|
1
|
|
|
# Author: Simon Blanke |
|
2
|
|
|
# Email: [email protected] |
|
3
|
|
|
# License: MIT License |
|
4
|
|
|
|
|
5
|
|
|
import numpy as np |
|
6
|
|
|
|
|
7
|
|
|
from ..base_optimizer import BaseOptimizer |
|
8
|
|
|
|
|
9
|
|
|
from numpy.random import normal, laplace, logistic, gumbel |
|
10
|
|
|
|
|
11
|
|
|
dist_dict = { |
|
12
|
|
|
"normal": normal, |
|
13
|
|
|
"laplace": laplace, |
|
14
|
|
|
"logistic": logistic, |
|
15
|
|
|
"gumbel": gumbel, |
|
16
|
|
|
} |
|
17
|
|
|
|
|
18
|
|
|
|
|
19
|
|
|
def max_list_idx(list_): |
|
20
|
|
|
max_item = max(list_) |
|
21
|
|
|
max_item_idx = [i for i, j in enumerate(list_) if j == max_item] |
|
22
|
|
|
return max_item_idx[-1:][0] |
|
23
|
|
|
|
|
24
|
|
|
|
|
25
|
|
|
class HillClimbingOptimizer(BaseOptimizer): |
|
26
|
|
|
name = "Hill Climbing" |
|
27
|
|
|
_name_ = "hill_climbing" |
|
28
|
|
|
__name__ = "HillClimbingOptimizer" |
|
29
|
|
|
|
|
30
|
|
|
optimizer_type = "local" |
|
31
|
|
|
computationally_expensive = False |
|
32
|
|
|
|
|
33
|
|
|
def __init__( |
|
34
|
|
|
self, *args, epsilon=0.03, distribution="normal", n_neighbours=3, **kwargs |
|
35
|
|
|
): |
|
36
|
|
|
super().__init__(*args, **kwargs) |
|
37
|
|
|
self.epsilon = epsilon |
|
38
|
|
|
self.distribution = distribution |
|
39
|
|
|
self.n_neighbours = n_neighbours |
|
40
|
|
|
|
|
41
|
|
|
def move_climb(self, pos, epsilon_mod=1): |
|
42
|
|
|
while True: |
|
43
|
|
|
sigma = self.conv.max_positions * self.epsilon * epsilon_mod |
|
44
|
|
|
pos_normal = dist_dict[self.distribution](pos, sigma, pos.shape) |
|
45
|
|
|
pos = self.conv2pos(pos_normal) |
|
46
|
|
|
|
|
47
|
|
|
if self.conv.not_in_constraint(pos): |
|
48
|
|
|
return pos |
|
49
|
|
|
epsilon_mod *= 1.01 |
|
50
|
|
|
|
|
51
|
|
|
@BaseOptimizer.track_new_pos |
|
52
|
|
|
@BaseOptimizer.random_iteration |
|
53
|
|
|
def iterate(self): |
|
54
|
|
|
return self.move_climb(self.pos_current) |
|
55
|
|
|
|
|
56
|
|
|
@BaseOptimizer.track_new_score |
|
57
|
|
|
def evaluate(self, score_new): |
|
58
|
|
|
BaseOptimizer.evaluate(self, score_new) |
|
59
|
|
|
if len(self.scores_valid) == 0: |
|
60
|
|
|
return |
|
61
|
|
|
|
|
62
|
|
|
modZero = self.nth_trial % self.n_neighbours == 0 |
|
63
|
|
|
if modZero: |
|
64
|
|
|
score_new_list_temp = self.scores_valid[-self.n_neighbours :] |
|
65
|
|
|
pos_new_list_temp = self.positions_valid[-self.n_neighbours :] |
|
66
|
|
|
|
|
67
|
|
|
idx = max_list_idx(score_new_list_temp) |
|
68
|
|
|
score = score_new_list_temp[idx] |
|
69
|
|
|
pos = pos_new_list_temp[idx] |
|
70
|
|
|
|
|
71
|
|
|
self._eval2current(pos, score) |
|
72
|
|
|
self._eval2best(pos, score) |
|
73
|
|
|
|
SimonBlanke /
Gradient-Free-Optimizers
