def __init__(self, worker_id, data, response_surface, maxeval, nsamples,
exp_design=None, sampling_method=None, extra=None):
"""Initialize the optimization strategy.
:param worker_id: Start ID in a multistart setting
:param data: Problem parameter data structure
:param response_surface: Surrogate model object
:param maxeval: Function evaluation budget
:param nsamples: Number of simultaneous fevals allowed
:param exp_design: Experimental design
:param search_procedure: Search procedure for finding
points to evaluate
:param extra: Points to be added to the experimental design
"""
self.worker_id = worker_id
self.data = data
self.fhat = response_surface
if self.fhat is None:
self.fhat = RBFInterpolant(surftype=CubicRBFSurface, maxp=maxeval)
self.maxeval = maxeval
self.nsamples = nsamples
self.extra = extra
# Default to generate sampling points using Symmetric Latin Hypercube
self.design = exp_design
if self.design is None:
if self.data.dim > 50:
self.design = LatinHypercube(data.dim, data.dim+1)
else:
self.design = SymmetricLatinHypercube(data.dim, 2*(data.dim+1))
self.xrange = np.asarray(data.xup - data.xlow)
# algorithm parameters
self.sigma_min = 0.005
self.sigma_max = 0.2
self.sigma_init = 0.2
self.failtol = max(5, data.dim)
self.succtol = 3
self.numeval = 0
self.status = 0
self.sigma = 0
self.resubmitter = RetryStrategy()
self.xbest = None
self.fbest = np.inf
self.fbest_old = None
# Set up search procedures and initialize
self.sampling = sampling_method
if self.sampling is None:
self.sampling = CandidateDYCORS(data)
self.check_input()
# Start with first experimental design
self.sample_initial()
def __init__(self,
worker_id,
data,
response_surface,
maxeval,
nsamples,
exp_design=None,
sampling_method=None,
extra=None):
"""Initialize the optimization strategy.
:param worker_id: Start ID in a multistart setting
:param data: Problem parameter data structure
:param response_surface: Surrogate model object
:param maxeval: Function evaluation budget
:param nsamples: Number of simultaneous fevals allowed
:param exp_design: Experimental design
:param sampling_method: Sampling method for finding
points to evaluate
:param extra: Points to be added to the experimental design
"""
self.worker_id = worker_id
self.data = data
self.fhat = response_surface
if self.fhat is None:
self.fhat = RBFInterpolant(surftype=CubicRBFSurface, maxp=maxeval)
self.maxeval = maxeval
self.nsamples = nsamples
self.extra = extra
# Default to generate sampling points using Symmetric Latin Hypercube
self.design = exp_design
if self.design is None:
if self.data.dim > 50:
self.design = LatinHypercube(data.dim, data.dim + 1)
else:
self.design = SymmetricLatinHypercube(data.dim,
2 * (data.dim + 1))
self.xrange = np.asarray(data.xup - data.xlow)
# algorithm parameters
self.sigma_min = 0.005
self.sigma_max = 0.2
self.sigma_init = 0.2
self.failtol = max(5, data.dim)
self.succtol = 3
self.numeval = 0
self.status = 0
self.sigma = 0
self.resubmitter = RetryStrategy()
self.xbest = None
self.fbest = np.inf
self.fbest_old = None
# Set up search procedures and initialize
self.sampling = sampling_method
if self.sampling is None:
self.sampling = CandidateDYCORS(data)
self.check_input()
# Start with first experimental design
self.sample_initial()
class SyncStrategyNoConstraints(BaseStrategy):
"""Parallel synchronous optimization strategy without non-bound constraints.
This class implements the parallel synchronous SRBF strategy
described by Regis and Shoemaker. After the initial experimental
design (which is embarrassingly parallel), the optimization
proceeds in phases. During each phase, we allow nsamples
simultaneous function evaluations. We insist that these
evaluations run to completion -- if one fails for whatever reason,
we will resubmit it. Samples are drawn randomly from around the
current best point, and are sorted according to a merit function
based on distance to other sample points and predicted function
values according to the response surface. After several
successive significant improvements, we increase the sampling
radius; after several failures to improve the function value, we
decrease the sampling radius. We restart once the sampling radius
decreases below a threshold.
"""
def __init__(self, worker_id, data, response_surface, maxeval, nsamples,
exp_design=None, sampling_method=None, extra=None):
"""Initialize the optimization strategy.
