArchiveBase.add`. Alternatively, use "single"
to add the solutions one at a time with
:meth:`~ribs.archives.ArchiveBase.add_single`. "single" mode is
included for legacy reasons, as it was the only mode of operation
in pyribs 0.4.0 and before. We highly recommend using "batch" mode
since it is significantly faster.
result_archive (ribs.archives.ArchiveBase): In some algorithms, such as
CMA-MAE, the archive does not store all the best-performing
solutions. The ``result_archive`` is a secondary archive where we
can store all the best-performing solutions.
Raises:
TypeError: The ``emitter_pool`` argument was not a list of emitters.
ValueError: Number of active emitters is less than one.
ValueError: Less emitters in the pool than the number of active
emitters.
ValueError: The emitters passed in do not have the same solution
dimensions.
ValueError: The same emitter instance was passed in multiple times.
Each emitter should be a unique instance (see the warning above).
ValueError: Invalid value for ``add_mode``.
ValueError: The ``result_archive`` and ``archive`` are the same object
(``result_archive`` should not be passed in in this case).
ValueError: The ``result_archive`` and ``archive`` have different
fields.
"""
def __init__(self,
archive,
emitter_pool,
num_active,
*,
reselect="terminated",
zeta=0.05,
result_archive=None,
add_mode="batch"):
if num_active < 1:
raise ValueError("num_active cannot be less than 1.")
try:
if len(emitter_pool) < num_active:
raise ValueError(f"The emitter pool must contain at least"
f"num_active emitters, but only"
f"{len(emitter_pool)} are given.")
except TypeError as exception:
# TypeError will be raised by len(). We avoid directly checking if
# `emitter_pool` is an instance of list since we do not want to be
# too restrictive.
raise TypeError("`emitter_pool` must be a list of emitter objects."
) from exception
emitter_ids = set(id(e) for e in emitter_pool)
if len(emitter_ids) != len(emitter_pool):
raise ValueError(
"Not all emitters passed in were unique (i.e. some emitters "
"had the same id). If emitters were created with something "
"like [EmitterClass(...)] * n, instead use "
"[EmitterClass(...) for _ in range(n)] so that all emitters "
"are unique instances.")
self._solution_dim = emitter_pool[0].solution_dim
for idx, emitter in enumerate(emitter_pool[1:]):
if emitter.solution_dim != self._solution_dim:
raise ValueError(
"All emitters must have the same solution dim, but "
f"Emitter {idx} has dimension {emitter.solution_dim}, "
f"while Emitter 0 has dimension {self._solution_dim}")
if reselect not in ["terminated", "all"]:
raise ValueError("add_mode must either be 'terminated' or 'all',"
f"but it was '{reselect}'")
if add_mode not in ["single", "batch"]:
raise ValueError("add_mode must either be 'batch' or 'single', but "
f"it was '{add_mode}'")
if archive is result_archive:
raise ValueError("`archive` has same id as `result_archive` -- "
"Note that `Scheduler.result_archive` already "
"defaults to be the same as `archive` if you pass "
"`result_archive=None`")
if (result_archive is not None and
set(archive.field_list) != set(result_archive.field_list)):
raise ValueError("`archive` and `result_archive` should have the "
"same set of fields. This may be the result of "
"passing extra_fields to archive but not to "
"result_archive.")
self._archive = archive
self._emitter_pool = np.array(emitter_pool)
self._num_active = num_active
self._add_mode = add_mode
self._result_archive = result_archive
self._reselect = reselect
# Boolean mask of the active emitters. Initializes to the first
# num_active emitters in the emitter pool.
self._active_arr = np.zeros_like(self._emitter_pool, dtype=bool)
# Used by UCB1 to select emitters.
self._success = np.zeros_like(self._emitter_pool, dtype=float)
self._selection = np.zeros_like(self._emitter_pool, dtype=float)
self._restarts = np.zeros_like(self._emitter_pool, dtype=int)
self._zeta = zeta
# Keeps track of whether the scheduler should be receiving a call to
# ask() or tell().
self._last_called = None
# The last set of solutions returned by ask().
self._cur_solutions = []
# The number of solutions created by each emitter.
self._num_emitted = np.array([None for _ in self._active_arr])
@property
def archive(self):
"""ribs.archives.ArchiveBase: Archive for storing solutions found in
this scheduler."""
return self._archive
@property
def emitters(self):
"""list of ribs.archives.EmitterBase: Emitters for generating solutions
in this scheduler."""
return self._active_arr
@property
def result_archive(self):
"""ribs.archives.ArchiveBase: Another archive for storing solutions
found in this optimizer.
If `result_archive` was not passed to the constructor, this property is
the same as :attr:`archive`.
"""
return (self._archive
if self._result_archive is None else self._result_archive)
[docs] def ask_dqd(self):
"""Generates a batch of solutions by calling ask_dqd() on all DQD
emitters.
This method is not supported for this scheduler and throws an error if
called.
Raises:
NotImplementedError: This method is not supported by this
scheduler.
"""
raise NotImplementedError("ask_dqd() is not supported by"
"BanditScheduler.")
[docs] def ask(self):
"""Generates a batch of solutions by calling ask() on all active
emitters.
The emitters used by ask are determined by the UCB1 algorithm. Briefly,
emitters that have never been selected before are prioritized, then
emitters are sorted in descending order based on the accurary of their
past prediction.
.. note:: The order of the solutions returned from this method is
important, so do not rearrange them.
Returns:
(batch_size, dim) array: An array of n solutions to evaluate. Each
row contains a single solution.
Raises:
RuntimeError: This method was called without first calling
:meth:`tell`.
