SVM 优化出来支持向量点的不等式约束不等于 1 是为什么?

    import numpy as np
    import random
    import matplotlib.pyplot as plt
    C=1
    def selectj(i,n):
        j = i
        while(j==i):
            j = random.randint(0,5)
        return j
	xt1 = np.array([[1,3],[3,5],[3,6]])
    xt2 = np.array([[0,2],[-3,0],[-4,-3]])
    plt.scatter(xt1[:,0], xt1[:,1], c='red', marker='o', s=30)
    plt.scatter(xt2[:,0], xt2[:,1], c='blue', marker='*', s=30)
    x = np.array([[1,3],[3,5],[3,6],[0,2],[-3,0],[-4,-3]])
    y = np.array([[1],[1],[1],[-1],[-1],[-1]])
    alpha = np.zeros((6, 1))
    w = np.dot((alpha*y).T, x)
    itera = 0
    b=0
    while (itera  < 10):
        itera = itera +1
        for i in range(6):
            w = np.dot((alpha * y).T, x)
            print(w)
            ev1 = np.dot(w, x[i]) + b - y[i]
            disold = 0
            if abs(ev1) > 0.001:
                # for j in range(6):
                #     ev2 = np.dot(w, x[j]) + b - y[j]
                #     dis = abs(float(ev1) - float(ev2))
                #     if dis > disold:
                #         disold = dis
                #         p = j
                p = selectj(i, 6)
                ev2 = float(np.dot(w, x[p]) + b - y[p])
                eta = -2 * np.dot(x[i, :], x[p, :]) + np.dot(x[i, :], x[i, :]) + np.dot(x[p, :], x[p, :])
    
                if y[i] != y[p]:
                    L= max(0, alpha[p] - alpha[i])
                    H= min(C, C + alpha[p] - alpha[i])
                else:
                    L=max(0, alpha[p] + alpha[i] - C)
                    H=min(C, alpha[p] + alpha[i])
                alphapold = alpha[p].copy()
                alphaiold = alpha[i].copy()
                alpha[p] = alpha[p] + y[p] * (float(ev1) - float(ev2)) / eta
                if alpha[p] < L:
                    alpha[p] = L
                elif alpha[p] > C:
                    alpha[p] = C
                alpha[i] += y[i] * y[p] * (alphapold-alpha[p])
    #            b1 = y[i] - np.dot(np.dot((alpha * y).T, x), x[i])
    #            b2 = y[p] - np.dot(np.dot((alpha * y).T, x), x[p])
                b1 = y[i] - np.dot((alpha * y).T, np.dot(x, x[i]))
                b2 = y[p] - np.dot((alpha * y).T, np.dot(x, x[p]))
                if 0 < alpha[i] < C:
                    b = float(b1)
                elif 0<alpha[p]<C:
                    b = float(b2)
                else:
                    b = (float(b1)+float(b2))/2
    
    for i in range(6):
        if alpha[i] < 0.01:
            alpha[i] = 0
    w = np.dot((alpha * y).T, x)
    b = -(y[0] * np.dot(w,x[0]) - y[3] * np.dot(w,x[3]))/2
    x231 = np.linspace(-3,3,200)
    y231 = -(w[0,0] * x231 + b)/w[0,1]
    plt.plot(x231,y231)
    print(y[0] * np.dot(w,x[0]))
    plt.show()

大家帮忙看看为什么优化出来,wx+b 不等于 1,其中 x 是支持向量


更多高质资源 尽在AIQ 机器学习大数据 知乎专栏 点击关注

转载请注明 AIQ - 最专业的机器学习大数据社区  http://www.6aiq.com