class MyEvComp(Component):
doit = Event(desc='Do It!')
doit2 = Event(desc='Do It Again!')
doit_count = Int(0, iotype='out')
doit2_count = Int(0, iotype='out')
some_int = Int(0, iotype='in')
def _doit_fired(self):
self.doit_count += 1
def _doit2_fired(self):
self.doit2_count += 1
def execute(self):
pass
class DrivenComponent(Component):
""" Just something to be driven and compute results. """
x0 = Float(1., iotype='in')
y0 = Float(1., iotype='in') # used just to get ParameterGroup
x1 = Float(1., iotype='in')
x2 = Float(1., iotype='in')
x3 = Float(1., iotype='in')
err_event = Event()
stop_exec = Bool(False, iotype='in')
rosen_suzuki = Float(0., iotype='out')
def __init__(self):
super(DrivenComponent, self).__init__()
self._raise_err = False
def _err_event_fired(self):
self._raise_err = True
def execute(self):
""" Compute results from input vector. """
self.rosen_suzuki = rosen_suzuki(self.x0, self.x1, self.x2, self.x3)
if self._raise_err:
self.raise_exception('Forced error', RuntimeError)
if self.stop_exec:
self.parent.driver.stop() # Only valid if sequential!
class Pareto_Min_Dist(Component):
"""Computes the probability that any given point from the primary concept
will interesect the pareto frontiers of some other concepts.
"""
pareto = List([], iotype="in",
desc="List of CaseIterators containing competing local Pareto points")
criteria = ListStr(iotype="in",dtype="str",
desc="Names of responses to maximize expected improvement around. "
"Must be NormalDistribution type.")
predicted_values = Array(iotype="in",dtype=NormalDistribution,
desc="CaseIterator which contains a NormalDistribution "
"for each response at a location where you wish to "
"calculate EI.")
dist = Float(0.0, iotype="out",
desc="minimum distance from a point to other pareto set ")
reset_pareto = Event()
def __init__(self):
super(Pareto_Min_Dist, self).__init__()
self.y_star_other = None
def _reset_pareto_fired(self):
self.y_star_other = None
def get_pareto(self):
y_star_other = []
c = []
for single_case_list in self.pareto:
for case in single_case_list:
for objective in case.outputs:
for crit in self.criteria:
if crit in objective[0]:
#TODO: criteria needs at least two things matching
#objective names in CaseIterator outputs, error otherwise
c.append(objective[2])
if c != [] :
y_star_other.append(c)
c = []
return y_star_other
def _calc_min_dist(self,p,y_star_other):
"""Computes the minimum distance from a candidate point
to other_pareto.
"""
dists = []
for y in y_star_other:
d = sqrt(sum([(A-B)**2 for A,B in zip(p,y)]))
dists.append(d)
return min(dists)
def execute(self):
mu = [objective.mu for objective in self.predicted_values]
if self.y_star_other == None:
self.y_star_other = self.get_pareto()
self.dist = self._calc_min_dist(mu,self.y_star_other)
class MetaModelBase(Component):
"""
Base class for functionality of a meta model.
Should be subclassed.
"""
# pylint: disable-msg=E1101
model = Slot(IComponent, allow_none=True,
desc='Slot for the Component or Assembly being '
'encapsulated.')
includes = List(Str, iotype='in',
desc='A list of names of variables to be included '
'in the public interface.')
excludes = List(Str, iotype='in',
desc='A list of names of variables to be excluded '
'from the public interface.')
default_surrogate = Slot(ISurrogate, allow_none=True,
desc="This surrogate will be used for all "
"outputs that don't have a specific surrogate "
"assigned to them in their sur_<name> slot.")
surrogates = Dict(key_trait=Str,
value_trait=Slot(ISurrogate),
desc='surrogates for output variables')
report_errors = Bool(True, iotype="in",
desc="If True, metamodel will report errors reported "
"from the component. If False, metamodel will swallow "
"the errors but log that they happened and "
"exclude the case from the training set.")
