class BO(object):
"""
Runner of the multi-attribute Bayesian optimization loop. This class wraps the optimization loop around the different handlers.
:param model: GPyOpt model class.
:param space: GPyOpt space class.
:param objective: GPyOpt objective class.
:param acquisition: GPyOpt acquisition class.
:param evaluator: GPyOpt evaluator class.
:param X_init: 2d numpy array containing the initial inputs (one per row) of the model.
:param Y_init: 2d numpy array containing the initial outputs (one per row) of the model.
:param cost: GPyOpt cost class (default, none).
:param normalize_Y: whether to normalize the outputs before performing any optimization (default, True).
:param model_update_interval: interval of collected observations after which the model is updated (default, 1).
:param de_duplication: GPyOpt DuplicateManager class. Avoids re-evaluating the objective at previous, pending or infeasible locations (default, False).
"""
def __init__(self,
model,
model_c,
space,
objective,
constraint,
acquisition,
evaluator,
X_init,
ref_point=None,
expensive=False,
Y_init=None,
C_init=None,
cost=None,
normalize_Y=False,
model_update_interval=1,
deterministic=True,
true_preference=0.5):
self.true_preference = true_preference
self.model_c = model_c
self.model = model
self.space = space
self.objective = objective
self.constraint = constraint
self.acquisition = acquisition
self.utility = acquisition.utility
self.evaluator = evaluator
self.normalize_Y = normalize_Y
self.model_update_interval = model_update_interval
self.X = X_init
self.Y = Y_init
self.C = C_init
self.ref_point = ref_point
self.deterministic = deterministic
self.cost = CostModel(cost)
self.model_parameters_iterations = None
self.expensive = expensive
try:
if acquisition.name == "Constrained_Thompson_Sampling":
self.sample_from_acq = True
self.tag_last_evaluation = False
else:
self.sample_from_acq = False
self.tag_last_evaluation = True
except:
print("name of acquisition function wasnt provided")
self.sample_from_acq = False
self.tag_last_evaluation = False
def suggest_next_locations(self,
context=None,
pending_X=None,
ignored_X=None):
"""
Run a single optimization step and return the next locations to evaluate the objective.
Number of suggested locations equals to batch_size.
:param context: fixes specified variables to a particular context (values) for the optimization run (default, None).
:param pending_X: matrix of input configurations that are in a pending state (i.e., do not have an evaluation yet) (default, None).
:param ignored_X: matrix of input configurations that the user black-lists, i.e., those configurations will not be suggested again (default, None).
"""
self.model_parameters_iterations = None
self.num_acquisitions = 0
self.context = context
self._update_model(self.normalization_type)
suggested_locations = self._compute_next_evaluations(
pending_zipped_X=pending_X, ignored_zipped_X=ignored_X)
return suggested_locations
# def _value_so_far(self):
# """
# Computes E_n[U(f(x_max))|f], where U is the utility function, f is the true underlying ojective function and x_max = argmax E_n[U(f(x))|U]. See
# function _marginal_max_value_so_far below.
# """
#
# output = 0
# support = self.utility.parameter_dist.support
# utility_dist = self.utility.parameter_dist.prob_dist
#
# a = np.reshape(self.objective.evaluate(self._marginal_max_value_so_far())[0],(self.objective.output_dim,))
#
# output += self.utility.eval_func(support[i],a)*utility_dist[i]
# #print(output)
# return output
#
#
# def _marginal_max_value_so_far(self):
# """
# Computes argmax E_n[U(f(x))|U] (The abuse of notation can be misleading; note that the expectation is with
# respect to the posterior distribution on f after n evaluations)
# """
#
# def val_func(X):
# X = np.atleast_2d(X)
# muX = self.model.posterior_mean(X)
# return muX
#
# def val_func_with_gradient(X):
# X = np.atleast_2d(X)
# muX = self.model.posterior_mean(X)
# dmu_dX = self.model.posterior_mean_gradient(X)
# valX = np.reshape( muX, (X.shape[0],1))
# dval_dX = dmu_dX
# return -valX, -dval_dX
#
#
# argmax = self.acquisition.optimizer.optimize_inner_func(f=val_func, f_df=val_func_with_gradient)[0]
# return argmax
#
#
def run_optimization(self,
max_iter=1,
max_time=np.inf,
rep=None,
last_step_evaluator=None,
eps=1e-8,
context=None,
verbosity=False,
path=None,
evaluations_file=None):
"""
Runs Bayesian Optimization for a number 'max_iter' of iterations (after the initial exploration data)
:param max_iter: exploration horizon, or number of acquisitions. If nothing is provided optimizes the current acquisition.
:param max_time: maximum exploration horizon in seconds.
:param eps: minimum distance between two consecutive x's to keep running the model.
:param context: fixes specified variables to a particular context (values) for the optimization run (default, None).
:param verbosity: flag to print the optimization results after each iteration (default, False).
:param evaluations_file: filename of the file where the evaluated points and corresponding evaluations are saved (default, None).
"""
self.last_step_evaluator = last_step_evaluator
if self.objective is None:
raise InvalidConfigError(
"Cannot run the optimization loop without the objective function"
)
# --- Save the options to print and save the results
self.verbosity = verbosity
self.evaluations_file = evaluations_file
self.context = context
self.path = path
self.rep = rep
# --- Setting up stop conditions
self.eps = eps
if (max_iter is None) and (max_time is None):
self.max_iter = 0
self.max_time = np.inf
elif (max_iter is None) and (max_time is not None):
self.max_iter = np.inf
self.max_time = max_time
elif (max_iter is not None) and (max_time is None):
self.max_iter = max_iter
self.max_time = np.inf
else:
self.max_iter = max_iter
self.max_time = max_time
# --- Initial function evaluation and model fitting
if self.X is not None and self.Y is None:
self.Y, cost_values = self.objective.evaluate(self.X)
if self.constraint is not None:
self.C, cost_values = self.constraint.evaluate(self.X)
if self.cost.cost_type == 'evaluation_time':
self.cost.update_cost_model(self.X, cost_values)
#self.model.updateModel(self.X,self.Y)
# --- Initialize iterations and running time
self.time_zero = time.time()
self.cum_time = 0
self.num_acquisitions = 0
self.suggested_sample = self.X
self.Y_new = self.Y
self.Opportunity_Cost = {"Hypervolume": np.array([])}
value_so_far = []
# --- Initialize time cost of the evaluations
print("MAIN LOOP STARTS")
self.true_best_stats = {
"true_best": [],
"mean_gp": [],
"std gp": [],
"pf": [],
"mu_pf": [],
"var_pf": [],
"residual_noise": []
}
self._update_model()
while (self.max_iter > self.num_acquisitions):
self.optimize_final_evaluation()
print("maKG optimizer")
start = time.time()
self.suggested_sample = self._compute_next_evaluations()
print("self.suggested_sample", self.suggested_sample)
if verbosity:
if self.constraint is not None:
self.verbosity_plot_2D_constrained()
else:
self.verbosity_plot_2D_unconstrained()
# self.suggested_sample = np.array([[0.1,0.0]])
finish = time.time()
print("time optimisation point X", finish - start)
self.X = np.vstack((self.X, self.suggested_sample))
# --- Evaluate *f* in X, augment Y and update cost function (if needed)
self.evaluate_objective()
self._update_model()
# --- Update current evaluation time and function evaluations
self.cum_time = time.time() - self.time_zero
self.num_acquisitions += 1
print("optimize_final_evaluation")
print("num acquired samples Main Alg: ", self.num_acquisitions)
print("self.X, self.Y, self.C , self.Opportunity_Cost", self.X,
self.Y, self.C, self.Opportunity_Cost)
return self.X, self.Y, self.C, self.Opportunity_Cost
# --- Print the desired result in files
#if self.evaluations_file is not None:
#self.save_evaluations(self.evaluations_file)
#file = open('test_file.txt','w')
#np.savetxt('test_file.txt',value_so_far)
def verbosity_plot_1D(self):
####plots
print("generating plots")
design_plot = np.linspace(0, 5, 100)[:, None]
# precision = []
# for i in range(20):
# kg_f = -self.acquisition._compute_acq(design_plot)
# precision.append(np.array(kg_f).reshape(-1))
# print("mean precision", np.mean(precision, axis=0), "std precision", np.std(precision, axis=0), "max precision", np.max(precision, axis=0), "min precision",np.min(precision, axis=0))
ac_f = self.expected_improvement(design_plot)
Y, _ = self.objective.evaluate(design_plot)
C, _ = self.constraint.evaluate(design_plot)
pf = self.probability_feasibility_multi_gp(design_plot,
self.model_c).reshape(
-1, 1)
mu_f = self.model.predict(design_plot)[0]
bool_C = np.product(np.concatenate(C, axis=1) < 0, axis=1)
func_val = Y * bool_C.reshape(-1, 1)
kg_f = -self.acquisition._compute_acq(design_plot)
design_plot = design_plot.reshape(-1)
fig, axs = plt.subplots(3, 2)
axs[0, 0].set_title('True Function')
axs[0, 0].plot(design_plot, np.array(func_val).reshape(-1))
axs[0, 0].scatter(self.X, self.Y, color="red", label="sampled")
# suggested_sample_value , _= self.objective.evaluate(self.suggested_sample)
# axs[0, 0].scatter(self.suggested_sample, suggested_sample_value, marker="x", color="red",
# label="suggested")
axs[0, 0].legend()
axs[0, 1].set_title('approximation Acqu Function')
axs[0, 1].plot(design_plot, np.array(ac_f).reshape(-1))
axs[0, 1].legend()
axs[1, 0].set_title("mu and pf separetely ")
axs[1, 0].plot(design_plot, np.array(mu_f).reshape(-1), label="mu")
axs[1, 0].plot(design_plot, np.array(pf).reshape(-1), label="pf")
axs[1, 0].legend()
axs[1, 1].set_title("mu pf")
axs[1, 1].plot(design_plot,
np.array(mu_f).reshape(-1) * np.array(pf).reshape(-1))
axs[1, 1].legend()
axs[2, 1].set_title('approximation kg Function')
axs[2, 1].plot(design_plot, np.array(kg_f).reshape(-1))
axs[2, 1].legend()
plt.show()
def verbosity_plot_2D_unconstrained(self):
####plots
print("generating plots")
design_plot = initial_design('random', self.space, 10000)
# precision = []
# for i in range(20):
# kg_f = -self.acquisition._compute_acq(design_plot)
# precision.append(np.array(kg_f).reshape(-1))
# print("mean precision", np.mean(precision, axis=0), "std precision", np.std(precision, axis=0), "max precision", np.max(precision, axis=0), "min precision",np.min(precision, axis=0))
func_val, _ = self.objective.evaluate(design_plot)
func_val = np.concatenate(func_val, axis=1)
mu_f = self.model.posterior_mean(design_plot)
var_f = self.model.posterior_variance(design_plot, noise=False)
mu_predicted_best = self.model.posterior_mean(self.suggested_sample)
HVI = self.acquisition._compute_acq(design_plot)
fig, axs = plt.subplots(2, 2)
axs[0, 0].set_title('True PF Function')
axs[0, 0].scatter(func_val[:, 0], func_val[:, 1])
print("self.suggested_sample", self.suggested_sample)
axs[0, 1].set_title("GP(X)")
axs[0,
1].scatter(
design_plot[:, 0],
design_plot[:, 1],
c=np.array(mu_f).reshape(-1)) #,c= np.array(HVI).reshape(-1))
axs[0, 1].scatter(self.suggested_sample[:, 0],
self.suggested_sample[:, 1],
color="magenta")
axs[0, 1].legend()
print("self.suggested_sample", self.suggested_sample)
axs[1, 0].set_title("Var[GP(X)]")
axs[1, 0].scatter(
design_plot[:, 0],
design_plot[:, 1],
c=np.array(var_f).reshape(-1)) #,c= np.array(HVI).reshape(-1))
axs[1, 0].scatter(self.suggested_sample[:, 0],
self.suggested_sample[:, 1],
color="magenta")
axs[1, 0].legend()
axs[1, 1].set_title("acq(X)")
axs[1,
1].scatter(
design_plot[:, 0],
design_plot[:, 1],
c=np.array(HVI).reshape(-1)) #,c= np.array(HVI).reshape(-1))
axs[1, 1].scatter(self.suggested_sample[:, 0],
self.suggested_sample[:, 1],
color="magenta")
axs[1, 1].legend()
# axs[1, 1].set_title("mu pf")
# axs[1, 1].scatter(design_plot[:,0],design_plot[:,1],c= np.array(mu_f).reshape(-1) * np.array(pf).reshape(-1))
# axs[1, 1].legend()
#
# axs[1, 0].set_title('Opportunity Cost')
# axs[1, 0].plot(range(len(self.Opportunity_Cost)), self.Opportunity_Cost)
# axs[1, 0].set_yscale("log")
# axs[1, 0].legend()
# axs[1, 1].set_title('True PF Function with sampled points')
# axs[1, 1].scatter(func_val[:, 0], func_val[:, 1])
# axs[1, 1].scatter(self.Y[0],self.Y[1], color="red", label="sampled")
# import os
# folder = "IMAGES"
# subfolder = "new_branin"
# cwd = os.getcwd()
# print("cwd", cwd)
# time_taken = time.time()
# path = cwd + "/" + folder + "/" + subfolder + '/im_' +str(time_taken) +str(self.X.shape[0]) + '.pdf'
# if os.path.isdir(cwd + "/" + folder + "/" + subfolder) == False:
# os.makedirs(cwd + "/" + folder + "/" + subfolder)
# plt.savefig(path)
plt.show()
def verbosity_plot_2D_constrained(self):
####plots
print("generating plots")
design_plot = initial_design('random', self.space, 1000)
# precision = []
# for i in range(20):
# kg_f = -self.acquisition._compute_acq(design_plot)
# precision.append(np.array(kg_f).reshape(-1))
# print("mean precision", np.mean(precision, axis=0), "std precision", np.std(precision, axis=0), "max precision", np.max(precision, axis=0), "min precision",np.min(precision, axis=0))
# self.acquisition._gradient_sanity_check_2D(f=self.acquisition._compute_acq, grad_f = self.acquisition.acquisition_Gradients, x_value = self.suggested_sample, delta=1e-4)
Y, _ = self.objective.evaluate(design_plot)