:param worker_id: Start ID in a multistart setting
:param data: Problem parameter data structure
:param response_surface: Surrogate model object
:param maxeval: Function evaluation budget
:param nsamples: Number of simultaneous fevals allowed
:param exp_design: Experimental design
:param search_procedure: Search procedure for finding
points to evaluate
:param extra: Points to be added to the experimental design
"""
self.worker_id = worker_id
self.data = data
self.fhat = response_surface
if self.fhat is None:
self.fhat = RBFInterpolant(surftype=CubicRBFSurface, maxp=maxeval)
self.maxeval = maxeval
self.nsamples = nsamples
self.extra = extra
# Default to generate sampling points using Symmetric Latin Hypercube
self.design = exp_design
if self.design is None:
if self.data.dim > 50:
self.design = LatinHypercube(data.dim, data.dim+1)
else:
self.design = SymmetricLatinHypercube(data.dim, 2*(data.dim+1))
self.xrange = np.asarray(data.xup - data.xlow)
# algorithm parameters
self.sigma_min = 0.005
self.sigma_max = 0.2
self.sigma_init = 0.2
self.failtol = max(5, data.dim)
self.succtol = 3
self.numeval = 0
self.status = 0
self.sigma = 0
self.resubmitter = RetryStrategy()
self.xbest = None
self.fbest = np.inf
self.fbest_old = None
# Set up search procedures and initialize
self.sampling = sampling_method
if self.sampling is None:
self.sampling = CandidateDYCORS(data)
self.check_input()
# Start with first experimental design
self.sample_initial()
def check_input(self):
assert not hasattr(self.data, "eval_ineq_constraints"), "Objective function has constraints,\n" \
"SyncStrategyNoConstraints can't handle constraints"
assert not hasattr(self.data, "eval_eq_constraints"), "Objective function has constraints,\n" \
"SyncStrategyNoConstraints can't handle constraints"
def proj_fun(self, x):
x = np.atleast_2d(x)
return round_vars(self.data, x)
def log_completion(self, record):
"""Record a completed evaluation to the log.
:param record: Record of the function evaluation
"""
xstr = np.array_str(record.params[0], max_line_width=np.inf,
precision=5, suppress_small=True)
logger.info("Feasible {:.3e} @ {}".format(record.value, xstr))
def adjust_step(self):
"""Adjust the sampling radius sigma.
After succtol successful steps, we cut the sampling radius;
after failtol failed steps, we double the sampling radius.
:ivar Fnew: Best function value in new step
:ivar fbest: Previous best function evaluation
"""
# Initialize if this is the first adaptive step
if self.fbest_old is None:
self.fbest_old = self.fbest
return
# Check if we succeeded at significant improvement
if self.fbest < self.fbest_old - 1e-3 * math.fabs(self.fbest_old):
self.status = max(1, self.status + 1)
else:
self.status = min(-1, self.status - 1)
self.fbest_old = self.fbest
# Check if step needs adjusting
if self.status <= -self.failtol:
self.status = 0
self.sigma /= 2
logger.info("Reducing sigma")
if self.status >= self.succtol:
self.status = 0
self.sigma = min([2.0 * self.sigma, self.sigma_max])
logger.info("Increasing sigma")
def sample_initial(self):
"""Generate and queue an initial experimental design.
"""
if self.numeval == 0:
logger.info("=== Start ===")
else:
logger.info("=== Restart ===")
self.fhat.reset()
self.sigma = self.sigma_init
self.status = 0
self.xbest = None
self.fbest_old = None
self.fbest = np.inf
self.fhat.reset()
start_sample = self.design.generate_points()
assert start_sample.shape[1] == self.data.dim, \
"Dimension mismatch between problem and experimental design"
start_sample = from_unit_box(start_sample, self.data)
# Add extra evaluation points provided by the user
if self.extra is not None:
start_sample = np.vstack((start_sample, self.extra))
for j in range(min(start_sample.shape[0], self.maxeval - self.numeval)):
start_sample[j, :] = self.proj_fun(start_sample[j, :]) # Project onto feasible region
proposal = self.propose_eval(start_sample[j, :])
self.resubmitter.rput(proposal)
self.sampling.init(start_sample, self.fhat, self.maxeval - self.numeval)
def sample_adapt(self):
"""Generate and queue samples from the search strategy
"""
self.adjust_step()
nsamples = min(self.nsamples, self.maxeval - self.numeval)
new_points = self.sampling.make_points(npts=nsamples, xbest=self.xbest, sigma=self.sigma,
proj_fun=self.proj_fun)
for i in range(nsamples):
proposal = self.propose_eval(np.ravel(new_points[i, :]))
self.resubmitter.rput(proposal)
def start_batch(self):
"""Generate and queue a new batch of points
"""
if self.sigma < self.sigma_min:
self.sample_initial()
else:
self.sample_adapt()
def propose_action(self):
"""Propose an action
"""
if self.numeval == self.maxeval:
return self.propose_terminate()
elif self.resubmitter.num_eval_outstanding == 0:
self.start_batch()
return self.resubmitter.get()
def on_complete(self, record):
"""Handle completed function evaluation.
