|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import numpy as np |
|
6
|
|
|
|
|
7
|
|
|
from ..smb_opt.smbo import SMBO |
|
8
|
|
|
from ...search import Search |
|
9
|
|
|
|
|
10
|
|
|
from scipy.spatial.distance import cdist |
|
11
|
|
|
|
|
12
|
|
|
|
|
13
|
|
|
class LipschitzFunction: |
|
14
|
|
|
def __init__(self, position_l): |
|
15
|
|
|
self.position_l = position_l |
|
16
|
|
|
|
|
17
|
|
|
def find_best_slope(self, X_sample, Y_sample): |
|
18
|
|
|
slopes = [ |
|
19
|
|
|
abs(y_sample1 - y_sample2) / abs(x_sample1 - x_sample2) |
|
20
|
|
|
for x_sample1, y_sample1 in zip(X_sample, Y_sample) |
|
21
|
|
|
for x_sample2, y_sample2 in zip(X_sample, Y_sample) |
|
22
|
|
|
if y_sample1 is not y_sample2 |
|
23
|
|
|
if np.prod((x_sample1 - x_sample2)) != 0 |
|
24
|
|
|
] |
|
25
|
|
|
|
|
26
|
|
|
if len(slopes) == 0: |
|
27
|
|
|
return 1 |
|
28
|
|
|
return np.max(slopes) |
|
29
|
|
|
|
|
30
|
|
|
def calculate(self, X_sample, Y_sample, score_best): |
|
31
|
|
|
lip_c = self.find_best_slope(X_sample, Y_sample) |
|
32
|
|
|
|
|
33
|
|
|
positions_np = np.array(self.position_l) |
|
34
|
|
|
samples_np = np.array(X_sample) |
|
35
|
|
|
|
|
36
|
|
|
pos_dist = cdist(positions_np, samples_np) * lip_c |
|
37
|
|
|
|
|
38
|
|
|
upper_bound_l = pos_dist |
|
39
|
|
|
upper_bound_l += np.array(Y_sample) |
|
40
|
|
|
|
|
41
|
|
|
mx = np.ma.masked_array(upper_bound_l, mask=upper_bound_l == 0) |
|
42
|
|
|
upper_bound_l = mx.min(1).reshape(1, -1).T |
|
43
|
|
|
upper_bound_l[upper_bound_l <= score_best] = -np.inf |
|
44
|
|
|
|
|
45
|
|
|
return upper_bound_l |
|
46
|
|
|
|
|
47
|
|
|
|
|
48
|
|
|
class LipschitzOptimizer(SMBO, Search): |
|
49
|
|
|
name = "Lipschitz Optimizer" |
|
50
|
|
|
_name_ = "lipschitz_optimizer" |
|
51
|
|
|
__name__ = "LipschitzOptimizer" |
|
52
|
|
|
|
|
53
|
|
|
optimizer_type = "sequential" |
|
54
|
|
|
computationally_expensive = True |
|
55
|
|
|
|
|
56
|
|
|
def __init__(self, *args, **kwargs): |
|
57
|
|
|
super().__init__(*args, **kwargs) |
|
58
|
|
|
|
|
59
|
|
|
def finish_initialization(self): |
|
60
|
|
|
self.all_pos_comb = self._all_possible_pos() |
|
61
|
|
|
return super().finish_initialization() |
|
62
|
|
|
|
|
63
|
|
|
@SMBO.track_new_pos |
|
64
|
|
|
@SMBO.track_X_sample |
|
65
|
|
|
def iterate(self): |
|
66
|
|
|
self.pos_comb = self._sampling(self.all_pos_comb) |
|
67
|
|
|
|
|
68
|
|
|
lip_func = LipschitzFunction(self.pos_comb) |
|
69
|
|
|
upper_bound_l = lip_func.calculate( |
|
70
|
|
|
self.X_sample, self.Y_sample, self.score_best |
|
71
|
|
|
) |
|
72
|
|
|
|
|
73
|
|
|
index_best = list(upper_bound_l.argsort()[::-1]) |
|
74
|
|
|
all_pos_comb_sorted = self.pos_comb[index_best] |
|
75
|
|
|
pos_best = all_pos_comb_sorted[0] |
|
76
|
|
|
|
|
77
|
|
|
return pos_best |
|
78
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