Y = np.concatenate(Y, axis=1)
C, _ = self.constraint.evaluate(design_plot)
pf = self.probability_feasibility_multi_gp(design_plot,
self.model_c).reshape(
-1, 1)
mu_f = self.model.posterior_mean(design_plot)
bool_C = np.product(np.concatenate(C, axis=1) < 0, axis=1)
bool_C = np.array(bool_C, dtype=bool)
func_val = Y[bool_C]
mu_predicted_best = self.model.posterior_mean(self.suggested_sample)
# mu_predicted_final_best = self.model.posterior_mean(self.suggested_final_evaluation)
feasable_mu_index = np.array(pf > 0.51, dtype=bool).reshape(-1)
# HVI = self.acquisition._compute_acq(design_plot[feasable_mu_index ])
# kg_f = -self.acquisition._compute_acq(design_plot)
# HVI_optimiser = self.acquisition._compute_acq(self.suggested_sample)
# print("optimiser best", HVI_optimiser, "discretisation best", np.max(np.array(HVI).reshape(-1)))
# print("x", design_plot[np.argmax(np.array(HVI).reshape(-1))])
# x_suggested_discretisation = design_plot[np.argmax(np.array(HVI).reshape(-1))]
# print("ac with grad info optimised", self.acquisition._compute_acq_withGradients(self.suggested_sample), "ac info optimised",self.acquisition._compute_acq(self.suggested_sample))
# print("best discretisation ac", self.acquisition._compute_acq(x_suggested_discretisation ))
# print("best discretisation ac with gradients", self.acquisition._compute_acq_withGradients(x_suggested_discretisation))
# print("mu predicted best opt", mu_predicted_best[0], mu_predicted_best[1])
# print("mu predicted best discretisation", mu_f[0][np.argmax(np.array(HVI).reshape(-1))], mu_f[1][np.argmax(np.array(HVI).reshape(-1))])
fig, axs = plt.subplots(2, 2)
axs[0, 0].set_title('True PF Function')
axs[0, 0].scatter(func_val[:, 0], func_val[:, 1])
axs[0, 1].set_title("HVI")
axs[0, 1].scatter(mu_f[0], mu_f[1],
color="green") # , c=np.array(HVI).reshape(-1))
axs[0, 1].scatter(mu_f[0][feasable_mu_index],
mu_f[1][feasable_mu_index],
color="blue") #, c=np.array(HVI).reshape(-1))
axs[0, 1].scatter(mu_predicted_best[0],
mu_predicted_best[1],
color="red",
label="optimiser best")
# axs[0, 1].scatter(mu_predicted_final_best[0], mu_predicted_final_best[1], color="red", label="optimiser final best")
#axs[0, 1].scatter(mu_f[0][np.argmax(np.array(HVI).reshape(-1))], mu_f[1][np.argmax(np.array(HVI).reshape(-1))], color="red",label="discretisation best")
axs[0, 1].legend()
axs[1, 0].set_title('Opportunity Cost')
axs[1, 0].plot(range(len(self.Opportunity_Cost["Hypervolume"])),
self.Opportunity_Cost["Hypervolume"])
axs[1, 0].set_yscale("log")
axs[1, 0].legend()
Y_reccomended, _ = self.objective.evaluate(self.suggested_sample)
# Y_reccomended = np.concatenate(Y_reccomended, axis=1)
axs[1, 1].set_title('True PF Function with sampled points')
axs[1, 1].scatter(func_val[:, 0], func_val[:, 1])
axs[1, 1].scatter(Y_reccomended[0],
Y_reccomended[1],
color="red",
label="sampled")
axs[1, 1].scatter(self.Y[0], self.Y[1], color="green")
# import os
# folder = "IMAGES"
# subfolder = "new_branin"
# cwd = os.getcwd()
# print("cwd", cwd)
# time_taken = time.time()
# path = cwd + "/" + folder + "/" + subfolder + '/im_' +str(time_taken) +str(self.X.shape[0]) + '.pdf'
# if os.path.isdir(cwd + "/" + folder + "/" + subfolder) == False:
# os.makedirs(cwd + "/" + folder + "/" + subfolder)
# plt.savefig(path)
plt.show()
def optimize_final_evaluation(self):
if self.last_step_evaluator is None:
if self.constraint is None:
sampled_Y = self.model.get_Y_values()
sampled_Y = np.concatenate(sampled_Y, axis=1).tolist()
sampled_hv = hypervolume(sampled_Y)
sampled_HV = sampled_hv.compute(ref_point=self.ref_point)
#self.Opportunity_Cost.append(sampled_HV)
feasable_Y = sampled_Y
self.store_results(feasable_Y)
else:
sampled_Y = self.model.get_Y_values()
sampled_Y = np.concatenate(sampled_Y, axis=1)
C_true, C_cost_new = self.constraint.evaluate(self.X,
true_val=True)
feasable_samples = np.product(
np.concatenate(C_true, axis=1) < 0, axis=1)
feasable_samples = np.array(feasable_samples, dtype=bool)
feasable_Y = sampled_Y[feasable_samples]
self.store_results(feasable_Y)
else:
if self.constraint is not None:
suggested_sample = self._compute_final_evaluations()
self.suggested_final_evaluation = suggested_sample
Y_new, cost_new = self.objective.evaluate(suggested_sample)
Y_new = np.concatenate(Y_new, axis=1)
C_new, C_cost_new = self.constraint.evaluate(suggested_sample)
C_new = np.concatenate(C_new, axis=1)
sampled_Y = self.model.get_Y_values()
sampled_Y = np.concatenate(sampled_Y, axis=1)
sampled_Y = np.vstack((sampled_Y, Y_new))
C_true, C_cost_new = self.constraint.evaluate(self.X,
true_val=True)
C_true = np.concatenate(C_true, axis=1)
C_true = np.vstack((C_true, C_new))
feasable_samples = np.product(C_true < 0, axis=1)
feasable_samples = np.array(feasable_samples, dtype=bool)
feasable_Y = sampled_Y[feasable_samples]
self.store_results(feasable_Y)
else:
suggested_sample = self._compute_final_evaluations()
# self.suggested_final_evaluation = suggested_sample
#
# Y_new, cost_new = self.objective.evaluate(suggested_sample)
# Y_new = np.concatenate(Y_new, axis=1)
#
# X_train = self.model.get_X_values()
# sampled_Y , cost_new = self.objective.evaluate(X_train)
# sampled_Y = np.concatenate(sampled_Y, axis=1)
# sampled_Y = np.vstack((sampled_Y, Y_new))
#
# feasable_Y = sampled_Y
# self.store_results(feasable_Y)
# design_plot = initial_design('random', self.space, 1000)
# ac_f = self.expected_improvement(design_plot)
# fig, axs = plt.subplots(2, 2)
# axs[0, 0].set_title('True Function')
# axs[0, 0].scatter(design_plot[:, 0], design_plot[:, 1], c=np.array(ac_f).reshape(-1))
#
# if self.tag_last_evaluation:
# start = time.time()
# self.acquisition.optimizer.context_manager = ContextManager(self.space, self.context)
# if self.constraint is None:
# out = self.acquisition.optimizer.optimize_inner_func(f=self.expected_improvement_unconstrained, duplicate_manager=None, num_samples=100)
# else:
# out = self.acquisition.optimizer.optimize_inner_func(f=self.expected_improvement_constrained,
# duplicate_manager=None, num_samples=100)
# print("out",out)
# suggested_sample = self.space.zip_inputs(out[0])
# stop = time.time()
# # axs[0, 0].scatter(suggested_sample[:, 0], suggested_sample[:, 1], color="red")
# # plt.show()
# print("time EI", stop - start)
# # print("self.suggested_sample",suggested_sample)
# # --- Evaluate *f* in X, augment Y and update cost function (if needed)
#
#
# if self.deterministic:
# if self.constraint is None:
# func_val , _ = self.objective.evaluate(suggested_sample, true_val=True)
# # print("Y",Y)
# Y_true, cost_new = self.objective.evaluate(self.X, true_val=True)
#
# feasable_Y_data = np.array(Y_true).reshape(-1)
# Y_aux = np.concatenate((func_val.reshape(-1), np.array(feasable_Y_data).reshape(-1)))
# else:
# Y, _ = self.objective.evaluate(suggested_sample, true_val=True)
# # print("Y",Y)
#
# C, _ = self.constraint.evaluate(suggested_sample, true_val=True)
# bool_C = np.product(np.concatenate(C, axis=1) < 0, axis=1)
# func_val = Y * bool_C.reshape(-1, 1)
#
# Y_true, cost_new = self.objective.evaluate(self.X ,true_val=True)
# C_true, C_cost_new = self.constraint.evaluate(self.X ,true_val=True)
#
# feasable_Y_data = np.array(Y_true).reshape(-1) * np.product(np.concatenate(C_true, axis=1) < 0, axis=1)
#
# Y_aux = np.concatenate((func_val.reshape(-1), np.array(feasable_Y_data).reshape(-1)))
#
# if self.expensive:
# self.Opportunity_Cost.append(np.array(np.abs(np.max(Y_aux))).reshape(-1))
# else:
# self.true_best_value()
# optimum = np.max(np.abs(self.true_best_stats["true_best"]))
# print("optimum", optimum)
# self.Opportunity_Cost.append(optimum - np.array(np.abs(np.max(Y_aux))).reshape(-1))
#
# else:
# # print("self.X,suggested_sample",self.X,suggested_sample)
#
# samples = np.concatenate((self.X,suggested_sample))
# # print("samples", samples)
# Y= self.model.posterior_mean(samples)
# # print("Y",Y)
# pf = self.probability_feasibility_multi_gp(samples, model=self.model_c)
#
# # print("pf", pf)
# func_val = np.array(Y).reshape(-1) * np.array(pf).reshape(-1)
#
# # print("Y", Y, "pf", pf, "func_val", func_val)
# suggested_final_sample = samples[np.argmax(func_val)]
# suggested_final_sample = np.array(suggested_final_sample).reshape(-1)
# suggested_final_sample = np.array(suggested_final_sample).reshape(1,-1)
# # print("suggested_final_sample", suggested_final_sample, "val",np.max(func_val) )
# Y_true, _ = self.objective.evaluate(suggested_final_sample, true_val=True)
# # print("Y_true", Y_true)
# C_true, _ = self.constraint.evaluate(suggested_final_sample, true_val=True)
# # print("C_true", C_true)
# bool_C_true = np.product(np.concatenate(C_true, axis=1) < 0, axis=1)
# func_val_true = Y_true * bool_C_true.reshape(-1, 1)
# # print("func_val_true",func_val_true)
#
# if self.expensive:
# self.Opportunity_Cost.append(np.array(np.abs(np.max(func_val_true))).reshape(-1))
# else:
# self.true_best_value()
# optimum = np.max(np.abs(self.true_best_stats["true_best"]))
# # print("optimum", optimum)
# self.Opportunity_Cost.append(optimum - np.array(np.abs(np.max(func_val_true))).reshape(-1))
# # print("OC_i", optimum - np.array(np.abs(np.max(func_val_true))).reshape(-1))
#
# else:
# samples = self.X
# # print("samples", samples)
# Y = self.model.posterior_mean(samples)
# # print("Y",Y)
# pf = self.probability_feasibility_multi_gp(samples, model=self.model_c)
#
# # print("pf", pf)
# func_val = np.array(Y).reshape(-1) * np.array(pf).reshape(-1)
#
# # print("Y", Y, "pf", pf, "func_val", func_val)
# suggested_final_sample = samples[np.argmax(func_val)]
# suggested_final_sample = np.array(suggested_final_sample).reshape(-1)
# suggested_final_sample = np.array(suggested_final_sample).reshape(1, -1)
# # print("suggested_final_sample", suggested_final_sample, "val",np.max(func_val) )
# Y_true, _ = self.objective.evaluate(suggested_final_sample, true_val=True)
# # print("Y_true", Y_true)
# C_true, _ = self.constraint.evaluate(suggested_final_sample, true_val=True)
# # print("C_true", C_true)
# bool_C_true = np.product(np.concatenate(C_true, axis=1) < 0, axis=1)
# func_val_true = Y_true * bool_C_true.reshape(-1, 1)
#
# if self.expensive:
# self.Opportunity_Cost.append(np.array(np.abs(np.max(func_val_true))).reshape(-1))
# else:
# self.true_best_value()
# optimum = np.max(np.abs(self.true_best_stats["true_best"]))
# # print("optimum", optimum)
# self.Opportunity_Cost.append(optimum - np.array(np.abs(np.max(func_val_true))).reshape(-1))
# # print("OC_i", optimum - np.array(np.abs(np.max(func_val_true))).reshape(-1))
#
def store_results(self, feasable_Y):
if len(feasable_Y) >= 1:
sampled_hv = hypervolume(-feasable_Y)
sampled_HV = 0 #sampled_hv.compute(ref_point=self.ref_point)
else:
sampled_HV = 0
self.Opportunity_Cost["Hypervolume"] = np.concatenate(
(self.Opportunity_Cost["Hypervolume"],
np.array(sampled_HV).reshape(-1)))
data = self.Opportunity_Cost
if self.path is not None:
gen_file = pd.DataFrame.from_dict(data)
results_folder = "HVI"
path = self.path + "/" + results_folder + '/it_' + str(
self.rep) + '.csv'
if os.path.isdir(self.path + "/" + results_folder) == False:
os.makedirs(self.path + "/" + results_folder)
gen_file.to_csv(path_or_buf=path)
def expected_improvement_unconstrained(self, X, offset=1e-4):
'''
Computes the EI at points X based on existing samples X_sample
and Y_sample using a Gaussian process surrogate model.
Args:
X: Points at which EI shall be computed (m x d).
X_sample: Sample locations (n x d).
Y_sample: Sample values (n x 1).
gpr: A GaussianProcessRegressor fitted to samples.
xi: Exploitation-exploration trade-off parameter.
Returns:
Expected improvements at points X.
'''
if self.deterministic:
#print("DETERMINISTIC LAST STEP")
mu = self.model.posterior_mean(X)
sigma = self.model.posterior_variance(X, noise=False)
sigma = np.sqrt(sigma).reshape(-1, 1)
mu = mu.reshape(-1, 1)
# Needed for noise-based model,
# otherwise use np.max(Y_sample).
# See also section 2.4 in [...]
func_val = self.Y #* bool_C.reshape(-1, 1)
mu_sample_opt = np.max(func_val) - offset
with np.errstate(divide='warn'):
imp = mu - mu_sample_opt
Z = imp / sigma
ei = imp * norm.cdf(Z) + sigma * norm.pdf(Z)
ei[sigma == 0.0] = 0.0
return -np.array(ei).reshape(-1)
else:
print("NOISY LAST STEP")
mu = self.model.posterior_mean(X)
mu = mu.reshape(-1, 1)
return -np.array(mu).reshape(-1)
def expected_improvement_constrained(self, X, offset=1e-4):
'''
Computes the EI at points X based on existing samples X_sample
and Y_sample using a Gaussian process surrogate model.
Args:
X: Points at which EI shall be computed (m x d).
X_sample: Sample locations (n x d).
Y_sample: Sample values (n x 1).
gpr: A GaussianProcessRegressor fitted to samples.
xi: Exploitation-exploration trade-off parameter.
Returns:
Expected improvements at points X.
