def model(self):
super(ProbabilisticScheduler, self).model()
overall_best_score = self._overall_best_score()
additional_parameters = self._additional_acquisition_function_parameters()
# Calculate acquisition functions by averaging over models
# If hyperparameters were optimised, there is only one model over with
# we average here. This means we don't have to treat sampling and
# optimisation separately
self.acquisition_functions = dict()
for algorithm, models_for_algorithm in self.models.items():
acquisition_functions = []
for score_model, time_model in zip(models_for_algorithm['score'], models_for_algorithm['time']):
# Call a with all available data
acquisition_function = self.__class__.a(self, overall_best_score, score_model.mean, score_model.std_dev, time_model.mean, time_model.std_dev, **additional_parameters)
acquisition_functions.append(acquisition_function)
avg_acquisition_function = (np.linspace(0,1,100), np.mean(acquisition_functions, axis=0))
if self._truncate_a():
max_x = 0
for datum in self.data[algorithm]:
max_x = max(datum.x, max_x)
xs_org, ys_org = avg_acquisition_function
xs, ys = truncate_func_at_x(max_x, xs_org, ys_org)
avg_acquisition_function = (xs, ys)
else:
xs_org, ys_org = avg_acquisition_function
xs, ys = truncate_func_at_x(0.01, xs_org, ys_org)
avg_acquisition_function = (xs, ys)
self.acquisition_functions[algorithm] = avg_acquisition_function
def model(self):
if self.decision == "SKIP":
print "Skip modelling"
else:
super(MinimiseUncertaintyAndExploitScheduler, self).model()
for algorithm, acquisition_function in self.acquisition_functions.items():
xs_org, ys_org = acquisition_function
# Truncate at the end of the burn in phase
xs, ys = truncate_func_at_x(self.burn_in_end, xs_org, ys_org)
acquisition_function = (xs, ys)
self.acquisition_functions[algorithm] = acquisition_function