When a function evaluation is completed we need to ask the constraint
handler if the function value should be modified which is the case for
say a penalty method. We also need to print the information to the
logfile, update the best value found so far and notify the GUI that
an evaluation has completed.
:param record: Evaluation record
"""
self.log_completion(record)
self.numeval += 1
record.worker_id = self.worker_id
record.worker_numeval = self.numeval
self.fhat.add_point(record.params[0], record.value)
if record.value < self.fbest:
self.xbest = record.params[0]
self.fbest = record.value
class SyncStrategyNoConstraints(BaseStrategy):
"""Parallel synchronous optimization strategy without non-bound constraints.
This class implements the parallel synchronous SRBF strategy
described by Regis and Shoemaker. After the initial experimental
design (which is embarrassingly parallel), the optimization
proceeds in phases. During each phase, we allow nsamples
simultaneous function evaluations. We insist that these
evaluations run to completion -- if one fails for whatever reason,
we will resubmit it. Samples are drawn randomly from around the
current best point, and are sorted according to a merit function
based on distance to other sample points and predicted function
values according to the response surface. After several
successive significant improvements, we increase the sampling
radius; after several failures to improve the function value, we
decrease the sampling radius. We restart once the sampling radius
decreases below a threshold.
"""
def __init__(self,
worker_id,
data,
response_surface,
maxeval,
nsamples,
exp_design=None,
sampling_method=None,
extra=None):
"""Initialize the optimization strategy.
:param worker_id: Start ID in a multistart setting
:param data: Problem parameter data structure
:param response_surface: Surrogate model object
:param maxeval: Function evaluation budget
:param nsamples: Number of simultaneous fevals allowed
:param exp_design: Experimental design
:param sampling_method: Sampling method for finding
points to evaluate
:param extra: Points to be added to the experimental design
"""
self.worker_id = worker_id
self.data = data
self.fhat = response_surface
if self.fhat is None:
self.fhat = RBFInterpolant(surftype=CubicRBFSurface, maxp=maxeval)
self.maxeval = maxeval
self.nsamples = nsamples
self.extra = extra
# Default to generate sampling points using Symmetric Latin Hypercube
self.design = exp_design
if self.design is None:
if self.data.dim > 50:
self.design = LatinHypercube(data.dim, data.dim + 1)
else:
self.design = SymmetricLatinHypercube(data.dim,
2 * (data.dim + 1))
self.xrange = np.asarray(data.xup - data.xlow)
# algorithm parameters
self.sigma_min = 0.005
self.sigma_max = 0.2
self.sigma_init = 0.2
self.failtol = max(5, data.dim)
self.succtol = 3
self.numeval = 0
self.status = 0
self.sigma = 0
self.resubmitter = RetryStrategy()
self.xbest = None
self.fbest = np.inf
self.fbest_old = None
# Set up search procedures and initialize
self.sampling = sampling_method
if self.sampling is None:
self.sampling = CandidateDYCORS(data)
self.check_input()
# Start with first experimental design
self.sample_initial()
def check_input(self):
self.check_common()
assert not hasattr(self.data, "eval_ineq_constraints"), "Objective function has constraints,\n" \
"SyncStrategyNoConstraints can't handle constraints"
assert not hasattr(self.data, "eval_eq_constraints"), "Objective function has constraints,\n" \
"SyncStrategyNoConstraints can't handle constraints"
def check_common(self):
# Check evaluation budget
if self.extra is None:
assert self.maxeval >= self.design.npts, \
"Experimental design is larger than the evaluation budget"
else:
assert self.maxeval >= self.design.npts + self.extra.shape[0], \
"Experimental design + extra points exceeds the evaluation budget"
# Check dimensionality
assert self.design.dim == self.data.dim, \
"Experimental design and optimization problem have different dimensions"
if self.extra is not None:
assert self.data.dim == self.extra.shape[1], \
"Extra point and optimization problem have different dimensions"
# Check that the optimization problem makes sense
check_opt_prob(self.data)
def proj_fun(self, x):
x = np.atleast_2d(x)
return round_vars(self.data, x)
def log_completion(self, record):
"""Record a completed evaluation to the log.