'''
if self.deterministic:
#print("DETERMINISTIC LAST STEP")
mu = self.model.posterior_mean(X)
sigma = self.model.posterior_variance(X, noise=False)
sigma = np.sqrt(sigma).reshape(-1, 1)
mu = mu.reshape(-1, 1)
# Needed for noise-based model,
# otherwise use np.max(Y_sample).
# See also section 2.4 in [...]
bool_C = np.product(np.concatenate(self.C, axis=1) < 0, axis=1)
func_val = self.Y * bool_C.reshape(-1, 1)
mu_sample_opt = np.max(func_val) - offset
with np.errstate(divide='warn'):
imp = mu - mu_sample_opt
Z = imp / sigma
ei = imp * norm.cdf(Z) + sigma * norm.pdf(Z)
ei[sigma == 0.0] = 0.0
pf = self.probability_feasibility_multi_gp(X,
self.model_c).reshape(
-1, 1)
return -(ei * pf)
else:
print("NOISY LAST STEP")
mu = self.model.posterior_mean(X)
mu = mu.reshape(-1, 1)
pf = self.probability_feasibility_multi_gp(X,
self.model_c).reshape(
-1, 1)
return -(mu * pf)
def probability_feasibility_multi_gp(self,
x,
model,
mean=None,
cov=None,
grad=False,
l=0):
# print("model",model.output)
x = np.atleast_2d(x)
Fz = []
for m in range(model.output_dim):
Fz.append(self.probability_feasibility(x, model.output[m], grad,
l))
Fz = np.product(Fz, axis=0)
return Fz
def probability_feasibility(self,
x,
model,
mean=None,
cov=None,
grad=False,
l=0):
model = model.model
# kern = model.kern
# X = model.X
mean = model.posterior_mean(x)
var = model.posterior_variance(x, noise=False)
# print("mean",mean,"var",var)
std = np.sqrt(var).reshape(-1, 1)
# print("std",std)
aux_var = np.reciprocal(var)
mean = mean.reshape(-1, 1)
norm_dist = norm(mean, std)
fz = norm_dist.pdf(l)
Fz = norm_dist.cdf(l)
if grad == True:
grad_mean, grad_var = model.predictive_gradients(x)
grad_std = (1 / 2.0) * grad_var
# cov = kern.K(X, X) + np.eye(X.shape[0]) * 1e-3
# L = scipy.linalg.cholesky(cov, lower=True)
# u = scipy.linalg.solve(L, np.eye(X.shape[0]))
# Ainv = scipy.linalg.solve(L.T, u)
dims = range(x.shape[1])
grad_Fz = []
for d in dims:
# K1 = np.diag(np.dot(np.dot(kern.dK_dX(x, X, d), Ainv), kern.dK_dX2(X, x, d)))
# K2 = np.diag(kern.dK2_dXdX2(x, x, d, d))
# var_grad = K2 - K1
# var_grad = var_grad.reshape(-1, 1)
grd_mean_d = grad_mean[:, d].reshape(-1, 1)
grd_std_d = grad_std[:, d].reshape(-1, 1)
# print("grd_mean ",grd_mean_d, "var_grad ",grd_std_d )
# print("std",std)
# print("aux_var", aux_var)
# print("fz",fz)
grad_Fz.append(fz * aux_var *
(mean * grd_std_d - grd_mean_d * std))
grad_Fz = np.stack(grad_Fz, axis=1)
return Fz.reshape(-1, 1), grad_Fz[:, :, 0]
else:
return Fz.reshape(-1, 1)
def evaluate_objective(self):
"""
Evaluates the objective
"""
print(1)
print(self.suggested_sample)
self.Y_new, cost_new = self.objective.evaluate(self.suggested_sample)
if self.constraint is not None:
self.C_new, C_cost_new = self.constraint.evaluate(
self.suggested_sample)
for k in range(self.constraint.output_dim):
self.C[k] = np.vstack((self.C[k], self.C_new[k]))
self.cost.update_cost_model(self.suggested_sample, cost_new)
for j in range(self.objective.output_dim):
print(self.Y_new[j])
self.Y[j] = np.vstack((self.Y[j], self.Y_new[j]))
def compute_current_best(self):
current_acqX = self.acquisition.current_compute_acq()
return current_acqX
def _distance_last_evaluations(self):
"""
Computes the distance between the last two evaluations.
"""
return np.sqrt(
sum((self.X[self.X.shape[0] - 1, :] -
self.X[self.X.shape[0] - 2, :])**2))
def _compute_next_evaluations(self,
pending_zipped_X=None,
ignored_zipped_X=None,
re_use=False):
"""
Computes the location of the new evaluation (optimizes the acquisition in the standard case).
:param pending_zipped_X: matrix of input configurations that are in a pending state (i.e., do not have an evaluation yet).
:param ignored_zipped_X: matrix of input configurations that the user black-lists, i.e., those configurations will not be suggested again.
:return:
"""
## --- Update the context if any
self.acquisition.optimizer.context_manager = ContextManager(
self.space,
self.context,
)
print("compute next evaluation")
if self.sample_from_acq:
print("suggest next location given THOMPSON SAMPLING")
candidate_points = initial_design('latin', self.space, 2000)
aux_var = self.acquisition._compute_acq(candidate_points)
else:
if self.constraint is not None:
aux_var = self.evaluator.compute_batch(duplicate_manager=None,
re_use=re_use,
constrained=True)
else:
aux_var = self.evaluator.compute_batch(duplicate_manager=None,
re_use=re_use,
constrained=False)
return self.space.zip_inputs(aux_var[0])
def _compute_final_evaluations(self,
pending_zipped_X=None,
ignored_zipped_X=None,
re_use=False):
"""
Computes the location of the new evaluation (optimizes the acquisition in the standard case).
:param pending_zipped_X: matrix of input configurations that are in a pending state (i.e., do not have an evaluation yet).
:param ignored_zipped_X: matrix of input configurations that the user black-lists, i.e., those configurations will not be suggested again.
:return:
"""
## --- Update the context if any
self.acquisition.optimizer.context_manager = ContextManager(
self.space,
self.context,
)
print("compute next evaluation")
if self.sample_from_acq:
print("suggest next location given THOMPSON SAMPLING")
candidate_points = initial_design('latin', self.space, 2000)
aux_var = self.acquisition._compute_acq(candidate_points)
else:
if self.constraint is not None:
aux_var = self.last_step_evaluator.compute_batch(
duplicate_manager=None, re_use=re_use, constrained=True)
else:
aux_var = self.last_step_evaluator.compute_batch(
duplicate_manager=None, re_use=re_use, constrained=False)
return self.space.zip_inputs(aux_var[0])
#return initial_design('random', self.space, 1)
def _update_model(self):
"""
Updates the model (when more than one observation is available) and saves the parameters (if available).
"""
if (self.num_acquisitions % self.model_update_interval) == 0:
### --- input that goes into the model (is unziped in case there are categorical variables)
X_inmodel = self.space.unzip_inputs(self.X)
Y_inmodel = list(self.Y)
Y_inmodel = -np.concatenate(Y_inmodel, axis=1)
scalarisation_rdn = self.utility.sample_parameter(
1) #np.array([[0.9055273, 0.0944727]])# np.array([[0.5,0.5]])
Utility = self.utility.func(parameter=scalarisation_rdn,
y=Y_inmodel)
Utility = [Utility]
if self.constraint is not None:
C_inmodel = list(self.C)
self.model_c.updateModel(X_inmodel, C_inmodel)
self.model.updateModel(X_inmodel, Utility)
### --- Save parameters of the model
#self._save_model_parameter_values()
def get_evaluations(self):
return self.X.copy(), self.Y.copy()
def true_best_value(self):
from scipy.optimize import minimize
X = initial_design('random', self.space, 1000)
fval = self.func_val(X)
anchor_point = np.array(X[np.argmin(fval)]).reshape(-1)
anchor_point = anchor_point.reshape(1, -1)
print("anchor_point", anchor_point)
best_design = minimize(self.func_val,
anchor_point,
method='Nelder-Mead',
tol=1e-8).x
self.true_best_stats["true_best"].append(self.func_val(best_design))
self.true_best_stats["mean_gp"].append(
self.model.posterior_mean(best_design))
self.true_best_stats["std gp"].append(
self.model.posterior_variance(best_design, noise=False))
self.true_best_stats["pf"].append(
self.probability_feasibility_multi_gp(best_design,
self.model_c).reshape(-1, 1))
mean = self.model_c.posterior_mean(best_design)
var = self.model_c.posterior_variance(best_design, noise=False)
residual_noise = self.model_c.posterior_variance(self.X[1],
noise=False)
self.true_best_stats["mu_pf"].append(mean)
self.true_best_stats["var_pf"].append(var)
self.true_best_stats["residual_noise"].append(residual_noise)
if False:
fig, axs = plt.subplots(3, 2)
N = len(np.array(self.true_best_stats["std gp"]).reshape(-1))
GAP = np.array(
np.abs(
np.abs(self.true_best_stats["true_best"]).reshape(-1) -
np.abs(self.true_best_stats["mean_gp"]).reshape(-1))
).reshape(-1)
print("GAP len", len(GAP))
print("N", N)
axs[0, 0].set_title('GAP')
axs[0, 0].plot(range(N), GAP)
axs[0, 0].set_yscale("log")
axs[0, 1].set_title('VAR')
axs[0,
1].plot(range(N),
np.array(self.true_best_stats["std gp"]).reshape(-1))
axs[0, 1].set_yscale("log")
axs[1, 0].set_title("PF")
axs[1, 0].plot(range(N),
np.array(self.true_best_stats["pf"]).reshape(-1))
axs[1, 1].set_title("mu_PF")
axs[1, 1].plot(
range(N),
np.abs(np.array(self.true_best_stats["mu_pf"]).reshape(-1)))
axs[1, 1].set_yscale("log")
axs[2, 1].set_title("std_PF")
axs[2, 1].plot(
range(N),
np.sqrt(np.array(self.true_best_stats["var_pf"]).reshape(-1)))
axs[2, 1].set_yscale("log")
axs[2, 0].set_title("Irreducible noise")
axs[2, 0].plot(
range(N),
np.sqrt(
np.array(
self.true_best_stats["residual_noise"]).reshape(-1)))
axs[2, 0].set_yscale("log")
plt.show()
def func_val(self, x):
if len(x.shape) == 1:
x = x.reshape(1, -1)
Y, _ = self.objective.evaluate(x, true_val=True)
C, _ = self.constraint.evaluate(x, true_val=True)
Y = np.array(Y).reshape(-1)
out = Y.reshape(-1) * np.product(np.concatenate(C, axis=1) < 0,
axis=1).reshape(-1)
out = np.array(out).reshape(-1)
return -out
class BO(object):
"""
Runner of the multi-attribute Bayesian optimization loop. This class wraps the optimization loop around the different handlers.
:param model: GPyOpt model class.
:param space: GPyOpt space class.
:param objective: GPyOpt objective class.
:param acquisition: GPyOpt acquisition class.
:param evaluator: GPyOpt evaluator class.
:param X_init: 2d numpy array containing the initial inputs (one per row) of the model.
:param Y_init: 2d numpy array containing the initial outputs (one per row) of the model.
:param cost: GPyOpt cost class (default, none).
:param normalize_Y: whether to normalize the outputs before performing any optimization (default, True).
:param model_update_interval: interval of collected observations after which the model is updated (default, 1).
:param de_duplication: GPyOpt DuplicateManager class. Avoids re-evaluating the objective at previous, pending or infeasible locations (default, False).
"""
def __init__(self, model, model_c,space, objective, constraint, acquisition, evaluator, X_init , Y_init=None, C_init=None, cost = None, normalize_Y = False, model_update_interval = 1, true_preference = 0.5):
self.true_preference = true_preference
self.model_c = model_c
self.model = model
self.space = space
self.objective = objective
self.constraint = constraint
self.acquisition = acquisition
self.utility = acquisition.utility
self.evaluator = evaluator
self.normalize_Y = normalize_Y
self.model_update_interval = model_update_interval
self.X = X_init
self.Y = Y_init
self.C = C_init
self.cost = CostModel(cost)
self.model_parameters_iterations = None
def suggest_next_locations(self, context = None, pending_X = None, ignored_X = None):
"""
Run a single optimization step and return the next locations to evaluate the objective.
Number of suggested locations equals to batch_size.
:param context: fixes specified variables to a particular context (values) for the optimization run (default, None).
:param pending_X: matrix of input configurations that are in a pending state (i.e., do not have an evaluation yet) (default, None).
:param ignored_X: matrix of input configurations that the user black-lists, i.e., those configurations will not be suggested again (default, None).
"""
self.model_parameters_iterations = None
self.num_acquisitions = 0
self.context = context
self._update_model(self.normalization_type)
suggested_locations = self._compute_next_evaluations(pending_zipped_X = pending_X, ignored_zipped_X = ignored_X)
return suggested_locations
# def _value_so_far(self):
# """
# Computes E_n[U(f(x_max))|f], where U is the utility function, f is the true underlying ojective function and x_max = argmax E_n[U(f(x))|U]. See
# function _marginal_max_value_so_far below.
# """
#
# output = 0
# support = self.utility.parameter_dist.support
# utility_dist = self.utility.parameter_dist.prob_dist
#
# a = np.reshape(self.objective.evaluate(self._marginal_max_value_so_far())[0],(self.objective.output_dim,))
#
# output += self.utility.eval_func(support[i],a)*utility_dist[i]
# #print(output)
# return output
#
#
# def _marginal_max_value_so_far(self):
# """
# Computes argmax E_n[U(f(x))|U] (The abuse of notation can be misleading; note that the expectation is with
# respect to the posterior distribution on f after n evaluations)
# """
#
# def val_func(X):
# X = np.atleast_2d(X)
# muX = self.model.posterior_mean(X)
# return muX
#
# def val_func_with_gradient(X):
# X = np.atleast_2d(X)
# muX = self.model.posterior_mean(X)
# dmu_dX = self.model.posterior_mean_gradient(X)
# valX = np.reshape( muX, (X.shape[0],1))
# dval_dX = dmu_dX
# return -valX, -dval_dX
#
#
# argmax = self.acquisition.optimizer.optimize_inner_func(f=val_func, f_df=val_func_with_gradient)[0]
# return argmax
#
#
def run_optimization(self, max_iter = 1, max_time = np.inf, eps = 1e-8, context = None, verbosity=False, evaluations_file = None):
"""
Runs Bayesian Optimization for a number 'max_iter' of iterations (after the initial exploration data)
:param max_iter: exploration horizon, or number of acquisitions. If nothing is provided optimizes the current acquisition.
:param max_time: maximum exploration horizon in seconds.
:param eps: minimum distance between two consecutive x's to keep running the model.
:param context: fixes specified variables to a particular context (values) for the optimization run (default, None).
:param verbosity: flag to print the optimization results after each iteration (default, False).
:param evaluations_file: filename of the file where the evaluated points and corresponding evaluations are saved (default, None).
