Java中的异或神经网络

Question

我正在尝试实现并训练一个五神经元神经网络，它具有用于Java中异或函数的反向传播。我的代码（请原谅它的丑）：

public class XORBackProp { 

private static final int MAX_EPOCHS = 500; 

//weights 
private static double w13, w23, w14, w24, w35, w45; 
private static double theta3, theta4, theta5; 
//neuron outputs 
private static double gamma3, gamma4, gamma5; 
//neuron error gradients 
private static double delta3, delta4, delta5; 
//weight corrections 
private static double dw13, dw14, dw23, dw24, dw35, dw45, dt3, dt4, dt5; 
//learning rate 
private static double alpha = 0.1; 
private static double error; 
private static double sumSqrError; 
private static int epochs = 0; 
private static boolean loop = true; 

private static double sigmoid(double exponent) 
{ 
    return (1.0/(1 + Math.pow(Math.E, (-1) * exponent))); 
} 

private static void activateNeuron(int x1, int x2, int gd5) 
{ 
    gamma3 = sigmoid(x1*w13 + x2*w23 - theta3); 
    gamma4 = sigmoid(x1*w14 + x2*w24 - theta4); 
    gamma5 = sigmoid(gamma3*w35 + gamma4*w45 - theta5); 

    error = gd5 - gamma5; 

    weightTraining(x1, x2); 
} 

private static void weightTraining(int x1, int x2) 
{ 
    delta5 = gamma5 * (1 - gamma5) * error; 
    dw35 = alpha * gamma3 * delta5; 
    dw45 = alpha * gamma4 * delta5; 
    dt5 = alpha * (-1) * delta5; 

    delta3 = gamma3 * (1 - gamma3) * delta5 * w35; 
    delta4 = gamma4 * (1 - gamma4) * delta5 * w45; 

    dw13 = alpha * x1 * delta3; 
    dw23 = alpha * x2 * delta3; 
    dt3 = alpha * (-1) * delta3; 
    dw14 = alpha * x1 * delta4; 
    dw24 = alpha * x2 * delta4; 
    dt4 = alpha * (-1) * delta4; 

    w13 = w13 + dw13; 
    w14 = w14 + dw14; 
    w23 = w23 + dw23; 
    w24 = w24 + dw24; 
    w35 = w35 + dw35; 
    w45 = w45 + dw45; 
    theta3 = theta3 + dt3; 
    theta4 = theta4 + dt4; 
    theta5 = theta5 + dt5; 
} 

public static void main(String[] args) 
{ 

    w13 = 0.5; 
    w14 = 0.9; 
    w23 = 0.4; 
    w24 = 1.0; 
    w35 = -1.2; 
    w45 = 1.1; 
    theta3 = 0.8; 
    theta4 = -0.1; 
    theta5 = 0.3; 

    System.out.println("XOR Neural Network"); 

    while(loop) 
    { 
     activateNeuron(1,1,0); 
     sumSqrError = error * error; 
     activateNeuron(0,1,1); 
     sumSqrError += error * error; 
     activateNeuron(1,0,1); 
     sumSqrError += error * error; 
     activateNeuron(0,0,0); 
     sumSqrError += error * error; 

     epochs++; 

     if(epochs >= MAX_EPOCHS) 
     { 
      System.out.println("Learning will take more than " + MAX_EPOCHS + " epochs, so program has terminated."); 
      System.exit(0); 
     } 

     System.out.println(epochs + " " + sumSqrError); 

     if (sumSqrError < 0.001) 
     { 
      loop = false; 
     } 
    } 
} 
} 

如果有帮助的任何，这里有一个diagram of the network。

所有权重和学习率的初始值都是从我的教科书中的示例中直接获得的。目标是训练网络，直到平方误差的总和小于.001。教科书还给出了第一次迭代之后所有权重的值（1,1,0），并且我测试了我的代码，其结果与教科书的结果完美匹配。但根据这本书，这应该只需要224个时代的融合。但是当我运行它时，它总是达到MAX_EPOCHS，除非它被设置为几千。我究竟做错了什么？

Answer 1

尝试在激活阶段使GAMMA3，gamma4，gamma5的舍入而instace：

if (gamma3 > 0.7) gamma3 = 1; 
if (gamma3 < 0.3) gamma3 = 0; 

和上升点点主义学习变量（阿尔法）在466个历元

alpha = 0.2; 

学习结束。

当然，如果你让大舍入和高级α-u盘u能达到更好的效果比224

Answer 2

//Add this in the constants declaration section. 
    private static double alpha = 3.8, g34 = 0.13, g5 = 0.21; 

    // Add this in activate neuron 
    gamma3 = sigmoid(x1 * w13 + x2 * w23 - theta3); 
    gamma4 = sigmoid(x1 * w14 + x2 * w24 - theta4);   
    if (gamma3 > 1 - g34) {gamma3 = 1;} 
    if (gamma3 < g34) {gamma3 = 0;} 
    if (gamma4 > 1- g34) {gamma4 = 1;} 
    if (gamma4 < g34) {gamma4 = 0;} 
    gamma5 = sigmoid(gamma3 * w35 + gamma4 * w45 - theta5); 
    if (gamma5 > 1 - g5) {gamma5 = 1;} 
    if (gamma5 < g5) {gamma5 = 0;} 

ANN应该学会在66次迭代，但是是发散的边缘。

Answer 3

这个网络的整个目的是展示如何处理组合时不是基于“top = yes，bottom = no”，而是有一个中心线（通过点（0,1 ）和（1,0）），如果值接近该行，则答案为“是”，而如果它很远，则答案为“否”。你不能将这样的系统只聚集一层。但是，两层就足够了。