TicTacToe/Brain_for_learning.py at main · DeepDeducing/TicTacToe · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
import numpy as np
from scipy.special import expit


class Brain(object):
    def __init__(self, network_size, slope, alpha, epoch_of_learning, drop_rate, momentum_rate):

        self.network_size                 = network_size
        self.number_of_layers             = self.network_size.shape[0]

        self.slope                        = slope
        self.alpha                        = alpha
        self.epoch_of_learning            = epoch_of_learning

        self.weight_list                  = self.initialize_weight_list()
        self.slope_list                   = self.initialize_slope_list()

        self.weight_list_momentum         = np.zeros_like(self.weight_list)
        self.slope_list_momentum          = np.zeros_like(self.slope_list )

        self.drop_rate                    = drop_rate

        self.momentum_rate                = momentum_rate


    def initialize_weight_list(self):
        weight_list = list()
        for i in range(self.number_of_layers - 1):
            weight                                              = (np.random.random((self.network_size[i]                                 , self.network_size[i+1]                          )) -0.5 ) * 0.1##001 + 0.05   # <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
            weight_list.append(weight)
        weight_list = np.asarray(weight_list)
        return  weight_list


    def initialize_slope_list(self):
        slope_list = list()
        for i in range(self.number_of_layers - 1):
            slope                                               = np.ones(self.network_size[i + 1]) * self.slope
            slope_list.append(slope)
        slope_list = np.asarray(slope_list)
        return slope_list


    def activator(self, x):
        return expit(x)


    def activator_output_to_derivative(self, output):
        return output * ( 1 - output)


    def generate_values_for_each_layer(self, input):

        layer_list                = list()

        layer                     = input

        layer_list.append(layer)

        binomial                  = np.atleast_2d(np.random.binomial(1, 1 - self.drop_rate, size=self.network_size[1]))

        layer                     = self.activator(np.dot(layer_list[-1]                          , self.weight_list[0]                                                          ) * self.slope_list[0] ) * binomial

        layer_list.append(layer)

        for i in range(self.number_of_layers - 3):

            binomial              = np.atleast_2d(np.random.binomial(1, 1 - self.drop_rate, size=self.network_size[i + 2]))

            layer                 = self.activator(np.dot(layer_list[-1]                          , self.weight_list[i + 1]                                                      ) * self.slope_list[i + 1] ) * binomial

            layer_list.append(layer)

        layer                 = self.activator(np.dot(layer_list[-1]                          , self.weight_list[-1]                                                      ) * self.slope_list[-1] )

        layer_list.append(layer)

        return   layer_list


    def train_for_weight(self,
                       layer_list,
                       output):

        layer_delta_list       = list()
        slope_delta_list       = list()

        layer_final_error      = output - layer_list[-1]

        layer_delta            = (layer_final_error                                                                                               )    * self.activator_output_to_derivative(layer_list[-1] )          *   self.slope_list[-1]
        slope_delta            = (layer_final_error                                                                                               )    * self.activator_output_to_derivative(layer_list[-1] )          * np.dot(layer_list[-2], self.weight_list[-1])

        layer_delta_list.append(layer_delta)
        slope_delta_list.append(slope_delta)

        for i in range(self.number_of_layers - 2):

            layer_delta        = (layer_delta_list[-1].dot(self.weight_list[- 1 - i].T                                                         ) )    * self.activator_output_to_derivative(layer_list[- 1 - 1 - i] )  *   self.slope_list[-1 -1 -i]
            slope_delta        = (layer_delta_list[-1].dot(self.weight_list[- 1 - i].T                                                         ) )    * self.activator_output_to_derivative(layer_list[- 1 - 1 - i] )  * np.dot(layer_list[-1 -1 -i -1], self.weight_list[-1 -i -1])

            layer_delta_list.append(layer_delta)
            slope_delta_list.append(slope_delta)

        for i in range(self.number_of_layers - 1):

            self.weight_list[i]                             += np.array(layer_list[i]                                                             ).T.dot(np.array(layer_delta_list[- 1 - i])                                          ) * self.alpha + self.momentum_rate * self.weight_list_momentum[i]
            self.weight_list_momentum[i]                     = np.array(layer_list[i]                                                             ).T.dot(np.array(layer_delta_list[- 1 - i])                                          ) * self.alpha + self.momentum_rate * self.weight_list_momentum[i]
            self.slope_list[i]                              += np.array(slope_delta_list[- 1 - i]).sum(axis=0)                                                                                                                           * self.alpha + self.momentum_rate * self.slope_list_momentum[i]
            self.slope_list_momentum[i]                      = np.array(slope_delta_list[- 1 - i]).sum(axis=0)                                                                                                                           * self.alpha + self.momentum_rate * self.slope_list_momentum[i]


    def learn_batch(self, input, output):

        layer_list = self.generate_values_for_each_layer(input)

        self.train_for_weight(
                   layer_list,
                   output)

        return self