|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
|
|
6
|
|
|
import random |
|
7
|
|
|
|
|
8
|
|
|
import numpy as np |
|
9
|
|
|
|
|
10
|
|
|
from .base_population_optimizer import BasePopulationOptimizer |
|
11
|
|
|
from ...search import Search |
|
12
|
|
|
from ..local import SimulatedAnnealingOptimizer |
|
13
|
|
|
|
|
14
|
|
|
|
|
15
|
|
|
class ParallelTemperingOptimizer(BasePopulationOptimizer, Search): |
|
16
|
|
|
def __init__(self, search_space, n_iter_swap=10): |
|
17
|
|
|
super().__init__(search_space) |
|
18
|
|
|
|
|
19
|
|
|
self.n_iter_swap = n_iter_swap |
|
20
|
|
|
|
|
21
|
|
|
self.systems = [] |
|
22
|
|
|
|
|
23
|
|
|
def _swap_pos(self): |
|
24
|
|
|
_systems_temp = self.systems[:] |
|
25
|
|
|
|
|
26
|
|
|
for _p1_ in self.systems: |
|
27
|
|
|
rand = random.uniform(0, 1) |
|
28
|
|
|
_p2_ = np.random.choice(_systems_temp) |
|
29
|
|
|
|
|
30
|
|
|
p_accept = self._accept_swap(_p1_, _p2_) |
|
31
|
|
|
if p_accept > rand: |
|
32
|
|
|
temp_temp = _p1_.temp # haha! |
|
33
|
|
|
_p1_.temp = _p2_.temp |
|
34
|
|
|
_p2_.temp = temp_temp |
|
35
|
|
|
|
|
36
|
|
|
def _accept_swap(self, _p1_, _p2_): |
|
37
|
|
|
denom = _p1_.score_current + _p2_.score_current |
|
38
|
|
|
|
|
39
|
|
|
if denom == 0: |
|
40
|
|
|
return 100 |
|
41
|
|
|
elif abs(denom) == np.inf: |
|
42
|
|
|
return 0 |
|
43
|
|
|
else: |
|
44
|
|
|
score_diff_norm = (_p1_.score_current - _p2_.score_current) / denom |
|
45
|
|
|
|
|
46
|
|
|
temp = (1 / _p1_.temp) - (1 / _p2_.temp) |
|
47
|
|
|
return np.exp(score_diff_norm * temp) |
|
48
|
|
|
|
|
49
|
|
|
def init_pos(self, pos): |
|
50
|
|
|
system = SimulatedAnnealingOptimizer(self.space_dim) |
|
51
|
|
|
self.systems.append(system) |
|
52
|
|
|
system.init_pos(pos) |
|
53
|
|
|
|
|
54
|
|
|
self.p_current = system |
|
55
|
|
|
|
|
56
|
|
|
def iterate(self): |
|
57
|
|
|
nth_iter = self._iterations(self.systems) |
|
58
|
|
|
print("nth_iter", nth_iter) |
|
59
|
|
|
self.p_current = self.systems[nth_iter % len(self.systems)] |
|
60
|
|
|
|
|
61
|
|
|
return self.p_current.iterate() |
|
62
|
|
|
|
|
63
|
|
|
def evaluate(self, score_new): |
|
64
|
|
|
nth_iter = self._iterations(self.systems) |
|
65
|
|
|
|
|
66
|
|
|
notZero = self.n_iter_swap != 0 |
|
67
|
|
|
modZero = nth_iter % self.n_iter_swap == 0 |
|
68
|
|
|
|
|
69
|
|
|
if notZero and modZero: |
|
70
|
|
|
self._swap_pos() |
|
71
|
|
|
|
|
72
|
|
|
self.p_current.score_new = score_new |
|
73
|
|
|
|
|
74
|
|
|
self.p_current._evaluate_new2current(score_new) |
|
75
|
|
|
self.p_current._evaluate_current2best() |
|
76
|
|
|
|
|
77
|
|
|
|
|
78
|
|
|
class System: |
|
79
|
|
|
def __init__(self, space_dim, _opt_args_, temp): |
|
80
|
|
|
super().__init__(space_dim, _opt_args_) |
|
81
|
|
|
self.temp = temp |
|
82
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
