|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import random |
|
6
|
|
|
import numpy as np |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
|
|
from . import HillClimbingOptimizer |
|
10
|
|
|
from ...search import Search |
|
11
|
|
|
from scipy.spatial.distance import euclidean |
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
def gaussian(distance, sig, sigma_factor=1): |
|
15
|
|
|
return ( |
|
16
|
|
|
sigma_factor |
|
17
|
|
|
* sig |
|
18
|
|
|
* np.exp(-np.power(distance, 2.0) / (sigma_factor * np.power(sig, 2.0))) |
|
19
|
|
|
) |
|
20
|
|
|
|
|
21
|
|
|
|
|
22
|
|
|
class TabuOptimizer(HillClimbingOptimizer, Search): |
|
23
|
|
|
def __init__(self, search_space, tabu_memory=3): |
|
24
|
|
|
super().__init__(search_space) |
|
25
|
|
|
|
|
26
|
|
|
self.tabus = [] |
|
27
|
|
|
self.tabu_memory = tabu_memory |
|
28
|
|
|
|
|
29
|
|
|
def add_tabu(self, tabu): |
|
30
|
|
|
self.tabus.append(tabu) |
|
31
|
|
|
|
|
32
|
|
|
if len(self.tabus) > self.tabu_memory: |
|
33
|
|
|
self.tabus.pop(0) |
|
34
|
|
|
|
|
35
|
|
|
def _move_climb(self, pos, epsilon_mod=1): |
|
36
|
|
|
sigma = 1 + self.space_dim * self.epsilon * epsilon_mod |
|
37
|
|
|
pos_normal = np.random.normal(pos, sigma, pos.shape) |
|
38
|
|
|
pos_new_int = np.rint(pos_normal) |
|
39
|
|
|
|
|
40
|
|
|
sigma_mod = 1 |
|
41
|
|
|
run = True |
|
42
|
|
|
while run: |
|
43
|
|
|
pos_normal = np.random.normal(pos, sigma * sigma_mod, pos.shape) |
|
44
|
|
|
pos_new_int = np.rint(pos_normal) |
|
45
|
|
|
|
|
46
|
|
|
p_discard_sum = [] |
|
47
|
|
|
for tabu in self.tabus: |
|
48
|
|
|
distance = euclidean(pos_new_int, tabu) |
|
49
|
|
|
sigma_mean = sigma.mean() |
|
50
|
|
|
p_discard = gaussian(distance, sigma_mean) |
|
51
|
|
|
|
|
52
|
|
|
p_discard_sum.append(p_discard) |
|
53
|
|
|
|
|
54
|
|
|
p_discard = np.array(p_discard_sum).sum() |
|
55
|
|
|
rand = random.uniform(0, 1) |
|
56
|
|
|
|
|
57
|
|
|
if p_discard < rand: |
|
58
|
|
|
run = False |
|
59
|
|
|
|
|
60
|
|
|
sigma_mod = sigma_mod * 1.01 |
|
61
|
|
|
|
|
62
|
|
|
n_zeros = [0] * len(self.space_dim) |
|
63
|
|
|
pos = np.clip(pos_new_int, n_zeros, self.space_dim) |
|
64
|
|
|
|
|
65
|
|
|
self.pos_new = pos.astype(int) |
|
66
|
|
|
|
|
67
|
|
|
return self.pos_new |
|
68
|
|
|
|
|
69
|
|
|
def evaluate(self, score_new): |
|
70
|
|
|
super().evaluate(score_new) |
|
71
|
|
|
|
|
72
|
|
|
if score_new < self.score_best: |
|
73
|
|
|
self.add_tabu(self.pos_new) |
|
74
|
|
|
|
|
75
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
