"""Provides the GradientArborescenceEmitter."""
import numbers
import numpy as np
from ribs._utils import check_shape, validate_batch
from ribs.emitters._emitter_base import EmitterBase
from ribs.emitters.opt import _get_es, _get_grad_opt
from ribs.emitters.rankers import _get_ranker
[docs]class GradientArborescenceEmitter(EmitterBase):
"""Generates solutions with a gradient arborescence, with coefficients
parameterized by an evolution strategy.
This emitter originates in `Fontaine 2021
<https://arxiv.org/abs/2106.03894>`_. It leverages the gradient information
of the objective and measure functions, generating new solutions around a
*solution point* :math:`\\boldsymbol{\\theta}` using *gradient
arborescence*, with coefficients drawn from a Gaussian distribution.
Essentially, this means that the emitter samples coefficients
:math:`\\boldsymbol{c_i} \\sim
\\mathcal{N}(\\boldsymbol{\\mu}, \\boldsymbol{\\Sigma})`
and creates new solutions :math:`\\boldsymbol{\\theta'_i}` according to
.. math::
\\boldsymbol{\\theta'_i} \\gets \\boldsymbol{\\theta} +
c_{i,0} \\boldsymbol{\\nabla} f(\\boldsymbol{\\theta}) +
\\sum_{j=1}^k c_{i,j}\\boldsymbol{\\nabla}m_j(\\boldsymbol{\\theta})
Where :math:`k` is the number of measures, and
:math:`\\boldsymbol{\\nabla} f(\\boldsymbol{\\theta})` and
:math:`\\boldsymbol{\\nabla} m_j(\\boldsymbol{\\theta})` are the objective
and measure gradients of the solution point :math:`\\boldsymbol{\\theta}`,
respectively.
Based on how the solutions are ranked after being inserted into the archive
(see ``ranker``), the solution point :math:`\\boldsymbol{\\theta}` is
updated with gradient ascent, and the coefficient distribution parameters
:math:`\\boldsymbol{\\mu}` and :math:`\\boldsymbol{\\Sigma}` are updated
with an ES (the default ES is CMA-ES).
.. note::
Unlike non-gradient emitters, GradientArborescenceEmitter requires
calling :meth:`ask_dqd` and :meth:`tell_dqd` (in this order) before
calling :meth:`ask` and :meth:`tell` to communicate the gradient
information to the emitter.
See Also:
Our DQD tutorial goes into detail on how to use this emitter:
:doc:`/tutorials/tom_cruise_dqd`
Args:
archive (ribs.archives.ArchiveBase): An archive to use when creating and
inserting solutions. For instance, this can be
:class:`ribs.archives.GridArchive`.
x0 (np.ndarray): Initial solution.
sigma0 (float): Initial step size / standard deviation of the
distribution of gradient coefficients.
lr (float): Learning rate for the gradient optimizer.
ranker (Callable or str): The ranker is a
:class:`~ribs.emitters.rankers.RankerBase` object that orders the
solutions after they have been evaluated in the environment. This
parameter may be a callable (e.g. a class or a lambda function) that
takes in no parameters and returns an instance of
:class:`~ribs.emitters.rankers.RankerBase`, or it may be a full or
abbreviated ranker name as described in
:mod:`ribs.emitters.rankers`.
selection_rule ("mu" or "filter"): Method for selecting parents in
CMA-ES. With "mu" selection, the first half of the solutions will be
selected as parents, while in "filter", any solutions that were
added to the archive will be selected.
restart_rule (int, "no_improvement", and "basic"): Method to use when
checking for restarts. If given an integer, then the emitter will
restart after this many iterations, where each iteration is a call
to :meth:`tell`. With "basic", only the default CMA-ES convergence
rules will be used, while with "no_improvement", the emitter will
restart when none of the proposed solutions were added to the
archive.
grad_opt (Callable or str): Gradient optimizer to use for the gradient
ascent step of the algorithm. The optimizer is a
:class:`~ribs.emitters.opt.GradientOptBase` object. This parameter
may be a callable (e.g. a class or a lambda function) which takes in
the ``theta0`` and ``lr`` arguments, or it may be a full or
abbreviated name as described in :mod:`ribs.emitters.opt`.
grad_opt_kwargs (dict): Additional arguments to pass to the gradient
optimizer. See the gradient-based optimizers in
:mod:`ribs.emitters.opt` for the arguments allowed by each
optimizer. Note that we already pass in ``theta0`` and ``lr``.
es (Callable or str): The evolution strategy is an
:class:`~ribs.emitters.opt.EvolutionStrategyBase` object that is
used to adapt the distribution from which new solutions are sampled.