"""
if self.objective is None:
raise InvalidConfigError("Cannot run the optimization loop without the objective function")
# --- Save the options to print and save the results
self.verbosity = verbosity
self.evaluations_file = evaluations_file
self.context = context
# --- Setting up stop conditions
self.eps = eps
if (max_iter is None) and (max_time is None):
self.max_iter = 0
self.max_time = np.inf
elif (max_iter is None) and (max_time is not None):
self.max_iter = np.inf
self.max_time = max_time
elif (max_iter is not None) and (max_time is None):
self.max_iter = max_iter
self.max_time = np.inf
else:
self.max_iter = max_iter
self.max_time = max_time
# --- Initial function evaluation and model fitting
if self.X is not None and self.Y is None:
self.Y, cost_values = self.objective.evaluate(self.X)
self.C, cost_values = self.constraint.evaluate(self.X)
if self.cost.cost_type == 'evaluation_time':
self.cost.update_cost_model(self.X, cost_values)
#self.model.updateModel(self.X,self.Y)
# --- Initialize iterations and running time
self.time_zero = time.time()
self.cum_time = 0
self.num_acquisitions = 0
self.suggested_sample = self.X
self.Y_new = self.Y
self.Opportunity_Cost = []
value_so_far = []
# --- Initialize time cost of the evaluations
print("MAIN LOOP STARTS")
Opportunity_Cost = []
while (self.max_iter > self.num_acquisitions ):
# self._update_model()
print("maKG optimizer")
start = time.time()
self.suggested_sample = initial_design('random', self.space, 1)
finish = time.time()
print("time optimisation point X", finish - start)
if verbosity:
self.verbosity_plot_2D()
print("self.Opportunity_Cost",self.Opportunity_Cost)
self.X = np.vstack((self.X,self.suggested_sample))
# --- Evaluate *f* in X, augment Y and update cost function (if needed)
self.evaluate_objective()
# --- Update current evaluation time and function evaluations
self.cum_time = time.time() - self.time_zero
self.num_acquisitions += 1
print("optimize_final_evaluation")
self.optimize_final_evaluation()
print("self.X, self.Y, self.C , self.Opportunity_Cost",self.X, self.Y, self.C , self.Opportunity_Cost)
return self.X, self.Y, self.C , self.Opportunity_Cost
# --- Print the desired result in files
#if self.evaluations_file is not None:
#self.save_evaluations(self.evaluations_file)
#file = open('test_file.txt','w')
#np.savetxt('test_file.txt',value_so_far)
def verbosity_plot_1D(self):
####plots
print("generating plots")
design_plot = np.linspace(0,5,100)[:,None]
# precision = []
# for i in range(20):
# kg_f = -self.acquisition._compute_acq(design_plot)
# precision.append(np.array(kg_f).reshape(-1))
# print("mean precision", np.mean(precision, axis=0), "std precision", np.std(precision, axis=0), "max precision", np.max(precision, axis=0), "min precision",np.min(precision, axis=0))
ac_f = self.expected_improvement(design_plot)
Y, _ = self.objective.evaluate(design_plot)
C, _ = self.constraint.evaluate(design_plot)
pf = self.probability_feasibility_multi_gp(design_plot, self.model_c).reshape(-1, 1)
mu_f = self.model.predict(design_plot)[0]
bool_C = np.product(np.concatenate(C, axis=1) < 0, axis=1)
func_val = Y * bool_C.reshape(-1, 1)
kg_f = -self.acquisition._compute_acq(design_plot)
design_plot = design_plot.reshape(-1)
fig, axs = plt.subplots(3, 2)
axs[0, 0].set_title('True Function')
axs[0, 0].plot(design_plot, np.array(func_val).reshape(-1))
axs[0, 0].scatter(self.X, self.Y, color="red", label="sampled")
suggested_sample_value , _= self.objective.evaluate(self.suggested_sample)
axs[0, 0].scatter(self.suggested_sample, suggested_sample_value, marker="x", color="red",
label="suggested")
axs[0, 0].legend()
axs[0, 1].set_title('approximation Acqu Function')
axs[0, 1].plot(design_plot, np.array(ac_f).reshape(-1))
axs[0, 1].legend()
axs[1, 0].set_title("mu and pf separetely ")
axs[1, 0].plot(design_plot, np.array(mu_f).reshape(-1) , label="mu")
axs[1, 0].plot(design_plot, np.array(pf).reshape(-1), label="pf")
axs[1, 0].legend()
axs[1, 1].set_title("mu pf")
axs[1, 1].plot(design_plot, np.array(mu_f).reshape(-1) * np.array(pf).reshape(-1))
axs[1, 1].legend()
axs[2, 1].set_title('approximation kg Function')
axs[2, 1].plot(design_plot, np.array(kg_f).reshape(-1))
axs[2, 1].legend()
plt.show()
def verbosity_plot_2D(self):
####plots
print("generating plots")
design_plot = initial_design('random', self.space, 1000)
# precision = []
# for i in range(20):
# kg_f = -self.acquisition._compute_acq(design_plot)
# precision.append(np.array(kg_f).reshape(-1))
# print("mean precision", np.mean(precision, axis=0), "std precision", np.std(precision, axis=0), "max precision", np.max(precision, axis=0), "min precision",np.min(precision, axis=0))
ac_f = self.expected_improvement(design_plot)
Y, _ = self.objective.evaluate(design_plot)
C, _ = self.constraint.evaluate(design_plot)
pf = self.probability_feasibility_multi_gp(design_plot, self.model_c).reshape(-1, 1)
mu_f = self.model.predict(design_plot)[0]
bool_C = np.product(np.concatenate(C, axis=1) < 0, axis=1)
func_val = Y * bool_C.reshape(-1, 1)
# kg_f = -self.acquisition._compute_acq(design_plot)
fig, axs = plt.subplots(2, 2)
axs[0, 0].set_title('True Function')
axs[0, 0].scatter(design_plot[:, 0], design_plot[:, 1], c=np.array(func_val).reshape(-1))
axs[0, 0].scatter(self.X[:, 0], self.X[:, 1], color="red", label="sampled")
#suggested_sample_value = self.objective.evaluate(self.suggested_sample)
axs[0, 0].scatter(self.suggested_sample[:,0], self.suggested_sample[:,1], marker="x", color="red",
label="suggested")
axs[0, 0].legend()
axs[0, 1].set_title('approximation Acqu Function')
axs[0, 1].scatter(design_plot[:,0],design_plot[:,1], c=np.array(ac_f).reshape(-1))
axs[0, 1].legend()
# axs[1, 0].set_title("KG")
# axs[1, 0].scatter(design_plot[:,0],design_plot[:,1],c= np.array(kg_f).reshape(-1))
# axs[1, 0].legend()
axs[1, 1].set_title("mu pf")
axs[1, 1].scatter(design_plot[:,0],design_plot[:,1],c= np.array(mu_f).reshape(-1) * np.array(pf).reshape(-1))
axs[1, 1].legend()
# axs[2, 1].set_title('approximation kg Function')
# axs[2, 1].scatter(design_plot, np.array(kg_f).reshape(-1))
# axs[2, 1].legend()
# import os
# folder = "IMAGES"
# subfolder = "new_branin"
# cwd = os.getcwd()
# print("cwd", cwd)
# time_taken = time.time()
# path = cwd + "/" + folder + "/" + subfolder + '/im_' +str(time_taken) +str(self.X.shape[0]) + '.pdf'
# if os.path.isdir(cwd + "/" + folder + "/" + subfolder) == False:
# os.makedirs(cwd + "/" + folder + "/" + subfolder)
# plt.savefig(path)
plt.show()
def optimize_final_evaluation(self):
# design_plot = initial_design('random', self.space, 1000)
# ac_f = self.expected_improvement(design_plot)
# fig, axs = plt.subplots(2, 2)
# axs[0, 0].set_title('True Function')
# axs[0, 0].scatter(design_plot[:, 0], design_plot[:, 1], c=np.array(ac_f).reshape(-1))
# start = time.time()
# self.acquisition.optimizer.context_manager = ContextManager(self.space, self.context)
# out = self.acquisition.optimizer.optimize(f=self.expected_improvement, duplicate_manager=None, re_use=False, num_samples=20, verbose=False)
# suggested_sample = self.space.zip_inputs(out[0])
# stop = time.time()
# # axs[0, 0].scatter(suggested_sample[:, 0], suggested_sample[:, 1], color="red")
# # plt.show()
# print("time EI", stop - start)
# # print("self.suggested_sample",suggested_sample)
# # --- Evaluate *f* in X, augment Y and update cost function (if needed)
# Y, _ = self.objective.evaluate(suggested_sample)
# # print("Y",Y)
# C, _ = self.constraint.evaluate(suggested_sample)
#
# bool_C = np.product(np.concatenate(C, axis=1) < 0, axis=1)
# func_val = Y * bool_C.reshape(-1, 1)
feasable_Y_data = np.array(self.Y).reshape(-1) * np.product(np.concatenate(self.C, axis=1) < 0, axis=1)
# print("suggested_sample",suggested_sample, "feasable_Y_data",func_val)
# print("C[-1, :]",C[-1, :])
# feasable_point = bool_C
Y_aux = np.array(feasable_Y_data).reshape(-1)
self.Opportunity_Cost.append(np.max(Y_aux))
def expected_improvement(self, X, offset=0.0):
'''
Computes the EI at points X based on existing samples X_sample
and Y_sample using a Gaussian process surrogate model.
Args:
X: Points at which EI shall be computed (m x d).
X_sample: Sample locations (n x d).
Y_sample: Sample values (n x 1).
gpr: A GaussianProcessRegressor fitted to samples.
xi: Exploitation-exploration trade-off parameter.
Returns:
Expected improvements at points X.
'''
mu, sigma = self.model.predict(X)
# sigma = np.sqrt(sigma).reshape(-1, 1)
mu = mu.reshape(-1,1)
# Needed for noise-based model,
# otherwise use np.max(Y_sample).
# See also section 2.4 in [...]
bool_C = np.product(np.concatenate(self.C, axis=1) < 0, axis=1)
func_val = self.Y * bool_C.reshape(-1, 1)
mu_sample_opt = np.max(func_val) - offset
# with np.errstate(divide='warn'):
# imp = mu - mu_sample_opt
# Z = imp / sigma
# ei = imp * norm.cdf(Z) + sigma * norm.pdf(Z)
# ei[sigma == 0.0] = 0.0
pf = self.probability_feasibility_multi_gp(X,self.model_c).reshape(-1,1)
pf[pf<0.50] = 0
return -(mu *pf )
def probability_feasibility_multi_gp(self, x, model, mean=None, cov=None, grad=False, l=0):
# print("model",model.output)
x = np.atleast_2d(x)
if grad == False:
Fz = []
for m in range(model.output_dim):
Fz.append(self.probability_feasibility( x, model.output[m], grad, l))
Fz = np.product(Fz,axis=0)
return Fz
else:
Fz = []
grad_Fz = []
for m in range(model.output_dim):
# print("model.output[m]",model.output[m])
# print("mean[m]",mean[m])
# print("cov[m]",cov[m])
Fz_aux, grad_Fz_aux = self.probability_feasibility( x, model.output[m])
Fz.append(Fz_aux)
grad_Fz.append(grad_Fz_aux)
# print("np.array(Fz)", np.array(Fz), "grad_Fz", np.array(grad_Fz))
grad_Fz = self.product_gradient_rule(func = np.array(Fz), grad = np.array(grad_Fz))
# print("output grad_Fz", grad_Fz)
Fz = np.product(Fz, axis=0)
return Fz, grad_Fz
def probability_feasibility(self, x, model, mean=None, cov=None, grad=False, l=0):
model = model.model
# kern = model.kern
# X = model.X
mean, cov = model.predict(x, full_cov=True)
var = np.diag(cov).reshape(-1, 1)
std = np.sqrt(var).reshape(-1, 1)
aux_var = np.reciprocal(var)
mean = mean.reshape(-1, 1)
norm_dist = norm(mean, std)
fz = norm_dist.pdf(l)
Fz = norm_dist.cdf(l)
if grad == True:
grad_mean , grad_var = model.predictive_gradients(x)
grad_std = (1/2.0)*grad_var
# cov = kern.K(X, X) + np.eye(X.shape[0]) * 1e-3
# L = scipy.linalg.cholesky(cov, lower=True)
# u = scipy.linalg.solve(L, np.eye(X.shape[0]))
# Ainv = scipy.linalg.solve(L.T, u)
dims = range(x.shape[1])
grad_Fz = []
for d in dims:
# K1 = np.diag(np.dot(np.dot(kern.dK_dX(x, X, d), Ainv), kern.dK_dX2(X, x, d)))
# K2 = np.diag(kern.dK2_dXdX2(x, x, d, d))
# var_grad = K2 - K1
# var_grad = var_grad.reshape(-1, 1)
grd_mean_d = grad_mean[:, d].reshape(-1, 1)
grd_std_d = grad_std[:, d].reshape(-1, 1)
# print("grd_mean ",grd_mean_d, "var_grad ",grd_std_d )
# print("std",std)
# print("aux_var", aux_var)
# print("fz",fz)
grad_Fz.append(fz * aux_var * (mean * grd_std_d - grd_mean_d * std))
grad_Fz = np.stack(grad_Fz, axis=1)
return Fz.reshape(-1, 1), grad_Fz[:, :, 0]
else:
return Fz.reshape(-1, 1)
def evaluate_objective(self):
"""
Evaluates the objective
"""
print(1)
print(self.suggested_sample)
self.Y_new, cost_new = self.objective.evaluate(self.suggested_sample)
self.C_new, C_cost_new = self.constraint.evaluate(self.suggested_sample)
self.cost.update_cost_model(self.suggested_sample, cost_new)
for j in range(self.objective.output_dim):
print(self.Y_new[j])
self.Y[j] = np.vstack((self.Y[j],self.Y_new[j]))
for k in range(self.constraint.output_dim):
print(self.C_new[k])
self.C[k] = np.vstack((self.C[k],self.C_new[k]))
def compute_current_best(self):
current_acqX = self.acquisition.current_compute_acq()
return current_acqX
def _distance_last_evaluations(self):
"""
Computes the distance between the last two evaluations.
"""
return np.sqrt(sum((self.X[self.X.shape[0]-1,:]-self.X[self.X.shape[0]-2,:])**2))
def _compute_next_evaluations(self, pending_zipped_X=None, ignored_zipped_X=None, re_use=False):
"""
Computes the location of the new evaluation (optimizes the acquisition in the standard case).
:param pending_zipped_X: matrix of input configurations that are in a pending state (i.e., do not have an evaluation yet).
:param ignored_zipped_X: matrix of input configurations that the user black-lists, i.e., those configurations will not be suggested again.
