|
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
|
|
|
from .smbo import SMBO |
|
10
|
|
|
|
|
11
|
|
|
|
|
12
|
|
|
def normalize(array): |
|
13
|
|
|
num = array - array.min() |
|
14
|
|
|
den = array.max() - array.min() |
|
15
|
|
|
|
|
16
|
|
|
if den == 0: |
|
17
|
|
|
return np.random.random_sample(array.shape) |
|
18
|
|
|
else: |
|
19
|
|
|
return ((num / den) + 0) / 1 |
|
20
|
|
|
|
|
21
|
|
|
|
|
22
|
|
|
class ExpectedImprovementBasedOptimization(SMBO): |
|
23
|
|
|
def __init__( |
|
24
|
|
|
self, |
|
25
|
|
|
*args, |
|
26
|
|
|
xi=0.01, |
|
27
|
|
|
warm_start_smbo=None, |
|
28
|
|
|
sampling={"random": 1000000}, |
|
29
|
|
|
warnings=100000000, |
|
30
|
|
|
**kwargs, |
|
31
|
|
|
): |
|
32
|
|
|
super().__init__(*args, **kwargs) |
|
33
|
|
|
self.new_positions = [] |
|
34
|
|
|
self.xi = xi |
|
35
|
|
|
self.warm_start_smbo = warm_start_smbo |
|
36
|
|
|
self.sampling = sampling |
|
37
|
|
|
self.warnings = warnings |
|
38
|
|
|
|
|
39
|
|
|
def _expected_improvement(self): |
|
40
|
|
|
all_pos_comb = self._all_possible_pos() |
|
41
|
|
|
self.pos_comb = self._sampling(all_pos_comb) |
|
42
|
|
|
|
|
43
|
|
|
mu, sigma = self.regr.predict(self.pos_comb, return_std=True) |
|
44
|
|
|
# mu_sample = self.regr.predict(self.X_sample) |
|
45
|
|
|
mu = mu.reshape(-1, 1) |
|
46
|
|
|
sigma = sigma.reshape(-1, 1) |
|
47
|
|
|
|
|
48
|
|
|
Y_sample = normalize(np.array(self.Y_sample)).reshape(-1, 1) |
|
49
|
|
|
imp = mu - np.max(Y_sample) - self.xi |
|
50
|
|
|
Z = np.divide(imp, sigma, out=np.zeros_like(sigma), where=sigma != 0) |
|
51
|
|
|
|
|
52
|
|
|
exploit = imp * norm.cdf(Z) |
|
53
|
|
|
explore = sigma * norm.pdf(Z) |
|
54
|
|
|
|
|
55
|
|
|
exp_imp = exploit + explore |
|
56
|
|
|
exp_imp[sigma == 0.0] = 0.0 |
|
57
|
|
|
|
|
58
|
|
|
return exp_imp[:, 0] |
|
59
|
|
|
|
|
60
|
|
|
def _propose_location(self): |
|
61
|
|
|
X_sample = np.array(self.X_sample) |
|
62
|
|
|
Y_sample = np.array(self.Y_sample) |
|
63
|
|
|
|
|
64
|
|
|
if len(Y_sample) == 0: |
|
65
|
|
|
return self.move_random() |
|
66
|
|
|
|
|
67
|
|
|
Y_sample = normalize(Y_sample).reshape(-1, 1) |
|
68
|
|
|
self.regr.fit(X_sample, Y_sample) |
|
69
|
|
|
|
|
70
|
|
|
exp_imp = self._expected_improvement() |
|
71
|
|
|
|
|
72
|
|
|
index_best = list(exp_imp.argsort()[::-1]) |
|
73
|
|
|
all_pos_comb_sorted = self.pos_comb[index_best] |
|
74
|
|
|
pos_best = all_pos_comb_sorted[0] |
|
75
|
|
|
|
|
76
|
|
|
return pos_best |
|
77
|
|
|
|
|
78
|
|
|
@SMBO.track_nth_iter |
|
79
|
|
|
@SMBO.track_X_sample |
|
80
|
|
|
def iterate(self): |
|
81
|
|
|
return self._propose_location() |
|
82
|
|
|
|
|
83
|
|
|
@SMBO.track_y_sample |
|
84
|
|
|
def evaluate(self, score_new): |
|
85
|
|
|
self.score_new = score_new |
|
86
|
|
|
|
|
87
|
|
|
self._evaluate_new2current(score_new) |
|
88
|
|
|
self._evaluate_current2best() |
|
89
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