This parameter may be a callable (e.g. a class or a lambda function)
that takes in the parameters of
:class:`~ribs.emitters.opt.EvolutionStrategyBase` along with kwargs
provided by the ``es_kwargs`` argument, or it may be a full or
abbreviated optimizer name as described in :mod:`ribs.emitters.opt`.
es_kwargs (dict): Additional arguments to pass to the evolution
strategy optimizer. See the evolution-strategy-based optimizers in
:mod:`ribs.emitters.opt` for the arguments allowed by each
optimizer.
normalize_grad (bool): If true (default), then gradient infomation will
be normalized. Otherwise, it will not be normalized.
bounds: This argument may be used for providing solution space bounds in
the future. This emitter does not currently support solution space
bounds, as bounding solutions for DQD algorithms such as CMA-MEGA is
an open problem. Hence, this argument must be set to None.
batch_size (int): Number of solutions to return in :meth:`ask`. If not
passed in, a batch size will be automatically calculated using the
default CMA-ES rules. This **does not** account for the **one**
solution returned by :meth:`ask_dqd`, which is the solution point
maintained by the gradient optimizer.
epsilon (float): For numerical stability, we add a small epsilon when
normalizing gradients in :meth:`tell_dqd` -- refer to the
implementation `here
<../_modules/ribs/emitters/_gradient_arborescence_emitter.html#GradientArborescenceEmitter.tell_dqd>`_.
Pass this parameter to configure that epsilon.
seed (int): Value to seed the random number generator. Set to None to
avoid a fixed seed.
Raises:
ValueError: There is an error in x0 or initial_solutions.
ValueError: ``bounds`` is set even though it is not currently supported.
ValueError: If ``restart_rule``, ``selection_rule``, or ``ranker`` is
invalid.
"""
def __init__(self,
archive,
*,
x0,
sigma0,
lr,
ranker="2imp",
selection_rule="filter",
restart_rule="no_improvement",
grad_opt="adam",
grad_opt_kwargs=None,
es="cma_es",
es_kwargs=None,
normalize_grad=True,
bounds=None,
batch_size=None,
epsilon=1e-8,
seed=None):
if bounds is not None:
raise ValueError(
"`bounds` must be set to None. The GradientArborescenceEmitter "
"does not currently support solution space bounds, as bounding "
"solutions for DQD algorithms such as CMA-MEGA is an open "
"problem.")
EmitterBase.__init__(
self,
archive,
solution_dim=archive.solution_dim,
bounds=bounds,
)
seed_sequence = (seed if isinstance(seed, np.random.SeedSequence) else
np.random.SeedSequence(seed))
opt_seed, ranker_seed = seed_sequence.spawn(2)
self._epsilon = epsilon
self._x0 = np.array(x0, dtype=archive.dtype)
check_shape(self._x0, "x0", archive.solution_dim,
"archive.solution_dim")
self._sigma0 = sigma0
self._normalize_grads = normalize_grad
self._jacobian_batch = None
self._ranker = _get_ranker(ranker, ranker_seed)
self._ranker.reset(self, archive)
if selection_rule not in ["mu", "filter"]:
raise ValueError(f"Invalid selection_rule {selection_rule}")
self._selection_rule = selection_rule
self._restart_rule = restart_rule
self._restarts = 0
self._itrs = 0
# Check if the restart_rule is valid, discard check_restart result.
_ = self._check_restart(0)
# We have a coefficient for each measure and an extra coefficient for
# the objective.
self._num_coefficients = archive.measure_dim + 1
# Initialize gradient optimizer.
self._grad_opt = _get_grad_opt(
grad_opt,
theta0=self._x0,
lr=lr,
**(grad_opt_kwargs if grad_opt_kwargs is not None else {}))
self._opt = _get_es(
es,
sigma0=sigma0,
batch_size=batch_size,
solution_dim=self._num_coefficients,
seed=opt_seed,
dtype=self.archive.dtype,
lower_bounds=-np.inf, # No bounds for gradient coefficients.
upper_bounds=np.inf,
**(es_kwargs if es_kwargs is not None else {}),
)
self._opt.reset(np.zeros(self._num_coefficients))
self._batch_size = self._opt.batch_size
self._itrs = 0
@property
def x0(self):
"""numpy.ndarray: Initial solution for the optimizer."""
return self._x0
@property
def batch_size(self):
"""int: Number of solutions to return in :meth:`ask`."""
return self._batch_size
@property
def batch_size_dqd(self):
"""int: Number of solutions to return in :meth:`ask_dqd`.
This is always 1, as we only return the solution point in
:meth:`ask_dqd`.
"""
return 1
@property
def restarts(self):
"""int: The number of restarts for this emitter."""
return self._restarts
@property
def itrs(self):
"""int: The number of iterations for this emitter."""
return self._itrs
@property
def epsilon(self):
"""int: The epsilon added for numerical stability when normalizing
gradients in :meth:`tell_dqd`."""
return self._epsilon
[docs] def ask_dqd(self):
"""Samples a new solution from the gradient optimizer.
**Call :meth:`ask_dqd` and :meth:`tell_dqd` (in this order) before
calling :meth:`ask` and :meth:`tell`.**
Returns:
a new solution to evaluate.
"""
return self._grad_opt.theta[None]
[docs] def ask(self):
"""Samples new solutions from a gradient arborescence parameterized by a
multivariate Gaussian distribution.
