def suggest_sample(self, number_of_samples=1):
"""
Returns a suggested next point to evaluate.
"""
# Get current evaluations
X_evaluated, Y_evaluated = self.model.get_XY()
self.X_aux = X_evaluated
self.Y_aux = Y_evaluated
# Auxiliary Bayesian optimization model to run ParEGO
weight = np.random.dirichlet(
np.ones(len(self.Y_aux))
) # self._sample_posterior_weight_for_chebyshev_scalarization()
self.Y_aux = np.reshape(self.Y_aux,
(len(self.Y_aux), self.Y_aux[0].shape[0]))
scalarized_fX = chebyshev_scalarization(self.Y_aux, weight)
scalarized_fX = np.reshape(scalarized_fX, (len(scalarized_fX), 1))
aux_model = MultiOutputGP(output_dim=1,
exact_feval=[True],
fixed_hyps=False)
def aux_utility_func(y, parameter):
return np.dot(parameter, y)
def aux_utility_gradient(y, parameter):
return parameter
aux_utility_parameter_support = np.ones((1, 1))
aux_utility_parameter_prob_distribution = np.ones((1, ))
aux_utility_param_distribution = UtilityDistribution(
support=aux_utility_parameter_support,
prob_dist=aux_utility_parameter_prob_distribution,
utility_func=aux_utility_func)
aux_utility = Utility(
func=aux_utility_func,
gradient=aux_utility_gradient,
parameter_distribution=aux_utility_param_distribution,
affine=True)
aux_acquisition_optimizer = AcquisitionOptimizer(
space=self.space,
model=aux_model,
utility=aux_utility,
optimizer=self.optimizer,
include_baseline_points=True)
aux_acquisition = uEI_affine(aux_model,
self.space,
optimizer=aux_acquisition_optimizer,
utility=aux_utility)
aux_evaluator = GPyOpt.core.evaluators.Sequential(aux_acquisition)
aux_sampling_policy = AcquisitionFunction(aux_model, self.space,
aux_acquisition,
aux_evaluator)
bopl = BOPL(aux_model,
self.space,
sampling_policy=aux_sampling_policy,
utility=aux_utility,
X_init=self.X_aux,
Y_init=[scalarized_fX],
dynamic_utility_parameter_distribution=False)
suggested_sample = bopl.suggest_next_points_to_evaluate()
return suggested_sample
# True underlying utility
true_underlying_utility_parameter = prior_sample_generator(
1, experiment_number)[0]
print('True underlying utility parameter: {}'.format(
true_underlying_utility_parameter))
def true_underlying_utility_func(y):
return utility_func(y, true_underlying_utility_parameter)
# Run full optimization loop
bopl = BOPL(model,
space,
attributes,
sampling_policy,
utility,
initial_design,
true_underlying_utility_func=true_underlying_utility_func,
dynamic_utility_parameter_distribution=
dynamic_utility_parameter_distribution)
bopl.run_optimization(max_iter=max_iter,
filename=filename,
report_evaluated_designs_only=True,
utility_distribution_update_interval=1,
compute_true_underlying_optimal_value=True,
compute_integrated_optimal_values=False,
compute_true_integrated_optimal_value=False)
else:
for i in range(1):
experiment_number = i
filename = [