def gen_sample_from_qei(cpp_gp,
cpp_search_domain,
sgd_params,
num_to_sample,
num_mc=1e4,
lhc_itr=2e4):
"""
:param cpp_gp: trained cpp version of GaussianProcess model
:param cpp_search_domain: cpp version of TensorProductDomain
:param sgd_params: GradientDescentParameters
:param num_to_sample: number of points to sample for the next iteration
:param num_mc: number of Monte Carlo iterations
:param lhc_itr: number of points used in latin hypercube search
:return: (points to sample next, expected improvement at this set of points)
"""
cpp_ei_evaluator = cppExpectedImprovement(gaussian_process=cpp_gp,
num_mc_iterations=int(num_mc))
#python_ei_evaluator = pythonExpectedImprovement(gaussian_process=cpp_gp, num_mc_iterations=int(num_mc))
#optimizer = pythonGradientDescentOptimizer(cpp_search_domain, python_ei_evaluator, sgd_params, int(lhc_itr))
optimizer = cppGradientDescentOptimizer(cpp_search_domain,
cpp_ei_evaluator, sgd_params,
int(lhc_itr))
points_to_sample_list = []
ei_list = []
points_to_sample_list.append(
multistart_expected_improvement_optimization(optimizer,
None,
num_to_sample,
use_gpu=False,
which_gpu=0,
max_num_threads=8))
cpp_ei_evaluator.set_current_point(points_to_sample_list[0])
ei_list.append(cpp_ei_evaluator.compute_expected_improvement())
return points_to_sample_list[numpy.argmax(ei_list)], numpy.amax(ei_list)
def gen_sample_from_qkg_mcmc(cpp_gp_mcmc, cpp_gp_list, inner_optimizer, cpp_search_domain, num_fidelity,
discrete_pts_list, sgd_params, num_to_sample, num_mc=10, lhc_itr=1e3):
"""
:param cpp_gp_mcmc: trained cpp version of GaussianProcess MCMC model
:param cpp_gp_list:
:param inner_optimizer:
:param cpp_search_domain: cpp version of TensorProductDomain
:param num_fidelity: number of fidelity control parameters
:param discrete_pts_list:
:param sgd_params: GradientDescentParameters
:param num_mc: number of Monte Carlo iterations
:param lhc_itr: number of points used in latin hypercube search
:return: (points to sample next, expected improvement at this set of points)
"""
cpp_kg_evaluator = cppKnowledgeGradientMCMC(gaussian_process_mcmc = cpp_gp_mcmc, gaussian_process_list=cpp_gp_list,
num_fidelity = num_fidelity, inner_optimizer = inner_optimizer, discrete_pts_list=discrete_pts_list,
num_to_sample = num_to_sample, num_mc_iterations=int(num_mc))
optimizer = cppGradientDescentOptimizer(cpp_search_domain, cpp_kg_evaluator, sgd_params, int(lhc_itr))
points_to_sample_list = []
kg_list = []
points_to_sample_list.append(multistart_knowledge_gradient_mcmc_optimization(optimizer, inner_optimizer, None, discrete_pts_list,
num_to_sample=num_to_sample,
num_pts=discrete_pts_list[0].shape[0],
max_num_threads=8))
cpp_kg_evaluator.set_current_point(points_to_sample_list[0])
kg_list.append(cpp_kg_evaluator.compute_objective_function())
return points_to_sample_list[numpy.argmax(kg_list)], numpy.amax(kg_list)
def gen_sample_from_qei_mcmc(cpp_gp_mcmc,
cpp_search_domain,
sgd_params,
start_points,
num_to_sample,
num_mc=1e4,
lhc_itr=2e4):
"""
:param cpp_gp_mcmc: trained cpp version of GaussianProcess MCMC model
:param cpp_search_domain: cpp version of TensorProductDomain
:param sgd_params: GradientDescentParameters
:param num_to_sample: number of points to sample for the next iteration
:param num_mc: number of Monte Carlo iterations
:param lhc_itr: number of points used in latin hypercube search
:return: (points to sample next, expected improvement at this set of points)
"""
cpp_ei_evaluator = cppExpectedImprovementMCMC(
gaussian_process_mcmc=cpp_gp_mcmc,
num_to_sample=num_to_sample,
num_mc_iterations=int(num_mc))
optimizer = cppGradientDescentOptimizer(cpp_search_domain,
cpp_ei_evaluator, sgd_params,
int(lhc_itr))
points_to_sample_list = []
ei_list = []
points_to_sample_list.append(
multistart_expected_improvement_mcmc_optimization(
ei_optimizer=optimizer,
num_multistarts=start_points.shape[0],
start_point_set=start_points,
num_to_sample=1,
max_num_threads=1))
cpp_ei_evaluator.set_current_point(points_to_sample_list[0])
ei_list.append(cpp_ei_evaluator.compute_objective_function())
return points_to_sample_list[numpy.argmax(ei_list)], numpy.amax(ei_list)
)
test = np.zeros(eval_pts.shape[0])
ps_evaluator = PosteriorMean(cpp_gp, num_fidelity)
for i, pt in enumerate(eval_pts):
ps_evaluator.set_current_point(
pt.reshape(
(1, cpp_gp_loglikelihood.dim - objective_func._num_fidelity)
)
)
test[i] = -ps_evaluator.compute_objective_function()
initial_point = eval_pts[np.argmin(test)]
ps_sgd_optimizer = cppGradientDescentOptimizer(
cpp_inner_search_domain, ps_evaluator, cpp_sgd_params_ps
)
report_point = posterior_mean_optimization(
ps_sgd_optimizer, initial_guess=initial_point, max_num_threads=4
)
ps_evaluator.set_current_point(
report_point.reshape(
(1, cpp_gp_loglikelihood.dim - objective_func._num_fidelity)
)
)
if -ps_evaluator.compute_objective_function() > np.min(test):
report_point = initial_point
discrete_pts_optima = np.reshape(
np.append(discrete_pts_optima, report_point),
hyper_domain = np.zeros((2 + dim + len(derivatives), 2))
hyper_domain[:(2 + dim), :] = objective_func._hyper_domain[:(2 + dim), :]
for g in range(len(derivatives)):
hyper_domain[g + 2 +
dim, :] = objective_func._hyper_domain[derivatives[g] + 2 +
dim, :]
hyper_search_domain = cppTensorProductDomain(
[ClosedInterval(bound[0], bound[1]) for bound in hyper_domain])
hyper_params = np.ones(2 + dim + len(derivatives))
cpp_gp = cppGaussianProcess(cpp_cov, hyper_params[(1 + objective_func._dim):],
init_data, derivatives)
cpp_gp_loglikelihood = cppGaussianProcessLogLikelihood(
cpp_cov, cpp_gp.get_historical_data_copy(), noise_variance, derivatives)
sgd_optimizer = cppGradientDescentOptimizer(hyper_search_domain,
cpp_gp_loglikelihood,
cpp_sgd_params_hyper)
hyper_params = multistart_hyperparameter_optimization(
log_likelihood_optimizer=sgd_optimizer,
num_multistarts=None,
max_num_threads=8)
discrete_pts_optima = python_search_domain.generate_uniform_random_points_in_domain(
1000)
points = sample_from_global_optima(cpp_gp, 1000, objective_func._search_domain,
discrete_pts_optima, 200)
print points
print cpp_gp.compute_mean_of_points(points)