:return:
"""
## --- Update the context if any
self.acquisition.optimizer.context_manager = ContextManager(self.space, self.context)
aux_var = self.evaluator.compute_batch(duplicate_manager=None, re_use=re_use)
### We zip the value in case there are categorical variables
return self.space.zip_inputs(aux_var[0])
#return initial_design('random', self.space, 1)
def _update_model(self):
"""
Updates the model (when more than one observation is available) and saves the parameters (if available).
"""
if (self.num_acquisitions%self.model_update_interval)==0:
### --- input that goes into the model (is unziped in case there are categorical variables)
X_inmodel = self.space.unzip_inputs(self.X)
Y_inmodel = list(self.Y)
C_inmodel = list(self.C)
self.model.updateModel(X_inmodel, Y_inmodel)
self.model_c.updateModel(X_inmodel, C_inmodel)
### --- Save parameters of the model
#self._save_model_parameter_values()
def get_evaluations(self):
return self.X.copy(), self.Y.copy()
class REMBO(BO):
"""
Args:
f (function): function to optimize.
domain (list | None): the description of the inputs variables
(See GpyOpt.core.space.Design_space class for details)
constraints (list | None): the description of the problem constraints
(See GpyOpt.core.space.Design_space class for details)
Attributes:
initial_design_numdata (int):
initial_design_type (string):
domain (dict | None):
constraints (dict | None):
space (Design_space):
model (BOModel):
acquisition (AcquisitionBase):
cost (CostModel):
"""
def __init__(self, f, domain=None, constraints=None, cost_withGradients=None, X=None,
Y=None, subspace_dim_size=0,
model_type='GP', initial_design_numdata=1, initial_design_type='random', acquisition_type='LCB',
normalize_Y=True, exact_feval=False, acquisition_optimizer_type='lbfgs', model_update_interval=1,
evaluator_type='sequential', batch_size=1, maximize=False, de_duplication=False):
if model_type == 'input_warped_GP':
raise NotImplementedError('input_warped_GP model is not implemented')
if acquisition_type in ['EI_MCMC', 'MPI_MCMC', 'LCB_MCMC']:
raise NotImplementedError('MCMC is not implemented')
if batch_size is not 1 or evaluator_type is not 'sequential':
raise NotImplementedError('only sequential evaluation is implemented')
if cost_withGradients is not None:
raise NotImplementedError('param cost is not implemented')
if constraints is not None:
raise NotImplementedError('param constraints is not implemented')
# private field
self._arguments_mng = ArgumentsManager(kwargs=dict())
self.subspace_dim_size = subspace_dim_size
self.cost_withGradients = cost_withGradients
self.initial_design_numdata = initial_design_numdata
self.initial_design_type = initial_design_type
self.model_type = model_type
self.acquisition_type = acquisition_type
self.evaluator_type = evaluator_type
self.model_update_interval = model_update_interval
self.maximize = maximize
self.normalize_Y = normalize_Y
self.de_duplication = de_duplication
self.subspace_dim_size = subspace_dim_size
self.original_domain = domain
# --- property injected in other methods.
self.verbosity = False
self.subspace = None
self.max_time = None
self.max_iter = None
self.cum_time = None
self.report_file = None
self.evaluations_file = None
self.models_file = None
self.eps = None
self.save_models_parameters = None
# --- unnecessary property
self.suggested_sample = None
self.Y_new = None
# --- BO class property in uncertain use
self.num_cores = 1
self.objective = SingleObjective(self._sign(f), batch_size, f.get_function_name())
self.cost = CostModel(cost_withGradients=cost_withGradients)
self.space = initialize_space(domain=domain, constraints=constraints)
self.embedding_matrix = np.random.normal(size=(self.dimensionality, subspace_dim_size))
subspace_domain = self.choose_subspace_domain(subspace_dim_size=subspace_dim_size)
self.subspace = initialize_space(domain=subspace_domain, constraints=constraints)
self.model = self._arguments_mng.model_creator(
model_type=self.model_type, exact_feval=exact_feval, space=self.subspace)
self.acquisition = self._arguments_mng.acquisition_creator(
acquisition_type=self.acquisition_type, model=self.model, space=self.subspace,
acquisition_optimizer=AcquisitionOptimizer(space=self.space, optimizer=acquisition_optimizer_type),
cost_withGradients=self.cost_withGradients
)
self.evaluator = self._arguments_mng.evaluator_creator(
evaluator_type=self.evaluator_type, acquisition=self.acquisition,
batch_size=self.batch_size, model_type=self.model_type, model=self.model,
space=self.space, acquisition_optimizer=self.acquisition_optimizer
)
self.X = X
self.Y = Y
self._set_initial_values()
super().__init__(
model=self.model,
space=self.subspace,
objective=self.objective,
acquisition=self.acquisition,
evaluator=self.evaluator,
X_init=self.X,
Y_init=self.Y,
cost=self.cost,
normalize_Y=self.normalize_Y,
model_update_interval=self.model_update_interval,
de_duplication=self.de_duplication
)
@property
def dimensionality(self):
return self.space.objective_dimensionality
@property
def subspace_domain(self):
return self.subspace.config_space
@property
def domain(self):
return self.space.config_space
@property
def constraints(self):
return self.space.constraints
@property
def f(self):
return self.objective.func
@property
def cost_type(self):
return self.cost.cost_withGradients
@property
def exact_feval(self):
return self.model.exact_feval
@property
def acquisition_optimizer_type(self):
return self.acquisition_optimizer.optimizer_name
@property
def acquisition_optimizer(self):
return self.acquisition.optimizer
@property
def batch_size(self):
return self.objective.n_procs
@property
def objective_name(self):
return self.objective.objective_name
def choose_subspace_domain(self, subspace_dim_size):
subspace_domain = list()
for i in range(subspace_dim_size):
subspace_domain.append({
'name': 'x' + str(i),
'type': 'continuous',
'domain': (-np.sqrt(self.dimensionality), np.sqrt(self.dimensionality)),
'dimensionality': 1
})
return subspace_domain
def map_to_original_space(self, x):
if x.shape != (1, self.subspace_dim_size):
raise ValueError('x.shape is not correct ' + str(x.shape))
x = np.array([np.dot(self.embedding_matrix, el) for el in x])
return x
def run_optimization(self, max_iter=0, max_time=np.inf, eps=1e-8, context=None,
verbosity=False, save_models_parameters=True, report_file=None,
evaluations_file=None, models_file=None):
if self.objective is None:
raise ValueError("Cannot run the optimization loop without the objective function")
# --- Save the options to print and save the results
self.verbosity = verbosity
self.save_models_parameters = save_models_parameters
self.report_file = report_file
self.evaluations_file = evaluations_file
self.models_file = models_file
self.model_parameters_iterations = None
self.context = context
# --- Check if we can save the model parameters in each iteration
if self.save_models_parameters:
if not (isinstance(self.model, GPModel)):
print('Models printout after each iteration is only available for GP and GP_MCMC models')
self.save_models_parameters = False
# --- Setting up stop conditions
self.eps = eps
if (max_iter is None) and (max_time is None):
self.max_iter = 0
self.max_time = np.inf
elif (max_iter is None) and (max_time is not None):
self.max_iter = np.inf
self.max_time = max_time
elif (max_iter is not None) and (max_time is None):
self.max_iter = max_iter
self.max_time = np.inf
else:
self.max_iter = max_iter
self.max_time = max_time
# --- Initialize iterations and running time
stopwatch = StopWatch()
self.num_acquisitions = self.initial_design_numdata
self.suggested_sample = self.X
self.Y_new = self.Y
self._compute_results()
self._run_optimization()
self.cum_time = stopwatch.passed_time()
# --- Stop messages and execution time
self._compute_results()
# --- Print the desired result in files
if self.report_file is not None:
self.save_report(self.report_file)
if self.evaluations_file is not None:
self.save_evaluations(self.evaluations_file)
if self.models_file is not None:
self.save_models(self.models_file)
self._save()
def _run_optimization(self):
while True:
print('.')
# --- update model
try:
self.update()
except np.linalg.LinAlgError:
print('np.linalg.LinAlgError')
break
if self.num_acquisitions >= self.max_iter:
break
self.next_point()
# --- Update current evaluation time and function evaluations
self.num_acquisitions += 1
def next_point(self):
self.suggested_sample = self._compute_next_evaluations()
# --- Augment X
self.X = np.vstack((self.X, self.suggested_sample))
# --- Evaluate *f* in X, augment Y and update cost function (if needed)
self.evaluate_objective()
def evaluate_objective(self):
"""
Evaluates the objective
"""
original_suggested_sample = self.map_to_original_space(x=self.suggested_sample)
self.Y_new, cost_new = self.objective.evaluate(original_suggested_sample)
self.cost.update_cost_model(self.suggested_sample, cost_new)
self.Y = np.vstack((self.Y, self.Y_new))
def get_best_point(self):
self._compute_results()
return self.x_opt, self.fx_opt
def update(self):
self._update_model(self.normalization_type)
self._update_acquisition()
self._update_evaluator()
def _sign(self, f):
if self.maximize:
f_copy = f
def f(x): return -f_copy(x)
return f
def _set_initial_values(self):
if self.X is None:
self.X = initial_design(self.initial_design_type, self.subspace, self.initial_design_numdata)
self.Y, _ = self.objective.evaluate(self.map_to_original_space(x=self.X))
elif self.X is not None and self.Y is None:
self.Y, _ = self.objective.evaluate(self.map_to_original_space(x=self.X))
# save initial values
self.initial_X = deepcopy(self.X)
if self.maximize:
self.initial_Y = -deepcopy(self.Y)
else:
self.initial_Y = deepcopy(self.Y)
def _update_acquisition(self):
self.acquisition = self._arguments_mng.acquisition_creator(
acquisition_type=self.acquisition_type, model=self.model, space=self.subspace,
acquisition_optimizer=AcquisitionOptimizer(space=self.subspace, optimizer=self.acquisition_optimizer_type),
cost_withGradients=self.cost_withGradients
)
def _update_evaluator(self):
self.evaluator.acquisition = self.acquisition
def _update_model(self, normalization_type='stats'):
if self.num_acquisitions % self.model_update_interval == 0:
self.model = self._arguments_mng.model_creator(
model_type=self.model_type, exact_feval=self.exact_feval, space=self.subspace)
X_inmodel, Y_inmodel = self._input_data(normalization_type=normalization_type)
self.model.updateModel(X_inmodel, Y_inmodel, None, None)
self.X_inmodel = X_inmodel
self.Y_inmodel = Y_inmodel
# Save parameters of the model
self._save_model_parameter_values()
def _input_data(self, normalization_type):
# input that goes into the model (is unziped in case there are categorical variables)
X_inmodel = self.subspace.unzip_inputs(self.X)
# Y_inmodel is the output that goes into the model
if self.normalize_Y:
Y_inmodel = normalize(self.Y, normalization_type)
else:
Y_inmodel = self.Y
return X_inmodel, Y_inmodel
def _compute_next_evaluations(self, pending_zipped_X=None, ignored_zipped_X=None):
# --- Update the context if any
self.acquisition.optimizer.context_manager = ContextManager(self.subspace, self.context)
# --- Activate de_duplication
if self.de_duplication:
duplicate_manager = DuplicateManager(
space=self.subspace, zipped_X=self.X, pending_zipped_X=pending_zipped_X,
ignored_zipped_X=ignored_zipped_X)
else:
duplicate_manager = None
# We zip the value in case there are categorical variables
suggested_ = self.subspace.zip_inputs(self.evaluator.compute_batch(
duplicate_manager=duplicate_manager,
context_manager=self.acquisition.optimizer.context_manager))
return suggested_
def _save(self):
mkdir_when_not_exist(abs_path=definitions.ROOT_DIR + '/storage/' + self.objective_name)
dir_name = definitions.ROOT_DIR + '/storage/' + self.objective_name + '/' + now_str() + ' ' + str(
len(self.original_domain)) + 'D REMBO_' + str(self.subspace_dim_size)
mkdir_when_not_exist(abs_path=dir_name)
self.save_report(report_file=dir_name + '/report.txt')
self.save_evaluations(evaluations_file=dir_name + '/evaluation.csv')
self.save_models(models_file=dir_name + '/model.csv')
def save_report(self, report_file=None):
with open(report_file, 'w') as file:
import GPyOpt
import time
file.write(
'-----------------------------' + ' GPyOpt Report file ' + '-----------------------------------\n')
file.write('GPyOpt Version ' + str(GPyOpt.__version__) + '\n')
file.write('Date and time: ' + time.strftime("%c") + '\n')
if self.num_acquisitions == self.max_iter:
file.write('Optimization completed: ' + 'YES, ' + str(self.X.shape[0]).strip(
'[]') + ' samples collected.\n')
file.write('Number initial samples: ' + str(self.initial_design_numdata) + ' \n')
else:
file.write('Optimization completed: ' + 'NO,' + str(self.X.shape[0]).strip(
'[]') + ' samples collected.\n')
file.write('Number initial samples: ' + str(self.initial_design_numdata) + ' \n')
file.write('Tolerance: ' + str(self.eps) + '.\n')
file.write('Optimization time: ' + str(self.cum_time).strip('[]') + ' seconds.\n')
file.write('\n')
file.write(
'--------------------------------' + ' Problem set up ' + '------------------------------------\n')
file.write('Problem name: ' + self.objective_name + '\n')
file.write('Problem dimension: ' + str(self.dimensionality) + '\n')
file.write('Number continuous variables ' + str(len(self.space.get_continuous_dims())) + '\n')
file.write('Number discrete variables ' + str(len(self.space.get_discrete_dims())) + '\n')
file.write('Number bandits ' + str(self.space.get_bandit().shape[0]) + '\n')
file.write('Noiseless evaluations: ' + str(self.exact_feval) + '\n')
file.write('Cost used: ' + self.cost.cost_type + '\n')
file.write('Constraints: ' + str(self.constraints == True) + '\n')
file.write('Subspace Dimension: ' + str(self.subspace_dim_size) + '\n')
file.write('\n')
file.write(
'------------------------------' + ' Optimization set up ' + '---------------------------------\n')
file.write('Normalized outputs: ' + str(self.normalize_Y) + '\n')
file.write('Model type: ' + str(self.model_type).strip('[]') + '\n')
file.write('Model update interval: ' + str(self.model_update_interval) + '\n')
file.write('Acquisition type: ' + str(self.acquisition_type).strip('[]') + '\n')
file.write(
'Acquisition optimizer: ' + str(self.acquisition_optimizer.optimizer_name).strip('[]') + '\n')
file.write('Acquisition type: ' + str(self.acquisition_type).strip('[]') + '\n')
if hasattr(self, 'acquisition_optimizer') and hasattr(self.acquisition_optimizer, 'optimizer_name'):
file.write(
'Acquisition optimizer: ' + str(self.acquisition_optimizer.optimizer_name).strip('[]') + '\n')
else:
file.write('Acquisition optimizer: None\n')
file.write('Evaluator type (batch size): ' + str(self.evaluator_type).strip('[]') + ' (' + str(
self.batch_size) + ')' + '\n')
file.write('Cores used: ' + str(self.num_cores) + '\n')
file.write('\n')
file.write(
'---------------------------------' + ' Summary ' + '------------------------------------------\n')
file.write('Initial X: ' + str(self.initial_X) + '\n')
file.write('Initial Y: ' + str(self.initial_Y) + '\n')
if self.maximize:
file.write('Value at maximum: ' + str(format(-min(self.Y)[0], '.20f')).strip('[]') + '\n')
file.write('Best found maximum location: ' + str(self.X[np.argmin(self.Y), :]).strip('[]') + '\n')
else:
file.write('Value at minimum: ' + str(format(min(self.Y)[0], '.20f')).strip('[]') + '\n')
file.write('Best found minimum location: ' + str(self.X[np.argmin(self.Y), :]).strip('[]') + '\n')
file.write(
'----------------------------------------------------------------------------------------------\n')
file.close()
class CBO(object):
"""
Runner of the multi-attribute Bayesian optimization loop. This class wraps the optimization loop around the different handlers.
