我的感知器可以对OR / AND进行分类，但不能对NOR / NAND进行分类

Question

这是我在python中的代码

def error(y_desired, y):
    return y_desired != y


def step_func(weighted_sum, theta):
    return 1 if ((weighted_sum - theta) >= 0) else 0


def weight_adjustment(error, alpha, x_element):
    return error*alpha*x_element


def weighted_sum(w, x_epoch):
    weighted_sum = 0
    for w, x_epoch in zip(w, x_epoch):
        weighted_sum += w*(x_epoch)
    return weighted_sum


def perceptron(x, y_desired, w, theta, alpha):
    cond = True
    epochs = 0
    while(cond == True):
        count = 0
        epochs += 1
        print(f'Epoch number - {epochs}')
        for epoch in range(len(x)):
            weighted_sums = round(weighted_sum(w, x[epoch]), 10)
            y = step_func(weighted_sums, theta)
            if error(y_desired[epoch], y):
                count += 1
                for weights in range(len(w)):
                    w[weights] = round(w[weights] + weight_adjustment(error = (y_desired[epoch] - y),
                                                   alpha = alpha,
                                                   x_element = x[epoch][weights]), 10)
            print(w)
        print('\n')
        if count == 0:
            cond = False
    print('Final Weights -')
    return w


x = [[0, 0], [0, 1], [1, 0], [1, 1]]
y_desired = [0, 0, 0, 1]
w = [0.3, -0.1]
perceptron(x, y_desired, w, theta = 0.2, alpha = 0.1)

OR / AND工作正常但是当我追求[1，0，0，0]它进入无限循环。 AND / NAND / OR / NOR的分类没有区别。全部都可以线性分类。

我是从根本上还是在减肥训练中错过了某些东西？我在哪里弄错了？如果可能的话，也分享学习材料。

Answer 1

首先，您要四舍五入，这是养成的坏习惯。其次，您要求错误为0，而分类通常基于概率。

我推荐https://stackabuse.com/creating-a-neural-network-from-scratch-in-python/由3篇文章组成的系列文章，在解释基础知识方面做得不错。