|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import random |
|
6
|
|
|
import numpy as np |
|
7
|
|
|
|
|
8
|
|
|
from ..base_optimizer import BaseOptimizer |
|
9
|
|
|
from ..local_opt.hill_climbing_optimizer import HillClimbingOptimizer |
|
10
|
|
|
|
|
11
|
|
|
|
|
12
|
|
|
def max_list_idx(list_): |
|
13
|
|
|
max_item = max(list_) |
|
14
|
|
|
max_item_idx = [i for i, j in enumerate(list_) if j == max_item] |
|
15
|
|
|
return max_item_idx[-1:][0] |
|
16
|
|
|
|
|
17
|
|
|
|
|
18
|
|
|
class PatternSearch(BaseOptimizer): |
|
19
|
|
|
name = "Pattern Search" |
|
20
|
|
|
_name_ = "pattern_search" |
|
21
|
|
|
__name__ = "PatternSearch" |
|
22
|
|
|
|
|
23
|
|
|
optimizer_type = "global" |
|
24
|
|
|
computationally_expensive = False |
|
25
|
|
|
|
|
26
|
|
|
def __init__( |
|
27
|
|
|
self, |
|
28
|
|
|
search_space, |
|
29
|
|
|
initialize={"grid": 4, "random": 2, "vertices": 4}, |
|
30
|
|
|
constraints=[], |
|
31
|
|
|
random_state=None, |
|
32
|
|
|
rand_rest_p=0, |
|
33
|
|
|
nth_process=None, |
|
34
|
|
|
n_positions=4, |
|
35
|
|
|
pattern_size=0.25, |
|
36
|
|
|
reduction=0.9, |
|
37
|
|
|
): |
|
38
|
|
|
super().__init__( |
|
39
|
|
|
search_space=search_space, |
|
40
|
|
|
initialize=initialize, |
|
41
|
|
|
constraints=constraints, |
|
42
|
|
|
random_state=random_state, |
|
43
|
|
|
rand_rest_p=rand_rest_p, |
|
44
|
|
|
nth_process=nth_process, |
|
45
|
|
|
) |
|
46
|
|
|
|
|
47
|
|
|
self.n_positions = n_positions |
|
48
|
|
|
self.pattern_size = pattern_size |
|
49
|
|
|
self.reduction = reduction |
|
50
|
|
|
|
|
51
|
|
|
self.n_positions_ = min(n_positions, self.conv.n_dimensions) |
|
52
|
|
|
self.pattern_size_tmp = pattern_size |
|
53
|
|
|
self.pattern_pos_l = [] |
|
54
|
|
|
|
|
55
|
|
|
def generate_pattern(self, current_position): |
|
56
|
|
|
pattern_pos_l = [] |
|
57
|
|
|
|
|
58
|
|
|
n_valid_pos = len(self.positions_valid) |
|
59
|
|
|
n_pattern_pos = int(self.n_positions_ * 2) |
|
60
|
|
|
n_pos_min = min(n_valid_pos, n_pattern_pos) |
|
61
|
|
|
|
|
62
|
|
|
best_in_recent_pos = any( |
|
63
|
|
|
np.array_equal(np.array(self.pos_best), pos) |
|
64
|
|
|
for pos in self.positions_valid[n_pos_min:] |
|
65
|
|
|
) |
|
66
|
|
|
if best_in_recent_pos: |
|
67
|
|
|
self.pattern_size_tmp *= self.reduction |
|
68
|
|
|
pattern_size = self.pattern_size_tmp |
|
69
|
|
|
|
|
70
|
|
|
for idx, dim_size in enumerate(self.conv.dim_sizes): |
|
71
|
|
|
pos_pattern_p = np.array(current_position) |
|
72
|
|
|
pos_pattern_n = np.array(current_position) |
|
73
|
|
|
|
|
74
|
|
|
pos_pattern_p[idx] += pattern_size * dim_size |
|
75
|
|
|
pos_pattern_n[idx] -= pattern_size * dim_size |
|
76
|
|
|
|
|
77
|
|
|
pos_pattern_p = self.conv2pos(pos_pattern_p) |
|
78
|
|
|
pos_pattern_n = self.conv2pos(pos_pattern_n) |
|
79
|
|
|
|
|
80
|
|
|
pattern_pos_l.append(pos_pattern_p) |
|
81
|
|
|
pattern_pos_l.append(pos_pattern_n) |
|
82
|
|
|
|
|
83
|
|
|
self.pattern_pos_l = list( |
|
84
|
|
|
random.sample(pattern_pos_l, self.n_positions_) |
|
85
|
|
|
) |
|
86
|
|
|
|
|
87
|
|
|
@BaseOptimizer.track_new_pos |
|
88
|
|
|
@BaseOptimizer.random_iteration |
|
89
|
|
|
def iterate(self): |
|
90
|
|
|
while True: |
|
91
|
|
|
pos_new = self.pattern_pos_l[0] |
|
92
|
|
|
self.pattern_pos_l.pop(0) |
|
93
|
|
|
|
|
94
|
|
|
if self.conv.not_in_constraint(pos_new): |
|
95
|
|
|
return pos_new |
|
96
|
|
|
return self.move_climb(pos_new) |
|
97
|
|
|
|
|
98
|
|
|
def finish_initialization(self): |
|
99
|
|
|
self.generate_pattern(self.pos_current) |
|
100
|
|
|
self.search_state = "iter" |
|
101
|
|
|
|
|
102
|
|
|
@BaseOptimizer.track_new_score |
|
103
|
|
|
def evaluate(self, score_new): |
|
104
|
|
|
BaseOptimizer.evaluate(self, score_new) |
|
105
|
|
|
if len(self.scores_valid) == 0: |
|
106
|
|
|
return |
|
107
|
|
|
|
|
108
|
|
|
modZero = self.nth_trial % int(self.n_positions_ * 2) == 0 |
|
109
|
|
|
|
|
110
|
|
|
if modZero or len(self.pattern_pos_l) == 0: |
|
111
|
|
|
if self.search_state == "iter": |
|
112
|
|
|
self.generate_pattern(self.pos_current) |
|
113
|
|
|
|
|
114
|
|
|
score_new_list_temp = self.scores_valid[-self.n_positions_ :] |
|
115
|
|
|
pos_new_list_temp = self.positions_valid[-self.n_positions_ :] |
|
116
|
|
|
|
|
117
|
|
|
idx = max_list_idx(score_new_list_temp) |
|
118
|
|
|
score = score_new_list_temp[idx] |
|
119
|
|
|
pos = pos_new_list_temp[idx] |
|
120
|
|
|
|
|
121
|
|
|
self._eval2current(pos, score) |
|
122
|
|
|
self._eval2best(pos, score) |
|
123
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