:param model: GPyOpt model class.
:param space: GPyOpt space class.
:param objective: GPyOpt objective class.
:param acquisition: GPyOpt acquisition class.
:param evaluator: GPyOpt evaluator class.
:param X_init: 2d numpy array containing the initial inputs (one per row) of the model.
:param Y_init: 2d numpy array containing the initial outputs (one per row) of the model.
:param cost: GPyOpt cost class (default, none).
:param normalize_Y: whether to normalize the outputs before performing any optimization (default, True).
:param model_update_interval: interval of collected observations after which the model is updated (default, 1).
:param de_duplication: GPyOpt DuplicateManager class. Avoids re-evaluating the objective at previous, pending or infeasible locations (default, False).
"""
def __init__(self, model, space, objective, acquisition, evaluator, X_init, Y_init=None, cost=None, normalize_Y = False, model_update_interval = 1, expectation_utility=None):
self.model = model
self.space = space
self.objective = objective
self.acquisition = acquisition
self.utility = acquisition.utility
self.expectation_utility = expectation_utility
self.evaluator = evaluator
self.X = X_init
self.Y = Y_init
self.normalize_Y = normalize_Y
self.cost = CostModel(cost)
self.model_update_interval = model_update_interval
self.historical_optimal_values = []
self.historical_time = []
self.n_attributes = self.model.output_dim
self.n_hyps_samples = min(1, self.model.number_of_hyps_samples())
self.n_parameter_samples = 10
self.full_parameter_support = self.utility.parameter_dist.use_full_support
self.evaluation_optimizer = GPyOpt.optimization.GeneralOptimizer(optimizer='lbfgs', space=space)
#
self.context = None
self.current_argmax = np.atleast_2d(X_init[0,:])
def _current_max_value(self):
"""
Computes E_n[U(f(x_max))|f], where U is the utility function, f is the true underlying ojective function and x_max = argmax E_n[U(f(x))|U]. See
function _marginal_max_value_so_far below.
"""
val = 0
if self.full_parameter_support:
utility_param_support = self.utility.parameter_dist.support
utility_param_dist = self.utility.parameter_dist.prob_dist
for i in range(len(utility_param_support)):
marginal_argmax = self._current_marginal_argmax(utility_param_support[i])
marginal_max_val = np.reshape(self.objective.evaluate(marginal_argmax)[0],(self.n_attributes,))
val += self.utility.eval_func(utility_param_support[i],marginal_max_val)*utility_param_dist[i]
else:
utility_param_samples = self.utility.parameter_dist.sample(self.n_parameter_samples)
for i in range(len(utility_param_samples)):
marginal_argmax = self._current_marginal_argmax(utility_param_samples[i])
marginal_max_val = np.reshape(self.objective.evaluate(marginal_argmax)[0],(self.n_attributes,))
val += self.utility.eval_func(utility_param_samples[i],marginal_max_val)
val /= len(utility_param_samples)
print('Current optimal value: {}'.format(val))
return val
def _current_max_value_and_var(self):
val = 0
var = 0
support = self.utility.parameter_dist.support
utility_dist = self.utility.parameter_dist.prob_dist
for i in range(len(support)):
marginal_argmax = self._current_marginal_argmax(support[i])
marginal_max_val = np.reshape(self.objective.evaluate(marginal_argmax)[0],(self.n_attributes,))
var_marginal_argmax = np.reshape(self.model.posterior_variance_noiseless(marginal_argmax),(self.n_attributes,))
var += self.utility.eval_func(support[i],var_marginal_argmax)*utility_dist[i]
val += self.utility.eval_func(support[i],marginal_max_val)*utility_dist[i]
print('Current optimal value: {}'.format(val))
return val, var
def _current_max_value_parallel(self):
"""
Computes E_n[U(f(x_max))|f], where U is the utility function, f is the true underlying ojective function and x_max = argmax E_n[U(f(x))|U]. See
function _marginal_max_value_so_far below.
"""
pool = Pool(4)
utility_param_samples = self.utility.parameter_dist.sample(self.n_parameter_samples)
val = sum(pool.map(self._current_marginal_max_value, utility_param_samples))/self.n_parameter_samples
print('Current optimal value')
print(val)
return val
def _current_marginal_max_value(self, parameter):
marginal_argmax = self._current_marginal_argmax(parameter)
marginal_max_val = np.reshape(self.objective.evaluate(marginal_argmax)[0], (self.objective.output_dim,))
return self.utility.eval_func(parameter, marginal_max_val)
def _current_marginal_argmax(self, parameter):
"""
Computes argmax E_n[U(f(x))|U] (The abuse of notation can be misleading; note that the expectation is with
respect to the posterior distribution on f after n evaluations)
"""
if self.utility.linear:
def val_func(X):
X = np.atleast_2d(X)
muX = self.model.posterior_mean(X)
valX = np.reshape(np.matmul(parameter, muX), (X.shape[0], 1))
return -valX
def val_func_with_gradient(X):
X = np.atleast_2d(X)
muX = self.model.posterior_mean(X)
dmu_dX = self.model.posterior_mean_gradient(X)
valX = np.reshape(np.matmul(parameter, muX), (X.shape[0], 1))
dval_dX = np.tensordot(parameter, dmu_dX, axes=1)
return -valX, -dval_dX
elif self.expectation_utility is not None:
# Note: the value of these functions is not normalized, i.e. it is not divided by the number of Z samples and GP hyps.
def val_func(X):
X = np.atleast_2d(X)
func_val = np.empty((X.shape[0], 1))
for h in range(self.n_hyps_samples):
self.model.set_hyperparameters(h)
mean, var = self.model.predict_noiseless(X)
for i in range(X.shape[0]):
func_val[i,0] = self.expectation_utility.func(parameter, mean[:,i], var[:,i])
return -func_val
def val_func_with_gradient(X):
X = np.atleast_2d(X)
func_val = np.zeros((X.shape[0], 1))
func_gradient = np.zeros(X.shape)
for h in range(self.n_hyps_samples):
self.model.set_hyperparameters(h)
mean, var = self.model.predict_noiseless(X)
dmean_dX = self.model.posterior_mean_gradient(X)
dvar_dX = self.model.posterior_variance_gradient(X)
aux = np.concatenate((dmean_dX,dvar_dX))
for i in range(X.shape[0]):
func_val[i,0] = self.expectation_utility.func(parameter, mean[:,i], var[:,i])
func_gradient[i,:] = np.matmul(self.expectation_utility.gradient(parameter,mean[:,i],var[:,i]),aux[:,i])
return -func_val, -func_gradient
else:
Z_samples = np.random.normal(size=(50, self.n_attributes))
def val_func(X):
X = np.atleast_2d(X)
func_val = np.zeros((X.shape[0], 1))
for h in range(self.n_hyps_samples):
self.model.set_hyperparameters(h)
mean, var = self.model.predict_noiseless(X)
std = np.sqrt(var)
for i in range(X.shape[0]):
for Z in Z_samples:
func_val[i,0] += self.utility.eval_func(parameter,mean[:,i] + np.multiply(std[:,i],Z))
return -func_val
def val_func_with_gradient(X):
X = np.atleast_2d(X)
func_val = np.zeros((X.shape[0], 1))
func_gradient = np.zeros(X.shape)
for h in range(self.n_hyps_samples):
self.model.set_hyperparameters(h)
mean, var = self.model.predict_noiseless(X)
std = np.sqrt(var)
dmean_dX = self.model.posterior_mean_gradient(X)
dstd_dX = self.model.posterior_variance_gradient(X)
for i in range(X.shape[0]):
for j in range(self.n_attributes):
dstd_dX[j,i,:] /= (2*std[j,i])
for Z in Z_samples:
aux1 = mean[:,i] + np.multiply(Z, std[:,i])
func_val[i,0] += self.utility.eval_func(parameter, aux1)
aux2 = dmean_dX[:,i,:] + np.multiply(dstd_dX[:,i,:].T, Z).T
func_gradient[i,:] += np.matmul(self.utility.eval_gradient(parameter, aux1), aux2)
return -func_val, -func_gradient
argmax = self.evaluation_optimizer.optimize(f=val_func, f_df=val_func_with_gradient, parallel=False)[0]
self.current_argmax = argmax
return argmax
def run_optimization(self, max_iter=1, parallel=False, plot=False, results_file=None, max_time=np.inf, eps=1e-8, context=None, verbosity=False):
"""
Runs Bayesian Optimization for a number 'max_iter' of iterations (after the initial exploration data)
:param max_iter: exploration horizon, or number of acquisitions. If nothing is provided optimizes the current acquisition.
:param max_time: maximum exploration horizon in seconds.
:param eps: minimum distance between two consecutive x's to keep running the model.
:param context: fixes specified variables to a particular context (values) for the optimization run (default, None).
:param verbosity: flag to print the optimization results after each iteration (default, False).
:param evaluations_file: filename of the file where the evaluated points and corresponding evaluations are saved (default, None).
"""
if self.objective is None:
raise InvalidConfigError("Cannot run the optimization loop without the objective function")
# --- Save the options to print and save the results
self.verbosity = verbosity
self.results_file = results_file
self.context = context
# --- Setting up stop conditions
self.eps = eps
if (max_iter is None) and (max_time is None):
self.max_iter = 0
self.max_time = np.inf
elif (max_iter is None) and (max_time is not None):
self.max_iter = np.inf
self.max_time = max_time
elif (max_iter is not None) and (max_time is None):
self.max_iter = max_iter
self.max_time = np.inf
else:
self.max_iter = max_iter
self.max_time = max_time
# --- Initial function evaluation
if self.X is not None and self.Y is None:
self.Y, cost_values = self.objective.evaluate(self.X)
if self.cost.cost_type == 'evaluation_time':
self.cost.update_cost_model(self.X, cost_values)
# --- Initialize model
self.model.updateModel(self.X,self.Y)
# --- Initialize iterations and running time
self.time_zero = time.clock()
self.cum_time = 0
self.num_acquisitions = 0
self.suggested_sample = self.X
self.Y_new = self.Y
# --- Initialize time cost of the evaluations
while (self.max_time > self.cum_time) and (self.num_acquisitions < self.max_iter):
#if not ((self.num_acquisitions < self.max_iter) and (self._distance_last_evaluations() > self.eps)):
tmp = self.suggested_sample
self.suggested_sample = self.compute_next_evaluations()
if np.all(self.suggested_sample == tmp):
self.suggested_sample = self._perturb(self.suggested_sample)
try:
self.acquisition.update_Z_samples()
except:
pass
# --- Augment X
self.X = np.vstack((self.X, self.suggested_sample))
# --- Evaluate *f* in X, augment Y and update cost function (if needed)
print('Acquisition {}'.format(self.num_acquisitions+1))
self.evaluate_objective()
# --- Update model
if (self.num_acquisitions%self.model_update_interval)==0:
self._update_model()
self.model.get_model_parameters_names()
self.model.get_model_parameters()
if parallel and (not self.full_parameter_support):
current_max_val = self._current_max_value_parallel()
else:
current_max_val = self._current_max_value()
self.historical_optimal_values.append(current_max_val)
# --- Update current evaluation time and function evaluations
self.cum_time = time.clock() - self.time_zero
self.historical_time.append(self.cum_time)
self.num_acquisitions += 1
if verbosity:
print("num acquisition: {}, time elapsed: {:.2f}s".format(
self.num_acquisitions, self.cum_time))
if results_file is not None:
self.save_results(results_file)
if plot:
self.plot_convergence(confidence_interval=True)
#self._plot_pareto_front()
# --- Print the desired result in files
#if self.evaluations_file is not None:
#self.save_evaluations(self.evaluations_file)
#file = open('test_file.txt','w')
#plt.plot(range(self.num_acquisitions),value_so_far)
#plt.show()
#np.savetxt('test_file.txt',value_so_far)
def best_evaluated(self):
if self.n_attributes > 1:
raise InvalidConfigError("This option is not avialable with multiple objectives")
else:
scores = self.Y[0].flatten()
x_best = self.X[np.argsort(-scores)[0],:]
fx_best = -np.sort(-scores)
return x_best, fx_best
def evaluate_objective(self):
"""
Evaluates the objective
"""
print('Suggested point to evaluate: {}'.format(self.suggested_sample))
self.Y_new, cost_new = self.objective.evaluate_w_noise(self.suggested_sample)
self.cost.update_cost_model(self.suggested_sample, cost_new)
for j in range(self.n_attributes):
#print(self.Y_new[j])
self.Y[j] = np.vstack((self.Y[j],self.Y_new[j]))
def _distance_last_evaluations(self):
"""
Computes the distance between the last two evaluations.
"""
return np.sqrt(sum((self.X[self.X.shape[0]-1,:]-self.X[self.X.shape[0]-2,:])**2))
def _perturb(self, x):
perturbed_x = np.copy(x)
while np.all(perturbed_x == x):
perturbed_x = x + np.random.normal(size=x.shape, scale=1e-2)
perturbed_x = self.space.round_optimum(perturbed_x)
return perturbed_x
def compute_next_evaluations(self, pending_zipped_X=None, ignored_zipped_X=None):
"""
Computes the location of the new evaluation (optimizes the acquisition in the standard case).
:param pending_zipped_X: matrix of input configurations that are in a pending state (i.e., do not have an evaluation yet).
:param ignored_zipped_X: matrix of input configurations that the user black-lists, i.e., those configurations will not be suggested again.
