|
1
|
|
|
<?php |
|
2
|
|
|
|
|
3
|
|
|
namespace MachineLearning\Application\Algorithm; |
|
4
|
|
|
|
|
5
|
|
|
use MachineLearning\Domain\Exception\DivergenceException; |
|
6
|
|
|
use MachineLearning\Domain\Model\Dataset; |
|
7
|
|
|
use MachineLearning\Domain\Hypothesis\HypothesisInterface; |
|
8
|
|
|
use MachineLearning\Domain\Model\Result; |
|
9
|
|
|
use MachineLearning\Domain\Model\ValueInterface; |
|
10
|
|
|
use MachineLearning\Domain\Model\Value\VectorValue; |
|
11
|
|
|
|
|
12
|
|
|
class GradientDescendent implements AlgorithmInterface |
|
13
|
|
|
{ |
|
14
|
|
|
const DEFAULT_LEARNING_RATE = 0.01; |
|
15
|
|
|
|
|
16
|
|
|
const DEFAULT_CONVERGENCE_CRITERIA = 0.0000001; |
|
17
|
|
|
|
|
18
|
|
|
const DEFAULT_DIVERGENCE_CRITERIA = 100; |
|
19
|
|
|
|
|
20
|
|
|
/** |
|
21
|
|
|
* @var HypothesisInterface |
|
22
|
|
|
*/ |
|
23
|
|
|
protected $hypothesis; |
|
24
|
|
|
|
|
25
|
|
|
/** |
|
26
|
|
|
* @var double |
|
27
|
|
|
*/ |
|
28
|
|
|
protected $learningRate; |
|
29
|
|
|
|
|
30
|
|
|
/** |
|
31
|
|
|
* @var double |
|
32
|
|
|
*/ |
|
33
|
|
|
protected $convergenceCriteria; |
|
34
|
|
|
|
|
35
|
|
|
/** |
|
36
|
|
|
* @var double |
|
37
|
|
|
*/ |
|
38
|
|
|
protected $divergenceCriteria; |
|
39
|
|
|
|
|
40
|
|
|
/** |
|
41
|
|
|
* @param HypothesisInterface $hypothesis |
|
42
|
|
|
* @param double $learningRate |
|
43
|
|
|
* @param double $convergenceCriteria |
|
44
|
|
|
* @param double $divergenceCriteria |
|
45
|
|
|
*/ |
|
46
|
1 |
|
public function __construct( |
|
47
|
|
|
HypothesisInterface $hypothesis, |
|
48
|
|
|
$learningRate = self::DEFAULT_LEARNING_RATE, |
|
49
|
|
|
$convergenceCriteria = self::DEFAULT_CONVERGENCE_CRITERIA, |
|
50
|
|
|
$divergenceCriteria = self::DEFAULT_DIVERGENCE_CRITERIA |
|
51
|
|
|
) { |
|
52
|
1 |
|
$this->hypothesis = $hypothesis; |
|
53
|
1 |
|
$this->learningRate = (double) $learningRate; |
|
54
|
1 |
|
$this->convergenceCriteria = (double) $convergenceCriteria; |
|
55
|
1 |
|
$this->divergenceCriteria = (double) $divergenceCriteria; |
|
56
|
1 |
|
} |
|
57
|
|
|
|
|
58
|
|
|
/** |
|
59
|
|
|
* @param Dataset $dataset |
|
60
|
|
|
* @return ValueInterface |
|
61
|
|
|
* @throws DivergenceException |
|
62
|
|
|
*/ |
|
63
|
1 |
|
public function train(Dataset $dataset) { |
|
64
|
1 |
|
$convergence = false; |
|
65
|
1 |
|
$divergence = false; |
|
66
|
|
|
|
|
67
|
1 |
|
$firstResult = $dataset->current(); |
|
68
|
1 |
|
$features = count($firstResult->getIndependentVariable()->getValue()); |
|
69
|
1 |
|
$total = count($dataset); |
|
70
|
|
|
|
|
71
|
1 |
|
$coefficientVector = array_fill(0, $features+1, 1); |
|
72
|
1 |
|
$coefficientVector[0] = 0; |
|
73
|
1 |
|
$incrementVector = array_fill(0, $features+1, 0); |
|
74
|
|
|
|
|
75
|
1 |
|
while(!$convergence && !$divergence) { |
|
76
|
1 |
|
list($convergence, $divergence, $coefficientVector) = $this->doStep( |
