|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import numpy as np |
|
6
|
|
|
from scipy.stats import norm |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
|
|
from .smbo import SMBO |
|
10
|
|
|
from .surrogate_models import ( |
|
11
|
|
|
GPR_linear, |
|
12
|
|
|
GPR, |
|
13
|
|
|
) |
|
14
|
|
|
from .acquisition_function import ExpectedImprovement |
|
15
|
|
|
|
|
16
|
|
|
|
|
17
|
|
|
gaussian_process = {"gp_nonlinear": GPR(), "gp_linear": GPR_linear()} |
|
18
|
|
|
|
|
19
|
|
|
|
|
20
|
|
|
def normalize(array): |
|
21
|
|
|
num = array - array.min() |
|
22
|
|
|
den = array.max() - array.min() |
|
23
|
|
|
|
|
24
|
|
|
if den == 0: |
|
25
|
|
|
return np.random.random_sample(array.shape) |
|
26
|
|
|
else: |
|
27
|
|
|
return ((num / den) + 0) / 1 |
|
28
|
|
|
|
|
29
|
|
|
|
|
30
|
|
|
class BayesianOptimizer(SMBO): |
|
31
|
|
|
name = "Bayesian Optimization" |
|
32
|
|
|
_name_ = "bayesian_optimization" |
|
33
|
|
|
__name__ = "BayesianOptimizer" |
|
34
|
|
|
|
|
35
|
|
|
optimizer_type = "sequential" |
|
36
|
|
|
computationally_expensive = True |
|
37
|
|
|
|
|
38
|
|
|
def __init__( |
|
39
|
|
|
self, |
|
40
|
|
|
*args, |
|
41
|
|
|
gpr=gaussian_process["gp_nonlinear"], |
|
42
|
|
|
xi=0.03, |
|
43
|
|
|
warm_start_smbo=None, |
|
44
|
|
|
max_sample_size=10000000, |
|
45
|
|
|
sampling={"random": 1000000}, |
|
46
|
|
|
warnings=100000000, |
|
47
|
|
|
**kwargs |
|
48
|
|
|
): |
|
49
|
|
|
super().__init__(*args, **kwargs) |
|
50
|
|
|
self.gpr = gpr |
|
51
|
|
|
self.regr = gpr |
|
52
|
|
|
self.xi = xi |
|
53
|
|
|
self.warm_start_smbo = warm_start_smbo |
|
54
|
|
|
self.max_sample_size = max_sample_size |
|
55
|
|
|
self.sampling = sampling |
|
56
|
|
|
self.warnings = warnings |
|
57
|
|
|
|
|
58
|
|
|
self.init_warm_start_smbo() |
|
59
|
|
|
|
|
60
|
|
|
def finish_initialization(self): |
|
61
|
|
|
self.all_pos_comb = self._all_possible_pos() |
|
62
|
|
|
return super().finish_initialization() |
|
63
|
|
|
|
|
64
|
|
|
def _expected_improvement(self): |
|
65
|
|
|
self.pos_comb = self._sampling(self.all_pos_comb) |
|
66
|
|
|
|
|
67
|
|
|
acqu_func = ExpectedImprovement(self.regr, self.pos_comb, self.xi) |
|
68
|
|
|
return acqu_func.calculate(self.X_sample, self.Y_sample) |
|
69
|
|
|
|
|
70
|
|
|
def _training(self): |
|
71
|
|
|
X_sample = np.array(self.X_sample) |
|
72
|
|
|
Y_sample = np.array(self.Y_sample) |
|
73
|
|
|
|
|
74
|
|
|
Y_sample = normalize(Y_sample).reshape(-1, 1) |
|
75
|
|
|
self.regr.fit(X_sample, Y_sample) |
|
76
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