"""
if self._last_called == "ask":
raise RuntimeError("ask cannot be called immediately after " +
self._last_called)
self._last_called = "ask"
if self._reselect == "terminated":
# Reselect terminated emitters. Emitters are terminated if their
# restarts attribute have incremented.
emitter_restarts = np.array([
emitter.restarts if hasattr(emitter, "restarts") else -1
for emitter in self._emitter_pool
])
reselect = emitter_restarts > self._restarts
# If the emitter does not have "restarts" attribute, assume it
# restarts every iteration.
reselect[emitter_restarts < 0] = True
self._restarts = emitter_restarts
else:
# Reselect all emitters.
reselect = self._active_arr.copy()
# If no emitters are active, activate the first num_active.
if not self._active_arr.any():
reselect[:] = False
self._active_arr[:self._num_active] = True
# Deactivate emitters to be reselected.
self._active_arr[reselect] = False
# Select emitters based on the UCB1 formula.
# The ranking of emitters also follows these rules:
# - Emitters that have never been selected are prioritized.
# - If reselect is "terminated", then only active emitters that have
# terminated/restarted will be reselected. Otherwise, if reselect is
# "all", then all emitters are reselected.
if reselect.any():
ucb1 = np.full_like(self._emitter_pool, np.inf)
update_ucb = self._selection != 0
if update_ucb.any():
ucb1[update_ucb] = (
self._success[update_ucb] / self._selection[update_ucb] +
self._zeta * np.sqrt(
np.log(self._success.sum()) /
self._selection[update_ucb]))
# Activate top emitters based on UCB1.
activate = np.argsort(ucb1)[-reselect.sum():]
self._active_arr[activate] = True
self._cur_solutions = []
for i in np.where(self._active_arr)[0]:
emitter = self._emitter_pool[i]
emitter_sols = emitter.ask()
self._cur_solutions.append(emitter_sols)
self._num_emitted[i] = len(emitter_sols)
# In case the emitters didn't return any solutions.
self._cur_solutions = np.concatenate(
self._cur_solutions, axis=0) if self._cur_solutions else np.empty(
(0, self._solution_dim))
return self._cur_solutions
def _check_length(self, name, arr):
"""Raises a ValueError if array does not have the same length as the
solutions."""
if len(arr) != len(self._cur_solutions):
raise ValueError(
f"{name} should have length {len(self._cur_solutions)} "
"(this is the number of solutions output by ask()) but "
f"has length {len(arr)}")
def _validate_tell_data(self, data):
"""Preprocesses data passed into tell methods."""
for name, arr in data.items():
data[name] = np.asarray(arr)
self._check_length(name, arr)
# Convenient to have solutions be part of data, so that everything is
# just one dict.
data["solution"] = self._cur_solutions
return data
[docs] def tell_dqd(self, objective, measures, jacobian):
"""Returns info for solutions from :meth:`ask_dqd`.
This method is not supported for this scheduler and throws an error if
called.
Raises:
NotImplementedError: This method is not supported by this
scheduler.
"""
raise NotImplementedError("tell_dqd() is not supported by"
"BanditScheduler.")
[docs] def tell(self, objective, measures, **fields):
"""Returns info for solutions from :meth:`ask`.
The emitters are the same with those used in the last call to
:meth:`ask`.
.. note:: The objective and measures arrays must be in the same order as
the solutions created by :meth:`ask_dqd`; i.e. ``objective[i]`` and
``measures[i]`` should be the objective and measures for
``solution[i]``.
Args:
objective_batch ((batch_size,) array): Each entry of this array
contains the objective function evaluation of a solution.
measures_batch ((batch_size, measures_dm) array): Each row of
this array contains a solution's coordinates in measure space.
fields (keyword arguments): Additional data for each solution. Each
argument should be an array with batch_size as the first
dimension.
Raises:
RuntimeError: This method is called without first calling
:meth:`ask`.
ValueError: One of the inputs has the wrong shape.
"""
if self._last_called != "ask":
raise RuntimeError("tell() was called without calling ask().")
self._last_called = "tell"
data = self._validate_tell_data({
"objective": objective,
"measures": measures,
**fields,
})
archive_empty_before = self.archive.empty
if self._result_archive is not None:
# Check self._result_archive here since self.result_archive is a
# property that always provides a proper archive.
result_archive_empty_before = self.result_archive.empty
# Add solutions to the archive.
if self._add_mode == "batch":
add_info = self.archive.add(**data)
# Add solutions to result_archive.
if self._result_archive is not None:
self._result_archive.add(**data)
elif self._add_mode == "single":
add_info = defaultdict(list)
for i in range(len(self._cur_solutions)):
single_data = {name: arr[i] for name, arr in data.items()}
single_info = self.archive.add_single(**single_data)
for name, val in single_info.items():
add_info[name].append(val)
# Add solutions to result_archive.
if self._result_archive is not None:
self._result_archive.add_single(**single_data)
for name, arr in add_info.items():
add_info[name] = np.asarray(arr)
# Warn the user if nothing was inserted into the archives.
if archive_empty_before and self.archive.empty:
warnings.warn(Scheduler.EMPTY_WARNING.format(name="archive"))
if self._result_archive is not None:
if result_archive_empty_before and self.result_archive.empty:
warnings.warn(
Scheduler.EMPTY_WARNING.format(name="result_archive"))
# Keep track of pos because emitters may have different batch sizes.
pos = 0
for i in np.where(self._active_arr)[0]:
emitter = self._emitter_pool[i]
n = self._num_emitted[i]
end = pos + n
self._selection[i] += n
self._success[i] += np.count_nonzero(add_info["status"][pos:end])
emitter.tell(
**{
name: arr[pos:end] for name, arr in data.items()
},
add_info={
name: arr[pos:end] for name, arr in add_info.items()
},
)
pos = end