|
77
|
1 |
|
$dataset, |
|
78
|
1 |
|
$coefficientVector, |
|
79
|
1 |
|
$features, |
|
80
|
1 |
|
$total, |
|
81
|
|
|
$incrementVector |
|
82
|
1 |
|
); |
|
83
|
1 |
|
} |
|
84
|
|
|
|
|
85
|
1 |
|
if ($divergence) { |
|
86
|
|
|
throw new DivergenceException(); |
|
87
|
|
|
} |
|
88
|
|
|
|
|
89
|
1 |
|
return new VectorValue($coefficientVector); |
|
90
|
|
|
} |
|
91
|
|
|
|
|
92
|
|
|
/** |
|
93
|
|
|
* @param Dataset $dataset |
|
94
|
|
|
* @param $coefficientVector |
|
95
|
|
|
* @param $features |
|
96
|
|
|
* @param $total |
|
97
|
|
|
* @param $incrementVector |
|
98
|
|
|
* @return array |
|
99
|
|
|
*/ |
|
100
|
1 |
|
protected function doStep(Dataset $dataset, $coefficientVector, $features, $total, $incrementVector) |
|
101
|
|
|
{ |
|
102
|
1 |
|
$coefficient = new VectorValue($coefficientVector); |
|
103
|
1 |
|
$costVector = array_fill(0, $features + 1, 0); |
|
104
|
|
|
|
|
105
|
1 |
|
foreach ($dataset as $result) { |
|
106
|
1 |
|
$costVector = $this->calculateStepCost($features, $coefficient, $result, $costVector); |
|
107
|
1 |
|
} |
|
108
|
|
|
|
|
109
|
1 |
|
for ($j = 0; $j < $features + 1; $j++) { |
|
110
|
1 |
|
$incrementVector[$j] = $this->learningRate * -(1 / ($total)) * $costVector[$j]; |
|
111
|
1 |
|
} |
|
112
|
|
|
|
|
113
|
1 |
|
$convergence = (bool)(abs(array_sum($incrementVector)) < $this->convergenceCriteria); |
|
114
|
1 |
|
$divergence = (bool)(abs(array_sum($incrementVector)) > $this->divergenceCriteria); |
|
115
|
|
|
|
|
116
|
1 |
|
for ($j = 0; $j < $features + 1; $j++) { |
|
117
|
1 |
|
$coefficientVector[$j] += $incrementVector[$j]; |
|
118
|
1 |
|
} |
|
119
|
1 |
|
return array($convergence, $divergence, $coefficientVector); |
|
120
|
|
|
} |
|
121
|
|
|
|
|
122
|
|
|
/** |
|
123
|
|
|
* @param int $features |
|
124
|
|
|
* @param ValueInterface $coefficient |
|
125
|
|
|
* @param Result $result |
|
126
|
|
|
* @param array $costVector |
|
127
|
|
|
*/ |
|
128
|
1 |
|
protected function calculateStepCost($features, ValueInterface $coefficient, Result $result, array $costVector) |
|
129
|
|
|
{ |
|
130
|
|
|
$firstOrderIncrement = ( |
|
131
|
1 |
|
(float)$this->hypothesis->calculate( |
|
132
|
1 |
|
$coefficient, |
|
133
|
1 |
|
$result->getIndependentVariable() |
|
134
|
1 |
|
)->getValue() |
|
135
|
1 |
|
- (float)$result->getDependentVariable()->getValue() |
|
136
|
1 |
|
); |
|
137
|
1 |
|
$costVector[0] += $firstOrderIncrement; |
|
138
|
|
|
|
|
139
|
1 |
|
for ($j = 1; $j < $features + 1; $j++) { |
|
140
|
|
|
$costVector[$j] += $firstOrderIncrement |
|
141
|
1 |
|
* (float)$this->hypothesis->derivative( |
|
142
|
1 |
|
$coefficient, |
|
143
|
1 |
|
$result->getIndependentVariable(), |
|
144
|
|
|
$j - 1 |
|
145
|
1 |
|
)->getValue(); |
|
146
|
1 |
|
} |
|
147
|
1 |
|
return $costVector; |
|
148
|
|
|
} |
|
149
|
|
|
|
|
150
|
|
|
} |
|
151
|
|
|
|
zeopix /
MachineLearning
