Java中的XOR神经网络

Question

我正在尝试用Java中的XOR函数实现和训练具有反向传播的五神经元神经网络.我的代码(请原谅它的丑陋):

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;
        }
    }
}
}

如果它有帮助,这是一个网络图.

所有权重和学习率的初始值都直接来自我的教科书中的一个例子.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(1,1,0)后所有权重的值,我测试了我的代码,其结果完全符合教科书的结果.但根据这本书,这应该只需要224个时代汇聚.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

尝试在激活阶段对 gamma3、gamma4、gamma5 进行舍入：

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

并增加一点学习变量（ alpha ）

alpha = 0.2;

学习在 466 个 epoch 后结束。

当然，如果你进行更大的舍入和更高的 alpha 设置，你可以获得比 224 更好的结果。