:param record: Record of the function evaluation
:param penalty: Penalty for the given point
"""
xstr = np.array_str(record.params[0],
max_line_width=np.inf,
precision=5,
suppress_small=True)
if record.feasible:
logger.info("{} {:.3e} @ {}".format("True", record.value, xstr))
else:
logger.info("{} {:.3e} @ {}".format("False", record.value, xstr))
def adjust_step(self):
"""Adjust the sampling radius sigma.
After succtol successful steps, we cut the sampling radius;
after failtol failed steps, we double the sampling radius.
:ivar Fnew: Best function value in new step
:ivar fbest: Previous best function evaluation
"""
# Initialize if this is the first adaptive step
if self.fbest_old is None:
self.fbest_old = self.fbest
return
# Check if we succeeded at significant improvement
if self.fbest < self.fbest_old - 1e-3 * math.fabs(self.fbest_old):
self.status = max(1, self.status + 1)
else:
self.status = min(-1, self.status - 1)
self.fbest_old = self.fbest
# Check if step needs adjusting
if self.status <= -self.failtol:
self.status = 0
self.sigma /= 2
logger.info("Reducing sigma")
if self.status >= self.succtol:
self.status = 0
self.sigma = min([2.0 * self.sigma, self.sigma_max])
logger.info("Increasing sigma")
def sample_initial(self):
"""Generate and queue an initial experimental design.
"""
if self.numeval == 0:
logger.info("=== Start ===")
else:
logger.info("=== Restart ===")
self.fhat.reset()
self.sigma = self.sigma_init
self.status = 0
self.xbest = None
self.fbest_old = None
self.fbest = np.inf
self.fhat.reset()
start_sample = self.design.generate_points()
assert start_sample.shape[1] == self.data.dim, \
"Dimension mismatch between problem and experimental design"
start_sample = from_unit_box(start_sample, self.data)
if self.extra is not None:
start_sample = np.vstack((start_sample, self.extra))
for j in range(min(start_sample.shape[0],
self.maxeval - self.numeval)):
start_sample[j, :] = self.proj_fun(
start_sample[j, :]) # Project onto feasible region
proposal = self.propose_eval(np.copy(start_sample[j, :]))
self.resubmitter.rput(proposal)
self.sampling.init(np.copy(start_sample), self.fhat,
self.maxeval - self.numeval)
def sample_adapt(self):
"""Generate and queue samples from the search strategy
"""
self.adjust_step()
nsamples = min(self.nsamples, self.maxeval - self.numeval)
new_points = self.sampling.make_points(npts=nsamples,
xbest=np.copy(self.xbest),
sigma=self.sigma,
proj_fun=self.proj_fun)
for i in range(nsamples):
proposal = self.propose_eval(np.copy(np.ravel(new_points[i, :])))
self.resubmitter.rput(proposal)
def start_batch(self):
"""Generate and queue a new batch of points
"""
if self.sigma < self.sigma_min:
self.sample_initial()
else:
self.sample_adapt()
def propose_action(self):
"""Propose an action
"""
if self.numeval == self.maxeval:
return self.propose_terminate()
elif self.resubmitter.num_eval_outstanding == 0:
self.start_batch()
return self.resubmitter.get()
def on_complete(self, record):
"""Handle completed function evaluation.
When a function evaluation is completed we need to ask the constraint
handler if the function value should be modified which is the case for
say a penalty method. We also need to print the information to the
logfile, update the best value found so far and notify the GUI that
an evaluation has completed.
:param record: Evaluation record
"""
self.numeval += 1
record.worker_id = self.worker_id
record.worker_numeval = self.numeval
record.feasible = True
self.log_completion(record)
self.fhat.add_point(np.copy(record.params[0]), record.value)
if record.value < self.fbest:
self.xbest = np.copy(record.params[0])
self.fbest = record.value