The multivariate Gaussian is parameterized by the evolution strategy
optimizer ``self._opt``.
This method returns ``batch_size`` solutions, even though one solution
is returned via ``ask_dqd``.
Returns:
(:attr:`batch_size`, :attr:`solution_dim`) array -- a batch of new
solutions to evaluate.
Raises:
RuntimeError: This method was called without first passing gradients
with calls to ask_dqd() and tell_dqd().
"""
if self._jacobian_batch is None:
raise RuntimeError("Please call ask_dqd() and tell_dqd() "
"before calling ask().")
grad_coeffs = self._opt.ask()[:, :, None]
return (self._grad_opt.theta +
np.sum(self._jacobian_batch * grad_coeffs, axis=1))
def _check_restart(self, num_parents):
"""Emitter-side checks for restarting the optimizer.
The optimizer also has its own checks.
Args:
num_parents (int): The number of solution to propagate to the next
generation from the solutions generated by CMA-ES.
Raises:
ValueError: If :attr:`restart_rule` is invalid.
"""
if isinstance(self._restart_rule, numbers.Integral):
return self._itrs % self._restart_rule == 0
if self._restart_rule == "no_improvement":
return num_parents == 0
if self._restart_rule == "basic":
return False
raise ValueError(f"Invalid restart_rule {self._restart_rule}")
[docs] def tell_dqd(self, solution, objective, measures, jacobian, add_info,
**fields):
"""Gives the emitter results from evaluating the gradient of the
solutions.
Args:
solution (array-like): (batch_size, :attr:`solution_dim`) array of
solutions generated by this emitter's :meth:`ask()` method.
objective (array-like): 1D array containing the objective function
value of each solution.
measures (array-like): (batch_size, measure space dimension) array
with the measure space coordinates of each solution.
jacobian (array-like): (batch_size, 1 + measure_dim, solution_dim)
array consisting of Jacobian matrices of the solutions obtained
from :meth:`ask_dqd`. Each matrix should consist of the
objective gradient of the solution followed by the measure
gradients.
add_info (dict): Data returned from the archive
:meth:`~ribs.archives.ArchiveBase.add` method.
fields (keyword arguments): Additional data for each solution. Each
argument should be an array with batch_size as the first
dimension.
"""
data, add_info, jacobian = validate_batch( # pylint: disable = unused-variable
self.archive,
{
"solution": solution,
"objective": objective,
"measures": measures,
**fields,
},
add_info,
jacobian,
)
if self._normalize_grads:
norms = (np.linalg.norm(jacobian, axis=2, keepdims=True) +
self._epsilon)
jacobian /= norms
self._jacobian_batch = jacobian
[docs] def tell(self, solution, objective, measures, add_info, **fields):
"""Gives the emitter results from evaluating solutions.
The solutions are ranked based on the `rank()` function defined by
`self._ranker`.
Args:
solution (array-like): (batch_size, :attr:`solution_dim`) array of
solutions generated by this emitter's :meth:`ask()` method.
objective (array-like): 1D array containing the objective function
value of each solution.
measures (array-like): (batch_size, measure space dimension) array
with the measure space coordinates of each solution.
add_info (dict): Data returned from the archive
:meth:`~ribs.archives.ArchiveBase.add` method.
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 was called without first passing gradients
with calls to ask_dqd() and tell_dqd().
"""
data, add_info = validate_batch(
self.archive,
{
"solution": solution,
"objective": objective,
"measures": measures,
**fields,
},
add_info,
)
if self._jacobian_batch is None:
raise RuntimeError("Please call ask_dqd(), tell_dqd(), and ask() "
"before calling tell().")
# Increase iteration counter.
self._itrs += 1
# Count number of new solutions.
new_sols = add_info["status"].astype(bool).sum()
# Sort the solutions using ranker.
indices, ranking_values = self._ranker.rank(self, self.archive, data,
add_info)
# Select the number of parents.
num_parents = (new_sols if self._selection_rule == "filter" else
self._batch_size // 2)
# Update Evolution Strategy.
self._opt.tell(indices, ranking_values, num_parents)
# Calculate a new mean in solution space. These weights are from CMA-ES.
parents = data["solution"][indices]
parents = parents[:num_parents]
weights = (np.log(num_parents + 0.5) -
np.log(np.arange(1, num_parents + 1)))
weights = weights / np.sum(weights) # Normalize weights
new_mean = np.sum(parents * np.expand_dims(weights, axis=1), axis=0)
# Use the mean to calculate a gradient step and step the optimizer.
gradient_step = new_mean - self._grad_opt.theta
self._grad_opt.step(gradient_step)
# Check for reset.
if (self._opt.check_stop(ranking_values[indices]) or
self._check_restart(new_sols)):
new_coeff = self.archive.sample_elites(1)["solution"][0]
self._grad_opt.reset(new_coeff)
self._opt.reset(np.zeros(self._num_coefficients))
self._ranker.reset(self, self.archive)
self._restarts += 1