class Metamodel():
"""
The Metamodel has four significant elements, the model, the surrogate,
the relevator and the history of evaluations.
At every step, the metamodel uses the relevator to decide wheter to use the
model or the surrogate and afterwards updates the history and its
components.
All methods and attributes that start with '_' should be treated as
private.
"""
def __init__(self, metamodel_kwargs,
model_kwargs, surrogate_kwargs,
relevator_kwargs, history_kwargs):
""" Initialize the Metamodel with the specified components. """
self.step_index = 0
self.model_evaluations = 0
self.model = Model(model_kwargs)
self.surrogate = Surrogate(self, surrogate_kwargs)
self.relevator = Relevator(self, relevator_kwargs)
self.history = History(self, history_kwargs)
self._random_seed = metamodel_kwargs.get("random_seed", None)
if self._random_seed is not None:
self.model.set_random_seed(self._random_seed)
self.surrogate.set_random_seed(self._random_seed)
self.relevator.set_random_seed(self._random_seed)
return
def evaluate(self, coords):
""" Evaluate the Metamodel at the given point. """
# Determine relevance.
relevance = self.relevator.evaluate(coords)
is_relevant = self.relevator.is_relevant(relevance)
# Decide which prediction to use.
if is_relevant:
prediction = self.model.evaluate(coords)
self.model_evaluations += 1
else:
prediction = self.surrogate.evaluate(coords)
# Update history and components.
self.history.update(coords, prediction, relevance, is_relevant)
# Update index and return result.
self.step_index += 1
return prediction
def get_toolbox(self, xtr, phi_failure, phi_parallel, global_failed_indices, skwargs, cfunc=None, \
cargs=(), ckwargs={}, verbose=True):
'''
Modify parent's toolbox method for single objective optimisation.
Parameters.
----------
xtr (np.array): training decision vectors.
skwargs (dict): keyword arguments for infill criterion; not used here.
cfunc (function): cheap constraint function.
cargs (tuple): arguments of cheap constraint function.
verbose (bool): whether to print verbose comments.
Returns a DEAP toolbox.
'''
ytr, xtr, xinds, local_failed_indices, success = self.scalarise(xtr, kwargs=skwargs)
self.current_hv = self.current_hpv()
surr = Surrogate(xtr, ytr, phi_failure, phi_parallel, local_failed_indices, self.kernel.copy(), verbose=verbose)
self.surr = surr
self.surr.success = success
#self.surr = phi_old
#for i in range (len(local_failed_indices)):
# self.surr.global_failed_indices.append(local_failed_indices[i])
print("surr.global_failed_indices", self.surr.global_failed_indices)
print("local_failed_indices", local_failed_indices)
#self.surr.global_failed_indices = np.concatenate([self.surr.global_failed_indices, local_failed_indices])
self.surr.xinds = xinds
########################
# Get bounds (APR added)
########################
lb=self.lower_bounds
ub=self.upper_bounds
xmean = self.surr.xbar
xsd = self.surr.xstd
ysd = self.surr.ystd
bounds = self.get_bounds((lb-xmean)/xsd, (ub-xmean)/xsd)
##################################
# Compute the gradient (APR added)
##################################
L = 1.0*(self.max_predictive_gradient(surr, bounds)) * (ysd / xsd)
self.surr.L = L
######################################################################
# This represents the best value from the surrogate model max(mean(x))
# It is not the rough approximation max_i{y_i} (APR added)
######################################################################
maxfevals = 20000
tx = np.linspace(lb, ub, maxfevals)[:,None]
pred_y, pred_s = self.surr.predict(tx)
Min = min(pred_y)
self.surr.best = Min
return self.init_deap(surr.penalized_acquisition, obj_sense=self.obj_sense[0], \
cfunc=cfunc, cargs=cargs, ckwargs=ckwargs, \
lb=self.lower_bounds, ub=self.upper_bounds)
def __init__(self, metamodel_kwargs,
model_kwargs, surrogate_kwargs,
relevator_kwargs, history_kwargs):
""" Initialize the Metamodel with the specified components. """
self.step_index = 0
self.model_evaluations = 0
self.model = Model(model_kwargs)
self.surrogate = Surrogate(self, surrogate_kwargs)
self.relevator = Relevator(self, relevator_kwargs)
self.history = History(self, history_kwargs)
self._random_seed = metamodel_kwargs.get("random_seed", None)
if self._random_seed is not None:
self.model.set_random_seed(self._random_seed)
self.surrogate.set_random_seed(self._random_seed)
self.relevator.set_random_seed(self._random_seed)
return
def get_toolbox(self, xtr, skwargs, cfunc=None, \
cargs=(), ckwargs={}, verbose=True):
"""
Generate a DEAP toolbox for the infill criterion optimiser.