recorder = Slot(ICaseRecorder,
desc='Records training cases')
# when fired, the next execution will train the metamodel
train_next = Event(desc='Train metamodel on next execution')
#when fired, the next execution will reset all training data
reset_training_data = Event(desc='Reset training data on next execution')
def __init__(self):
super(MetaModelBase, self).__init__()
self._surrogate_input_names = None
self._surrogate_output_names = None
self._surrogate_overrides = set() # keeps track of which sur_<name> slots are full
self._training_data = {}
self._training_input_history = []
self._const_inputs = {} # dict of constant training inputs indices and their values
self._train = False
self._new_train_data = False
self._failed_training_msgs = []
self._default_surrogate_copies = {} # need to maintain separate copy of
# default surrogate for each sur_*
# that doesn't have a surrogate
# defined
# the following line will work for classes that inherit from MetaModel
# as long as they declare their traits in the class body and not in
# the __init__ function. If they need to create traits dynamically
# during initialization they'll have to provide the value of
# _mm_class_traitnames
self._mm_class_traitnames = set(self.traits(iotype=not_none).keys())
self.on_trait_change(self._surrogate_updated, "surrogates_items")
def _train_next_fired(self):
self._train = True
self._new_train_data = True
def _reset_training_data_fired(self):
self._training_input_history = []
self._const_inputs = {}
self._failed_training_msgs = []
# remove output history from training_data
for name in self._training_data:
self._training_data[name] = []
def _warm_start_data_changed(self, oldval, newval):
self.reset_training_data = True
# build list of inputs
for case in newval:
if self.recorder:
self.recorder.record(case)
inputs = []
for inp_name in self.surrogate_input_names():
var_name = '.'.join([self.name, inp_name])
try:
inp_val = case[var_name]
except KeyError:
pass
#self.raise_exception('The variable "%s" was not '
#'found as an input in one of the cases provided '
#'for warm_start_data.' % var_name, ValueError)
else:
if inp_val is not None:
inputs.append(inp_val)
self._training_input_history.append(inputs)
for output_name in self.surrogate_output_names():
#grab value from case data
var_name = '.'.join([self.name, output_name])
try:
val = case.get_output(var_name)
except KeyError:
self.raise_exception('The output "%s" was not found '
'in one of the cases provided for '
'warm_start_data' % var_name, ValueError)
else: # save to training output history
self._training_data[output_name].append(val)
self._new_train_data = True
def child_run_finished(self, childname, outs=None):
pass
def check_config(self):
'''Called as part of pre_execute.'''
# 1. model must be set
if self.model is None:
self.raise_exception("MetaModel object must have a model!",
RuntimeError)
# 2. can't have both includes and excludes
if self.excludes and self.includes:
self.raise_exception("includes and excludes are mutually exclusive",
RuntimeError)
# 3. the includes and excludes must match actual inputs and outputs of the model
input_names = self.surrogate_input_names()
output_names = self.surrogate_output_names()
input_and_output_names = input_names + output_names
for include in self.includes:
if include not in input_and_output_names:
self.raise_exception('The include "%s" is not one of the '
'model inputs or outputs ' % include, ValueError)
for exclude in self.excludes:
if exclude not in input_and_output_names:
self.raise_exception('The exclude "%s" is not one of the '
'model inputs or outputs ' % exclude, ValueError)
# 4. Either there are no surrogates set and no default surrogate
# ( just do passthrough )
# or
# all outputs must have surrogates assigned either explicitly
# or through the default surrogate
if self.default_surrogate is None:
no_sur = []
for name in self.surrogate_output_names():
if not self.surrogates[name]:
no_sur.append(name)
if len(no_sur) > 0 and len(no_sur) != len(self._surrogate_output_names):
self.raise_exception("No default surrogate model is defined and"
" the following outputs do not have a"
" surrogate model: %s. Either specify"
" default_surrogate, or specify a"
" surrogate model for all outputs." %
no_sur, RuntimeError)
# 5. All the explicitly set surrogates[] should match actual outputs of the model
for surrogate_name in self.surrogates.keys():
if surrogate_name not in output_names:
self.raise_exception('The surrogate "%s" does not match one of the '
'model outputs ' % surrogate_name, ValueError)
def execute(self):
"""If the training flag is set, train the metamodel. Otherwise,
predict outputs.
"""
if self._train:
try:
inputs = self.update_model_inputs()
self.model.run(force=True)
except Exception as err:
if self.report_errors:
raise err
else:
self._failed_training_msgs.append(str(err))
else: # if no exceptions are generated, save the data
self._training_input_history.append(inputs)
self.update_outputs_from_model()
case_outputs = []
for name, output_history in self._training_data.items():
case_outputs.append(('.'.join([self.name, name]),
output_history[-1]))
# save the case, making sure to add out name to the local input
# name since this Case is scoped to our parent Assembly
case_inputs = [('.'.join([self.name, name]), val)
for name, val in zip(self.surrogate_input_names(),
inputs)]
if self.recorder:
self.recorder.record(Case(inputs=case_inputs,
outputs=case_outputs))
self._train = False
else:
# NO surrogates defined. just run model and get outputs
if self.default_surrogate is None and not self._surrogate_overrides:
inputs = self.update_model_inputs()
self.model.run()
self.update_outputs_from_model()
return
if self._new_train_data:
if len(self._training_input_history) < 2:
self.raise_exception("ERROR: need at least 2 training points!",
RuntimeError)
# figure out if we have any constant training inputs
tcases = self._training_input_history
in_hist = tcases[0][:]
# start off assuming every input is constant
idxlist = range(len(in_hist))
self._const_inputs = dict(zip(idxlist, in_hist))
for i in idxlist:
val = in_hist[i]
for case in range(1, len(tcases)):
if val != tcases[case][i]:
del self._const_inputs[i]
break
if len(self._const_inputs) == len(in_hist):
self.raise_exception("ERROR: all training inputs are constant.")