:return:
"""
# --- Initial function evaluation
if self.X is not None and self.Y is None:
self.Y, cost_values = self.objective.evaluate(self.X)
if self.cost.cost_type == 'evaluation_time':
self.cost.update_cost_model(self.X, cost_values)
# --- Initialize model
self.model.updateModel(self.X, self.Y)
## --- Update the context if any
self.acquisition.optimizer.context_manager = ContextManager(self.space, self.context)
### We zip the value in case there are categorical variables
x_baseline = self.current_argmax + np.random.normal(loc=0.0, scale=0.1, size=self.current_argmax.shape)
return self.space.zip_inputs(self.evaluator.compute_batch(duplicate_manager=None, x_baseline=self.current_argmax))
#return self.space.zip_inputs(self.evaluator.compute_batch(duplicate_manager=None))
#return initial_design('random', self.space, 1)
def _update_model(self):
"""
Updates the model (when more than one observation is available) and saves the parameters (if available).
"""
### --- input that goes into the model (is unziped in case there are categorical variables)
X_inmodel = self.space.unzip_inputs(self.X)
Y_inmodel = list(self.Y)
#print(X_inmodel)
#print(Y_inmodel)
self.model.updateModel(X_inmodel, Y_inmodel)
### --- Save parameters of the model
#self._save_model_parameter_values()
def convergence_assesment(self, n_iter=10, attribute=0, context=None):
if self.objective is None:
raise InvalidConfigError("Cannot run the optimization loop without the objective function")
#self.model_parameters_iterations = None
self.context = context
# --- Initial function evaluation
if self.X is not None and self.Y is None:
self.Y, cost_values = self.objective.evaluate(self.X)
if self.cost.cost_type == 'evaluation_time':
self.cost.update_cost_model(self.X, cost_values)
self._update_model()
for i in range(n_iter):
self.suggested_sample = self.compute_next_evaluations()
filename = './experiments/1d' + str(i) + '.eps'
model_to_plot = deepcopy(self.model)
integrated_plot(self.acquisition.space.get_bounds(),
self.X.shape[1],
model_to_plot,
self.X,
self.Y,
self.acquisition.acquisition_function,
self.suggested_sample,
attribute,
filename)
self.X = np.vstack((self.X,self.suggested_sample))
self.evaluate_objective()
self._update_model()
#self.model.get_model_parameters_names()
#self.model.get_model_parameters()
#print('Acquisition value at previously evaluated points:')
#print(self.acquisition.acquisition_function(self.X))
#print('Posterior mean and variance')
#print(self.model.predict(self.X))
#print(self.Y)
self.historical_optimal_values.append(self._current_max_value())
def one_step_assesment(self, attribute=0, context=None):
"""
"""
if self.objective is None:
raise InvalidConfigError("Cannot run the optimization loop without the objective function")
#self.model_parameters_iterations = None
self.context = context
# --- Initial function evaluation
if self.X is not None and self.Y is None:
self.Y, cost_values = self.objective.evaluate(self.X)
if self.cost.cost_type == 'evaluation_time':
self.cost.update_cost_model(self.X, cost_values)
self._update_model()
self.suggested_sample = self.compute_next_evaluations()
model_to_plot = deepcopy(self.model)
integrated_plot(self.acquisition.space.get_bounds(),
self.X.shape[1],
model_to_plot,
self.X,
self.Y,
self.acquisition.acquisition_function,
self.suggested_sample,
attribute,
None)
self.X = np.vstack((self.X,self.suggested_sample))
self.evaluate_objective()
self._update_model()
self.historical_optimal_values.append(self._current_max_value())
def integrated_plot(self, attribute=0, filename=None):
"""
Plots the model and the acquisition function.
if self.input_dim = 1: Plots data, mean and variance in one plot and the acquisition function in another plot
if self.input_dim = 2: as before but it separates the mean and variance of the model in two different plots
:param filename: name of the file where the plot is saved
"""
from copy import deepcopy
model_to_plot = deepcopy(self.model)
integrated_plot(self.acquisition.space.get_bounds(),
self.X.shape[1],
model_to_plot,
self.X,
self.Y,
self.acquisition.acquisition_function,
self.suggest_next_locations(),
attribute,
filename)
def plot_acquisition(self,filename=None):
"""
Plots the acquisition function.
"""
return plot_acquisition(self.acquisition.space.get_bounds(),
self.X.shape[1],
self.acquisition.acquisition_function,
filename)
def plot_convergence(self, confidence_interval=False, filename=None):
"""
Makes twp plots to evaluate the convergence of the model:
plot 1: Iterations vs. distance between consecutive selected x's
plot 2: Iterations vs. the mean of the current model in the selected sample.
:param filename: name of the file where the plot is saved
"""
return plot_convergence(self.historical_optimal_values, self.var_at_historical_optima, confidence_interval, filename)
def get_evaluations(self):
return self.X.copy(), self.Y.copy()
def save_results(self, filename):
results = np.zeros((len(self.historical_optimal_values),2))
results[:, 0] = np.atleast_1d(self.historical_optimal_values)
results[:, 1] = np.atleast_1d(self.historical_time)
np.savetxt(filename,results)
class BO(object):
"""
Runner of the multi-attribute Bayesian optimization loop. This class wraps the optimization loop around the different handlers.
:param model: GPyOpt model class.
:param space: GPyOpt space class.
:param objective: GPyOpt objective class.
:param acquisition: GPyOpt acquisition class.
:param evaluator: GPyOpt evaluator class.
:param X_init: 2d numpy array containing the initial inputs (one per row) of the model.
:param Y_init: 2d numpy array containing the initial outputs (one per row) of the model.
:param cost: GPyOpt cost class (default, none).
:param normalize_Y: whether to normalize the outputs before performing any optimization (default, True).
:param model_update_interval: interval of collected observations after which the model is updated (default, 1).
:param de_duplication: GPyOpt DuplicateManager class. Avoids re-evaluating the objective at previous, pending or infeasible locations (default, False).
"""
def __init__(self,
model,
model_c,
space,
objective,
constraint,
acquisition,
evaluator,
X_init,
penalty_tag="fixed",
penalty_value=0.0,
expensive=False,
Y_init=None,
C_init=None,
cost=None,
normalize_Y=False,
model_update_interval=1,
true_preference=0.5):
self.true_preference = true_preference
self.model_c = model_c
self.model = model
self.space = space
self.objective = objective
self.constraint = constraint
self.acquisition = acquisition
self.utility = acquisition.utility
self.evaluator = evaluator
self.normalize_Y = normalize_Y
self.model_update_interval = model_update_interval
self.X = X_init
self.Y = Y_init
self.C = C_init
self.cost = CostModel(cost)
self.model_parameters_iterations = None
self.expensive = expensive
self.penalty_tag = penalty_tag
self.penalty_value = penalty_value
def suggest_next_locations(self,
context=None,
pending_X=None,
ignored_X=None):
"""
Run a single optimization step and return the next locations to evaluate the objective.
Number of suggested locations equals to batch_size.
:param context: fixes specified variables to a particular context (values) for the optimization run (default, None).
:param pending_X: matrix of input configurations that are in a pending state (i.e., do not have an evaluation yet) (default, None).
:param ignored_X: matrix of input configurations that the user black-lists, i.e., those configurations will not be suggested again (default, None).
"""
self.model_parameters_iterations = None
self.num_acquisitions = 0
self.context = context
self._update_model(self.normalization_type)
suggested_locations = self._compute_next_evaluations(
pending_zipped_X=pending_X, ignored_zipped_X=ignored_X)
return suggested_locations
# def _value_so_far(self):
# """
# Computes E_n[U(f(x_max))|f], where U is the utility function, f is the true underlying ojective function and x_max = argmax E_n[U(f(x))|U]. See
# function _marginal_max_value_so_far below.
# """
#
# output = 0
# support = self.utility.parameter_dist.support
# utility_dist = self.utility.parameter_dist.prob_dist
#
# a = np.reshape(self.objective.evaluate(self._marginal_max_value_so_far())[0],(self.objective.output_dim,))
#
# output += self.utility.eval_func(support[i],a)*utility_dist[i]
# #print(output)
# return output
#
#
# def _marginal_max_value_so_far(self):
# """
# Computes argmax E_n[U(f(x))|U] (The abuse of notation can be misleading; note that the expectation is with
# respect to the posterior distribution on f after n evaluations)
# """
#
# def val_func(X):
# X = np.atleast_2d(X)
# muX = self.model.posterior_mean(X)
# return muX
#
# def val_func_with_gradient(X):
# X = np.atleast_2d(X)
# muX = self.model.posterior_mean(X)
# dmu_dX = self.model.posterior_mean_gradient(X)
# valX = np.reshape( muX, (X.shape[0],1))
# dval_dX = dmu_dX
# return -valX, -dval_dX
#
#
# argmax = self.acquisition.optimizer.optimize_inner_func(f=val_func, f_df=val_func_with_gradient)[0]
# return argmax
#
#
def run_optimization(self,
max_iter=1,
max_time=np.inf,
eps=1e-8,
context=None,
verbosity=False,
evaluations_file=None):
"""
Runs Bayesian Optimization for a number 'max_iter' of iterations (after the initial exploration data)
:param max_iter: exploration horizon, or number of acquisitions. If nothing is provided optimizes the current acquisition.
:param max_time: maximum exploration horizon in seconds.
:param eps: minimum distance between two consecutive x's to keep running the model.
:param context: fixes specified variables to a particular context (values) for the optimization run (default, None).
:param verbosity: flag to print the optimization results after each iteration (default, False).
:param evaluations_file: filename of the file where the evaluated points and corresponding evaluations are saved (default, None).
"""
self.verbosity = verbosity
if self.objective is None:
raise InvalidConfigError(
"Cannot run the optimization loop without the objective function"
)
# --- Save the options to print and save the results
self.verbosity = verbosity
self.evaluations_file = evaluations_file
self.context = context
# --- Setting up stop conditions
self.eps = eps
if (max_iter is None) and (max_time is None):
self.max_iter = 0
self.max_time = np.inf
elif (max_iter is None) and (max_time is not None):
self.max_iter = np.inf
self.max_time = max_time
elif (max_iter is not None) and (max_time is None):
self.max_iter = max_iter
self.max_time = np.inf
else:
self.max_iter = max_iter
self.max_time = max_time
# print("------------------------TRAINING HYPERS----------------------")
# self._get_hyperparameters()
# --- Initial function evaluation and model fitting
if self.X is not None and self.Y is None:
self.Y, cost_values = self.objective.evaluate(self.X)
if self.constraint is not None:
self.C, cost_values = self.constraint.evaluate(self.X)
if self.cost.cost_type == 'evaluation_time':
self.cost.update_cost_model(self.X, cost_values)
#self.model.updateModel(self.X,self.Y)
# --- Initialize iterations and running time
self.time_zero = time.time()
self.cum_time = 0
self.num_acquisitions = 0
self.suggested_sample = self.X
self.Y_new = self.Y
self.Opportunity_Cost = []
value_so_far = []
# --- Initialize time cost of the evaluations
print("-----------------------MAIN LOOP STARTS----------------------")
Opportunity_Cost = []
self.true_best_stats = {
"true_best": [],
"mean_gp": [],
"std gp": [],
"pf": [],
"mu_pf": [],
"var_pf": [],
"residual_noise": []
}
while (self.max_iter > self.num_acquisitions):
self._update_model()
if self.constraint is None:
self.Opportunity_Cost_caller_unconstrained()
else:
self.Opportunity_Cost_caller_constrained()
print("maKG optimizer")
start = time.time()
self.suggested_sample = self._compute_next_evaluations()
finish = time.time()
print("time optimisation point X", finish - start)
if verbosity:
####plots
design_plot = initial_design('random', self.space, 1000)
ac_f = self.expected_improvement(design_plot)
Y, _ = self.objective.evaluate(design_plot)
C, _ = self.constraint.evaluate(design_plot)
pf = self.probability_feasibility_multi_gp(
design_plot, self.model_c).reshape(-1, 1)
mu_f = self.model.predict(design_plot)[0]
bool_C = np.product(np.concatenate(C, axis=1) < 0, axis=1)
func_val = Y * bool_C.reshape(-1, 1)
print("self.suggested_sample", self.suggested_sample)
fig, axs = plt.subplots(2, 2)
axs[0, 0].set_title('True Function')
axs[0, 0].scatter(design_plot[:, 0],
design_plot[:, 1],
c=np.array(func_val).reshape(-1))
axs[0, 0].scatter(self.X[:, 0],
self.X[:, 1],
color="red",
label="sampled")
axs[0, 0].scatter(self.suggested_sample[:, 0],
self.suggested_sample[:, 1],
marker="x",
color="red",
label="suggested")
axs[0, 1].set_title('approximation Acqu Function')
axs[0, 1].scatter(design_plot[:, 0],
design_plot[:, 1],
c=np.array(ac_f).reshape(-1))
axs[1, 0].set_title("convergence")
axs[1, 0].plot(range(len(self.Opportunity_Cost)),
np.array(self.Opportunity_Cost).reshape(-1))
axs[1, 0].set_yscale("log")
axs[1, 1].set_title("mu")
axs[1, 1].scatter(design_plot[:, 0],
design_plot[:, 1],
c=np.array(mu_f).reshape(-1) *
np.array(pf).reshape(-1))
plt.show()
self.X = np.vstack((self.X, self.suggested_sample))
# --- Evaluate *f* in X, augment Y and update cost function (if needed)
self.evaluate_objective()
print("X", self.X, "Y", self.Y, "C", self.C, "OC",
self.Opportunity_Cost)
# --- Update current evaluation time and function evaluations
self.cum_time = time.time() - self.time_zero
self.num_acquisitions += 1
return self.X, self.Y, self.C, self.Opportunity_Cost
# --- Print the desired result in files
#if self.evaluations_file is not None:
#self.save_evaluations(self.evaluations_file)
#file = open('test_file.txt','w')
#np.savetxt('test_file.txt',value_so_far)
def Opportunity_Cost_caller_constrained(self):
# design_plot = initial_design('random', self.space, 1000)
# ac_f = self.expected_improvement(design_plot)
# fig, axs = plt.subplots(2, 2)
# axs[0, 0].set_title('True Function')
# axs[0, 0].scatter(design_plot[:, 0], design_plot[:, 1], c=np.array(ac_f).reshape(-1))
# axs[0, 0].scatter(suggested_sample[:, 0], suggested_sample[:, 1], color="red")
# plt.show()
# print("self.suggested_sample",suggested_sample)
# --- Evaluate *f* in X, augment Y and update cost function (if needed)
samples = self.X
print("samples", samples)
Y = self.model.posterior_mean(samples)
# print("Y",Y)
pf = self.probability_feasibility_multi_gp(samples, model=self.model_c)
func_val = np.array(Y).reshape(-1) * np.array(pf).reshape(-1)
print("Y", Y, "pf", pf)
suggested_final_sample = samples[np.argmax(func_val)]
suggested_final_sample = np.array(suggested_final_sample).reshape(-1)
suggested_final_sample = np.array(suggested_final_sample).reshape(
1, -1)
print("suggested_final_sample", suggested_final_sample, "max",
np.max(func_val))
Y_true, _ = self.objective.evaluate(suggested_final_sample,
true_val=True)
C_true, _ = self.constraint.evaluate(suggested_final_sample,
true_val=True)
bool_C_true = np.product(np.concatenate(C_true, axis=1) < 0, axis=1)
func_val_true = Y_true * bool_C_true.reshape(-1, 1)
print("suggested_final_sample", suggested_final_sample,
"func_val_true", func_val_true)
if self.expensive:
self.Opportunity_Cost.append(
np.array(np.abs(np.max(func_val_true))).reshape(-1))
else:
self.true_best_value()
optimum = np.max(np.abs(self.true_best_stats["true_best"]))
# print("optimum", optimum)
self.Opportunity_Cost.append(
optimum - np.array(np.abs(np.max(func_val_true))).reshape(-1))
# print("OC_i", optimum - np.array(np.abs(np.max(func_val_true))).reshape(-1))
def Opportunity_Cost_caller_unconstrained(self):
# design_plot = initial_design('random', self.space, 1000)
# ac_f = self.expected_improvement(design_plot)
# fig, axs = plt.subplots(2, 2)
# axs[0, 0].set_title('True Function')
# axs[0, 0].scatter(design_plot[:, 0], design_plot[:, 1], c=np.array(ac_f).reshape(-1))
# axs[0, 0].scatter(suggested_sample[:, 0], suggested_sample[:, 1], color="red")
# plt.show()
# print("self.suggested_sample",suggested_sample)
# --- Evaluate *f* in X, augment Y and update cost function (if needed)
samples = self.X
print("samples", samples)
Y = self.model.posterior_mean(samples)
# print("Y",Y)
#pf = self.probability_feasibility_multi_gp(samples, model=self.model_c)
func_val = np.array(Y).reshape(-1) #* np.array(pf).reshape(-1)
#print("Y", Y, "pf", pf)
suggested_final_sample = samples[np.argmax(func_val)]
suggested_final_sample = np.array(suggested_final_sample).reshape(-1)
suggested_final_sample = np.array(suggested_final_sample).reshape(
1, -1)
print("suggested_final_sample", suggested_final_sample, "max",
np.max(func_val))
Y_true, _ = self.objective.evaluate(suggested_final_sample,
true_val=True)
#C_true, _ = self.constraint.evaluate(suggested_final_sample, true_val=True)
#bool_C_true = np.product(np.concatenate(C_true, axis=1) < 0, axis=1)
func_val_true = Y_true #* bool_C_true.reshape(-1, 1)
print("suggested_final_sample", suggested_final_sample,
"func_val_true", func_val_true)
if self.expensive:
self.Opportunity_Cost.append(
np.array(np.abs(np.max(func_val_true))).reshape(-1))
else:
self.true_best_value()
optimum = np.max(np.abs(self.true_best_stats["true_best"]))
# print("optimum", optimum)
self.Opportunity_Cost.append(
optimum - np.array(np.abs(np.max(func_val_true))).reshape(-1))
# print("OC_i", optimum - np.array(np.abs(np.max(func_val_true))).reshape(-1))
def optimize_final_evaluation(self):
out = self.acquisition.optimizer.optimize(f=self.current_best,
duplicate_manager=None)
print("out", out)
print("out", out[1])
self.best_mu_all = -1 * np.array(out[1]).reshape(-1)
self.acquisition.optimizer.context_manager = ContextManager(
self.space, self.context)
out = self.acquisition.optimizer.optimize(f=self.expected_improvement,
duplicate_manager=None)
suggested_sample = self.space.zip_inputs(out[0])
# suggested_sample = suggested_sample.astype("int")
print("suggested_sample", suggested_sample)
return suggested_sample
def current_best(self, X):
# X = X.astype("int")
mu = self.model.posterior_mean(X)
mu = mu.reshape(-1, 1)
return -mu.reshape(-1)
def expected_improvement(self, X, offset=0.0):
'''
Computes the EI at points X based on existing samples X_sample
and Y_sample using a Gaussian process surrogate model.