Parameters.
-----------
xtr (np.array): traing decision vectors.
skwargs (dict): options for infill criterion calculation; varies with
technique.
cfunc (function): cheap constraint function.
cargs (tuple): argumetns for constraint function.
ckwargs (dict): keyword arguments for constraint function.
verbose (bool): whether to print more comments.
Returns a DEAP toolbox.
"""
ytr = self.scalarise(xtr, kwargs=skwargs)
self.current_hv = self.current_hpv()
surr = Surrogate(xtr, ytr, self.kernel.copy(), verbose=verbose)
#########################################
# Add surrogate to base class (APR added)
#########################################
self.surr = surr
########################
# Get bounds (APR added)
########################
lb = self.lower_bounds
ub = self.upper_bounds
xmean = self.surr.xbar
xsd = self.surr.xstd
ysd = self.surr.ystd
bounds = self.get_bounds((lb - xmean) / xsd, (ub - xmean) / xsd)
##################################
# Compute the gradient (APR added)
##################################
L = (self.max_predictive_gradient(surr, bounds)) * (ysd / xsd)
self.surr.L = L
#######################################################
# Set best based on minimum observed so far (APR added)
#######################################################
self.surr.best = min(ytr)
return self.init_deap(surr.penalized_acquisition, obj_sense=self.obj_sense[0], cfunc=cfunc, cargs=cargs, \
ckwargs=ckwargs, lb=self.lower_bounds, ub=self.upper_bounds)
def get_toolbox(self, xtr, skwargs, cfunc=None, \
cargs=(), ckwargs={}, verbose=True):
"""
Generate a DEAP toolbox for the infill criterion optimiser.
Parameters.
-----------
xtr (np.array): traing decision vectors.
skwargs (dict): options for infill criterion calculation; varies with
technique.
cfunc (function): cheap constraint function.
cargs (tuple): argumetns for constraint function.
ckwargs (dict): keyword arguments for constraint function.
verbose (bool): whether to print more comments.
Returns a DEAP toolbox.
"""
ytr = self.scalarise(xtr, kwargs=skwargs)
self.current_hv = self.current_hpv()
surr = Surrogate(xtr, ytr, self.kernel.copy(), verbose=verbose)
return self.init_deap(surr.expected_improvement, obj_sense=1, cfunc=cfunc, cargs=cargs, \
ckwargs=ckwargs, lb=self.lower_bounds, ub=self.upper_bounds)
def get_toolbox(self, xtr, skwargs, cfunc=None, \
cargs=(), ckwargs={}, verbose=True):
'''
Modify parent's toolbox method for single objective optimisation.
Parameters.
----------
xtr (np.array): training decision vectors.
skwargs (dict): keyword arguments for infill criterion; not used here.
cfunc (function): cheap constraint function.
cargs (tuple): arguments of cheap constraint function.
verbose (bool): whether to print verbose comments.
Returns a DEAP toolbox.
'''
ytr = self.scalarise(xtr, kwargs=skwargs)
self.current_hv = self.current_hpv()
surr = Surrogate(xtr, ytr, self.kernel.copy(), verbose=verbose)
self.surr = surr
return self.init_deap(surr.expected_improvement, obj_sense=self.obj_sense[0], \
cfunc=cfunc, cargs=cargs, ckwargs=ckwargs, \
lb=self.lower_bounds, ub=self.upper_bounds)