elif len(self._const_inputs) > 0:
# some inputs are constant, so we have to remove them from the training set
training_input_history = []
for inputs in self._training_input_history:
training_input_history.append([val for i, val in enumerate(inputs)
if i not in self._const_inputs])
else:
training_input_history = self._training_input_history
for name, output_history in self._training_data.items():
surrogate = self._get_surrogate(name)
if surrogate is not None:
surrogate.train(training_input_history, output_history)
self._new_train_data = False
inputs = []
for i, name in enumerate(self.surrogate_input_names()):
val = self.get(name)
cval = self._const_inputs.get(i, _missing)
if cval is _missing:
inputs.append(val)
elif val != cval:
self.raise_exception("ERROR: training input '%s' was a"
" constant value of (%s) but the value"
" has changed to (%s)." %
(name, cval, val), ValueError)
for name in self._training_data:
surrogate = self._get_surrogate(name)
# copy output to boundary
if surrogate is None:
self._set_output(name, self.model.get(name))
else:
self._set_output(name, surrogate.predict(inputs))
class MultiObjExpectedImprovementBase(Component):
criteria = Array(
iotype="in",
desc="Names of responses to maximize expected improvement around. \
Must be NormalDistribution type.")
predicted_values = Array(
[0, 0],
iotype="in",
dtype=NormalDistribution,
desc="CaseIterator which contains NormalDistributions for each \
response at a location where you wish to calculate EI."
)
n = Int(1000,
iotype="in",
desc="Number of Monte Carlo Samples with \
which to calculate probability of improvement.")
calc_switch = Enum("PI", ["PI", "EI"],
iotype="in",
desc="Switch to use either \
probability (PI) or expected (EI) improvement.")
PI = Float(0.0,
iotype="out",
desc="The probability of improvement of the next_case.")
EI = Float(0.0,
iotype="out",
desc="The expected improvement of the next_case.")
reset_y_star = Event(desc='Reset Y* on next execution')
def __init__(self):
super(MultiObjExpectedImprovementBase, self).__init__()
self.y_star = None
def _reset_y_star_fired(self):
self.y_star = None
def get_y_star(self):
criteria_count = len(self.criteria)
flat_crit = self.criteria.ravel()
try:
y_star = zip(*[self.best_cases[crit] for crit in self.criteria])
except KeyError:
self.raise_exception(
'no cases in the provided case_set had output '
'matching the provided criteria, %s' % self.criteria,
ValueError)
#sort list on first objective
y_star = array(y_star)[array([i[0] for i in y_star]).argsort()]
return y_star
def _2obj_PI(self, mu, sigma):
"""Calculates the multi-objective probability of improvement
for a new point with two responses. Takes as input a
pareto frontier, mean and sigma of new point."""
y_star = self.y_star
PI1 = (0.5 + 0.5 * erf(
(1 / (2**0.5)) * ((y_star[0][0] - mu[0]) / sigma[0])))
PI3 = (1-(0.5+0.5*erf((1/(2**0.5))*((y_star[-1][0]-mu[0])/sigma[0]))))\
*(0.5+0.5*erf((1/(2**0.5))*((y_star[-1][1]-mu[1])/sigma[1])))
PI2 = 0
if len(y_star) > 1:
for i in range(len(y_star) - 1):
PI2=PI2+((0.5+0.5*erf((1/(2**0.5))*((y_star[i+1][0]-mu[0])/sigma[0])))\
-(0.5+0.5*erf((1/(2**0.5))*((y_star[i][0]-mu[0])/sigma[0]))))\
*(0.5+0.5*erf((1/(2**0.5))*((y_star[i+1][1]-mu[1])/sigma[1])))
mcpi = PI1 + PI2 + PI3
return mcpi
def _2obj_EI(self, mu, sigma):
"""Calculates the multi-criteria expected improvement
for a new point with two responses. Takes as input a
pareto frontier, mean and sigma of new point."""