Args:
X: Points at which EI shall be computed (m x d).
X_sample: Sample locations (n x d).
Y_sample: Sample values (n x 1).
gpr: A GaussianProcessRegressor fitted to samples.
xi: Exploitation-exploration trade-off parameter.
Returns:
Expected improvements at points X.
'''
# X = X.astype("int")
mu = self.model.posterior_mean(X)
sigma = self.model.posterior_variance(X, noise=False)
sigma = np.sqrt(sigma).reshape(-1, 1)
mu = mu.reshape(-1, 1)
# Needed for noise-based model,
# otherwise use np.max(Y_sample).
# See also section 2.4 in [...]
bool_C = np.product(np.concatenate(self.C, axis=1) < 0, axis=1)
func_val = self.Y * bool_C.reshape(-1, 1)
mu_sample_opt = np.max(func_val) - offset
with np.errstate(divide='warn'):
imp = mu - mu_sample_opt
Z = imp / sigma
ei = imp * norm.cdf(Z) + sigma * norm.pdf(Z)
ei[sigma == 0.0] = 0.0
pf = self.probability_feasibility_multi_gp(X, self.model_c).reshape(
-1, 1)
if self.penalty_tag == "fixed":
self.penalty = self.penalty_value
else:
dif = np.array(mu).reshape(-1) - np.array(
self.best_mu_all).reshape(-1)
constraints = self.model_c.posterior_mean(
X) # MAYBE CHECK FOR SEVERAL CONSTRAINTS
sum_constraints = np.sum(constraints, axis=0).reshape(-1)
self.penalty = dif * sum_constraints
return -(ei.reshape(-1) * pf.reshape(-1) +
(1 - pf.reshape(-1)) * self.penalty)
def probability_feasibility_multi_gp(self,
x,
model,
mean=None,
cov=None,
grad=False,
l=0):
# print("model",model.output)
x = np.atleast_2d(x)
if grad == False:
Fz = []
for m in range(model.output_dim):
Fz.append(
self.probability_feasibility(x, model.output[m], grad, l))
Fz = np.product(Fz, axis=0)
return Fz
else:
Fz = []
grad_Fz = []
for m in range(model.output_dim):
# print("model.output[m]",model.output[m])
# print("mean[m]",mean[m])
# print("cov[m]",cov[m])
Fz_aux, grad_Fz_aux = self.probability_feasibility(
x, model.output[m])
Fz.append(Fz_aux)
grad_Fz.append(grad_Fz_aux)
# print("np.array(Fz)", np.array(Fz), "grad_Fz", np.array(grad_Fz))
grad_Fz = self.product_gradient_rule(func=np.array(Fz),
grad=np.array(grad_Fz))
# print("output grad_Fz", grad_Fz)
Fz = np.product(Fz, axis=0)
return Fz, grad_Fz
def probability_feasibility(self,
x,
model,
mean=None,
cov=None,
grad=False,
l=0):
model = model.model
# kern = model.kern
# X = model.X
mean = model.posterior_mean(x)
var = model.posterior_variance(x, noise=False)
std = np.sqrt(var).reshape(-1, 1)
aux_var = np.reciprocal(var)
mean = mean.reshape(-1, 1)
norm_dist = norm(mean, std)
fz = norm_dist.pdf(l)
Fz = norm_dist.cdf(l)
if grad == True:
grad_mean, grad_var = model.predictive_gradients(x)
grad_std = (1 / 2.0) * grad_var
# cov = kern.K(X, X) + np.eye(X.shape[0]) * 1e-3
# L = scipy.linalg.cholesky(cov, lower=True)
# u = scipy.linalg.solve(L, np.eye(X.shape[0]))
# Ainv = scipy.linalg.solve(L.T, u)
dims = range(x.shape[1])
grad_Fz = []
for d in dims:
# K1 = np.diag(np.dot(np.dot(kern.dK_dX(x, X, d), Ainv), kern.dK_dX2(X, x, d)))
# K2 = np.diag(kern.dK2_dXdX2(x, x, d, d))
# var_grad = K2 - K1
# var_grad = var_grad.reshape(-1, 1)
grd_mean_d = grad_mean[:, d].reshape(-1, 1)
grd_std_d = grad_std[:, d].reshape(-1, 1)
# print("grd_mean ",grd_mean_d, "var_grad ",grd_std_d )
# print("std",std)
# print("aux_var", aux_var)
# print("fz",fz)
grad_Fz.append(fz * aux_var *
(mean * grd_std_d - grd_mean_d * std))
grad_Fz = np.stack(grad_Fz, axis=1)
return Fz.reshape(-1, 1), grad_Fz[:, :, 0]
else:
return Fz.reshape(-1, 1)
def evaluate_objective(self):
"""
Evaluates the objective
"""
print(1)
print(self.suggested_sample)
self.Y_new, cost_new = self.objective.evaluate(self.suggested_sample)
self.C_new, C_cost_new = self.constraint.evaluate(
self.suggested_sample)
self.cost.update_cost_model(self.suggested_sample, cost_new)
for j in range(self.objective.output_dim):
print(self.Y_new[j])
self.Y[j] = np.vstack((self.Y[j], self.Y_new[j]))
for k in range(self.constraint.output_dim):
print(self.C_new[k])
self.C[k] = np.vstack((self.C[k], self.C_new[k]))
def compute_current_best(self):
current_acqX = self.acquisition.current_compute_acq()
return current_acqX
def _distance_last_evaluations(self):
"""
Computes the distance between the last two evaluations.
"""
return np.sqrt(
sum((self.X[self.X.shape[0] - 1, :] -
self.X[self.X.shape[0] - 2, :])**2))
def _compute_next_evaluations(self,
pending_zipped_X=None,
ignored_zipped_X=None,
re_use=False):
"""
Computes the location of the new evaluation (optimizes the acquisition in the standard case).
:param pending_zipped_X: matrix of input configurations that are in a pending state (i.e., do not have an evaluation yet).
:param ignored_zipped_X: matrix of input configurations that the user black-lists, i.e., those configurations will not be suggested again.
:return:
"""
## --- Update the context if any
self.acquisition.optimizer.context_manager = ContextManager(
self.space, self.context)
aux_var = self.evaluator.compute_batch(duplicate_manager=None,
re_use=re_use)
### We zip the value in case there are categorical variables
return self.space.zip_inputs(aux_var[0])
#return initial_design('random', self.space, 1)
def _get_hyperparameters(self):
"""
Updates the model (when more than one observation is available) and saves the parameters (if available).
"""
X_basis = GPyOpt.experiment_design.initial_design(
'latin', self.space, 2000)
X_train = X_basis
# N=4
# for i in range(N - 1):
# X_train = np.concatenate([X_train, X_basis])
Y_train, cost_train = self.objective.evaluate(X_train)
C_train, C_cost_train = self.constraint.evaluate(X_train)
### --- input that goes into the model (is unziped in case there are categorical variables)
X_inmodel = self.space.unzip_inputs(X_train)
Y_inmodel = list(Y_train)
C_inmodel = list(C_train)
print("X", X_inmodel, len(X_inmodel), "Y", Y_inmodel, len(Y_inmodel))
self.model.trainModel(X_inmodel, Y_inmodel)
self.model_c.trainModel(X_inmodel, C_inmodel)
def _update_model(self):
"""
Updates the model (when more than one observation is available) and saves the parameters (if available).
"""
if (self.num_acquisitions % self.model_update_interval) == 0:
### --- input that goes into the model (is unziped in case there are categorical variables)
X_inmodel = self.space.unzip_inputs(self.X)
Y_inmodel = list(self.Y)
if self.constraint is not None:
C_inmodel = list(self.C)
self.model_c.updateModel(X_inmodel, C_inmodel)
print("X", X_inmodel, len(X_inmodel), "Y", Y_inmodel,
len(Y_inmodel))
self.model.updateModel(X_inmodel, Y_inmodel)
### --- Save parameters of the model
#self._save_model_parameter_values()
def get_evaluations(self):
return self.X.copy(), self.Y.copy()
def true_best_value(self):
from scipy.optimize import minimize
X = initial_design('random', self.space, 1000)
fval = self.func_val(X)
anchor_point = np.array(X[np.argmin(fval)]).reshape(-1)
anchor_point = anchor_point.reshape(1, -1)
print("anchor_point", anchor_point)
best_design = minimize(self.func_val,
anchor_point,
method='Nelder-Mead',
tol=1e-8).x
self.true_best_stats["true_best"].append(self.func_val(best_design))
self.true_best_stats["mean_gp"].append(
self.model.posterior_mean(best_design))
self.true_best_stats["std gp"].append(
self.model.posterior_variance(best_design, noise=False))
self.true_best_stats["pf"].append(
self.probability_feasibility_multi_gp(best_design,
self.model_c).reshape(-1, 1))
mean = self.model_c.posterior_mean(best_design)
var = self.model_c.posterior_variance(best_design, noise=False)
residual_noise = self.model_c.posterior_variance(self.X[1],
noise=False)
self.true_best_stats["mu_pf"].append(mean)
self.true_best_stats["var_pf"].append(var)
self.true_best_stats["residual_noise"].append(residual_noise)
if False:
fig, axs = plt.subplots(3, 2)
N = len(np.array(self.true_best_stats["std gp"]).reshape(-1))
GAP = np.array(
np.abs(
np.abs(self.true_best_stats["true_best"]).reshape(-1) -
np.abs(self.true_best_stats["mean_gp"]).reshape(-1))
).reshape(-1)
print("GAP len", len(GAP))
print("N", N)
axs[0, 0].set_title('GAP')
axs[0, 0].plot(range(N), GAP)
axs[0, 0].set_yscale("log")
axs[0, 1].set_title('VAR')
axs[0,
1].plot(range(N),
np.array(self.true_best_stats["std gp"]).reshape(-1))
axs[0, 1].set_yscale("log")
axs[1, 0].set_title("PF")
axs[1, 0].plot(range(N),
np.array(self.true_best_stats["pf"]).reshape(-1))
axs[1, 1].set_title("mu_PF")
axs[1, 1].plot(
range(N),
np.abs(np.array(self.true_best_stats["mu_pf"]).reshape(-1)))
axs[1, 1].set_yscale("log")
axs[2, 1].set_title("std_PF")
axs[2, 1].plot(
range(N),
np.sqrt(np.array(self.true_best_stats["var_pf"]).reshape(-1)))
axs[2, 1].set_yscale("log")
axs[2, 0].set_title("Irreducible noise")
axs[2, 0].plot(
range(N),
np.sqrt(
np.array(
self.true_best_stats["residual_noise"]).reshape(-1)))
axs[2, 0].set_yscale("log")
plt.show()
def func_val(self, x):
if len(x.shape) == 1:
x = x.reshape(1, -1)
Y, _ = self.objective.evaluate(x, true_val=True)
C, _ = self.constraint.evaluate(x, true_val=True)
Y = np.array(Y).reshape(-1)
out = Y.reshape(-1) * np.product(np.concatenate(C, axis=1) < 0,
axis=1).reshape(-1)
out = np.array(out).reshape(-1)
return -out