1

我正在尝试建立一个神经网络来解决 XOR 问题。但我做不到。总是给出错误的结果。也许我在你的数学中犯了一个错误。网络不学习。结果总是相似的。

我没有使用偏见。

Note: execute function = (feed-forward + backpropagation)
ALPHA = 0.5

这是代码：

//main.cpp

#include <iostream>

#include "neural_network.h"

int main(int argc, char const *argv[])
{
 srand(time(NULL));

 double array[][3] = {{0.0, 0.0, 0.0},
                      {0.0, 1.0, 1.0},
                      {1.0, 0.0, 1.0},
                      {1.0, 1.0, 0.0}};


neural_network* nn = new neural_network(3, 2, 2, 1, 1.0);
nn->create_network();

for(int i = 0; i < 15000; i++)
{
    int index = rand() % 4;
#if DEBUG
    std::cout<<"Inputs :"<<array[index][0]<<" , "<<array[index][1]<<std::endl;
    std::cout<<"Outputs :"<<array[index][2]<<std::endl;
#endif    
    nn->execute(array[index], &array[index][2]);
}

nn->print_weight();

nn->execute(array[0], &array[0][2]);
nn->print_output();

nn->execute(array[1], &array[1][2]);
nn->print_output();

nn->execute(array[2], &array[2][2]);
nn->print_output();

nn->execute(array[3], &array[3][2]);
nn->print_output();

return 0;
}

//前馈函数

void neural_network::feed_forward(double* inputs)
{
int index = 0;

for(int i = 0; i < neural_network::input_layer_size; i++)
    neural_network::input_neuron[i] = inputs[i];

for(int i = 0; i < neural_network::hidden_layer_size; i++)
{
    for(int j = 0; j < neural_network::input_layer_size; j++)
    {
        neural_network::hidden_neuron[i] += neural_network::input_neuron[j] * weight_I_H[index++];
    }
    neural_network::hidden_neuron[i] = neural_network::activation_func(neural_network::hidden_neuron[i]);
}    

index = 0;

for(int i = 0; i < neural_network::output_layer_size; i++)
{
    for(int j = 0; j < neural_network::hidden_layer_size; j++)
    {
        neural_network::output_neuron[i] += neural_network::hidden_neuron[j] * weight_H_O[index++];
    }
    neural_network::output_neuron[i] = neural_network::activation_func(neural_network::output_neuron[i]);
}

}

//反向传播函数

void neural_network::back_propagation(double* outputs)
{
    int index;

for(int i = 0; i < neural_network::output_layer_size; i++)
    neural_network::err_output[i] = (outputs[i] - neural_network::output_neuron[i]);

for(int i = 0; i < neural_network::hidden_layer_size; i++)
{
    index = i;
    for(int j = 0; j < neural_network::output_layer_size; j++)
    {
        neural_network::err_hidden[i] += neural_network::weight_H_O[index] * neural_network::err_output[j] * neural_network::derivative_act_func(neural_network::output_neuron[j]);
        neural_network::weight_H_O[index] += ALPHA * neural_network::err_output[j] * neural_network::derivative_act_func(neural_network::output_neuron[j]) * neural_network::hidden_neuron[i];
        index += neural_network::hidden_layer_size;
    }
}

for(int i = 0; i < neural_network::input_layer_size; i++)
{
    index = i;
    for(int j = 0; j < neural_network::hidden_layer_size; j++)
    {
        neural_network::weight_I_H[index] += ALPHA * neural_network::err_hidden[j] * neural_network::derivative_act_func(neural_network::hidden_neuron[j]) * neural_network::input_neuron[i];
        index += neural_network::input_layer_size;
    }
}
}

//输出

Input To Hidden :
H-1 :
    Weight :-13.269
    Weight :-13.2705
H-2 :
    Weight :-12.5172
    Weight :-12.5195

Hidden To Output :
O-1 :
    Weight :-5.37707
    Weight :-2.93218


Outputs for (0,0):
    O-1 :0.0294265

Outputs for (0,1):
    O-1 :0.507348

Outputs for (1,0):
    O-1 :0.62418

Outputs for (1,1):
    O-1 :0.651169
4

1 回答 1

0

真的不可能没有 keras 没有我基于 Furye 变换（比 keras 更强大）开发的网络真正决定这个 XOR 任务。我测试了这两个 ANN 的非常准确。最大识别是 4 的 3 个示例（acc = 0.75- >75%）。没有人回答 1 xor 1=0。看来现在有人真的认真地测试了这个案例。（ANN 是多层的）

于 2020-07-30T18:28:25.753 回答