y_star = self.y_star
ybar11 = mu[0]*(0.5+0.5*erf((1/(2**0.5))*((y_star[0][0]-mu[0])/sigma[0])))\
-sigma[0]*(1/((2*pi)**0.5))*exp(-0.5*((y_star[0][0]-mu[0])**2/sigma[0]**2))
ybar13 = (mu[0]*(0.5+0.5*erf((1/(2**0.5))*((y_star[-1][0]-mu[0])/sigma[0])))\
-sigma[0]*(1/((2*pi)**0.5))*exp(-0.5*((y_star[-1][0]-mu[0])**2/sigma[0]**2)))\
*(0.5+0.5*erf((1/(2**0.5))*((y_star[-1][1]-mu[1])/sigma[1])))
ybar12 = 0
if len(y_star) > 1:
for i in range(len(y_star) - 1):
ybar12 = ybar12+((mu[0]*(0.5+0.5*erf((1/(2**0.5))*((y_star[i+1][0]-mu[0])/sigma[0])))\
-sigma[0]*(1/((2*pi)**0.5))*exp(-0.5*((y_star[i+1][0]-mu[0])**2/sigma[0]**2)))\
-(mu[0]*(0.5+0.5*erf((1/(2**0.5))*((y_star[i][0]-mu[0])/sigma[0])))\
-sigma[0]*(1/((2*pi)**0.5))*exp(-0.5*((y_star[i][0]-mu[0])**2/sigma[0]**2))))\
*(0.5+0.5*erf((1/(2**0.5))*((y_star[i+1][1]-mu[1])/sigma[1])))
ybar1 = (ybar11 + ybar12 + ybar13) / self.PI
ybar21 = mu[1]*(0.5+0.5*erf((1/(2**0.5))*((y_star[0][1]-mu[1])/sigma[1])))\
-sigma[1]*(1/((2*pi)**0.5))*exp(-0.5*((y_star[0][1]-mu[1])**2/sigma[1]**2))
ybar23 = (mu[1]*(0.5+0.5*erf((1/(2**0.5))*((y_star[-1][1]-mu[1])/sigma[1])))\
-sigma[1]*(1/((2*pi)**0.5))*exp(-0.5*((y_star[-1][1]-mu[1])**2/sigma[1]**2)))\
*(0.5+0.5*erf((1/(2**0.5))*((y_star[-1][0]-mu[0])/sigma[0])))
ybar22 = 0
if len(y_star) > 1:
for i in range(len(y_star) - 1):
ybar22 = ybar22+((mu[1]*(0.5+0.5*erf((1/(2**0.5))*((y_star[i+1][1]-mu[1])/sigma[1])))\
-sigma[1]*(1/((2*pi)**0.5))*exp(-0.5*((y_star[i+1][1]-mu[1])**2/sigma[1]**2)))\
-(mu[1]*(0.5+0.5*erf((1/(2**0.5))*((y_star[i][1]-mu[1])/sigma[1])))\
-sigma[1]*(1/((2*pi)**0.5))*exp(-0.5*((y_star[i][1]-mu[1])**2/sigma[1]**2))))\
*(0.5+0.5*erf((1/(2**0.5))*((y_star[i+1][0]-mu[0])/sigma[0])))
ybar2 = (ybar21 + ybar22 + ybar23) / self.PI
dists = [((ybar1 - point[0])**2 + (ybar2 - point[1])**2)**0.5
for point in y_star]
mcei = self.PI * min(dists)
if isnan(mcei):
mcei = 0
return mcei
def _dom(self, a, b):
"""determines if a completely dominates b
returns True is if does
"""
comp = [c1 < c2 for c1, c2 in zip(a, b)]
if sum(comp) == len(self.criteria):
return True
return False
def _nobj_PI(self, mu, sigma):
cov = diag(array(sigma)**2)
rands = random.multivariate_normal(mu, cov, self.n)
num = 0 # number of cases that dominate the current Pareto set
for random_sample in rands:
for par_point in self.y_star:
#par_point = [p[2] for p in par_point.outputs]
if self._dom(par_point, random_sample):
num = num + 1
break
pi = (self.n - num) / float(self.n)
return pi
def execute(self):
""" Calculates the expected improvement or
probability of improvement of a candidate
point given by a normal distribution.
"""
mu = [objective.mu for objective in self.predicted_values]
sig = [objective.sigma for objective in self.predicted_values]
if self.y_star == None:
self.y_star = self.get_y_star()
n_objs = len(self.criteria)
if n_objs == 2:
"""biobjective optimization"""
self.PI = self._2obj_PI(mu, sig)
if self.calc_switch == 'EI':
"""execute EI calculations"""
self.EI = self._2obj_EI(mu, sig)
if n_objs > 2:
"""n objective optimization"""
self.PI = self._nobj_PI(mu, sig)
if self.calc_switch == 'EI':
"""execute EI calculations"""
self.raise_exception(
"EI calculations not supported"
" for more than 2 objectives", ValueError)