|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
|
|
6
|
|
|
import numpy as np |
|
7
|
|
|
from scipy.stats import norm |
|
8
|
|
|
|
|
9
|
|
|
|
|
10
|
|
|
from .sbom import SBOM |
|
11
|
|
|
|
|
12
|
|
|
|
|
13
|
|
|
class BayesianOptimizer(SBOM): |
|
14
|
|
|
def __init__(self, _main_args_, _opt_args_): |
|
15
|
|
|
super().__init__(_main_args_, _opt_args_) |
|
16
|
|
|
self.gpr = self._opt_args_.gpr |
|
17
|
|
|
|
|
18
|
|
|
def expected_improvement(self): |
|
19
|
|
|
mu, sigma = self.gpr.predict(self.all_pos_comb) |
|
20
|
|
|
mu_sample, _ = self.gpr.predict(self.X_sample) |
|
21
|
|
|
|
|
22
|
|
|
sigma = sigma.reshape(-1, 1) |
|
23
|
|
|
mu_sample_opt = np.max(mu_sample) |
|
24
|
|
|
|
|
25
|
|
|
imp = mu - mu_sample_opt - self._opt_args_.xi |
|
26
|
|
|
Z = np.divide(imp, sigma, out=np.zeros_like(sigma), where=sigma != 0) |
|
27
|
|
|
exp_imp = imp * norm.cdf(Z) + sigma * norm.pdf(Z) |
|
28
|
|
|
exp_imp[sigma == 0.0] = 0.0 |
|
29
|
|
|
|
|
30
|
|
|
return exp_imp |
|
31
|
|
|
|
|
32
|
|
|
def propose_location(self, cand): |
|
33
|
|
|
self.gpr.fit(self.X_sample, self.Y_sample) |
|
34
|
|
|
exp_imp = self.expected_improvement() |
|
35
|
|
|
exp_imp = exp_imp[:, 0] |
|
36
|
|
|
|
|
37
|
|
|
index_best = list(exp_imp.argsort()[::-1]) |
|
38
|
|
|
|
|
39
|
|
|
all_pos_comb_sorted = self.all_pos_comb[index_best] |
|
40
|
|
|
pos_best = all_pos_comb_sorted[0] |
|
41
|
|
|
|
|
42
|
|
|
return pos_best |
|
43
|
|
|
|
|
44
|
|
|
def _iterate(self, i, _cand_, _p_): |
|
45
|
|
|
_p_.pos_new = self.propose_location(_cand_) |
|
46
|
|
|
_p_.score_new = _cand_.eval_pos(_p_.pos_new) |
|
47
|
|
|
|
|
48
|
|
|
if _p_.score_new > _cand_.score_best: |
|
49
|
|
|
_cand_, _p_ = self._update_pos(_cand_, _p_) |
|
50
|
|
|
|
|
51
|
|
|
self.X_sample = np.vstack((self.X_sample, _p_.pos_new)) |
|
52
|
|
|
self.Y_sample = np.vstack((self.Y_sample, _p_.score_new)) |
|
53
|
|
|
|
|
54
|
|
|
return _cand_ |
|
55
|
|
|
|
SimonBlanke /
Hyperactive
