局部加权学习算法(LWR)
局部加权回归(LWR)是非参数学习方法。 首先参数学习方法是这样一种方法:在训练完成所有数据后得到一系列训练参数,然后根据训练参数来预测新样本的值,这时不再依赖之前的训练数据了,参数值是确定的。而非参数学习方法是这样一种算法:在预测新样本值时候每次都会重新训练数据得到新的参数值,也就是说每次预测新样本都会依赖训练数据集合,所以每次得到的参数值是不确定的。 局部加权回归(LWR)是我们遇到的第一个non-parametric(非参数)学习算法,而线性回归则是我们遇到的以一个parametric(参数)学习算法。因为参数学习算法它有固定的明确的参数,所以参数一旦确定,就不会改变了,我们不需要在保留训练集中的训练样本。而非参数学习算法,每进行一次预测,就需要重新学习一组,是变化的,所以需要一直保留训练样本。因而,当训练集的容量较大时,非参数学习算法需要占用更多的存储空间,计算速度也较慢。所以有得必有失,效果好了,计算速度却降下来了。
应用案例
# coding: utf-8
# linear_regression/regression.pydef JLwr(theta, X, y, x, c):"""局部加权线性回归的代价函数计算式Args:theta: 相关系数矩阵X: 样本集矩阵y: 标签集矩阵x: 待预测输入c: tauReturns:预测代价"""m,n = X.shapesummerize = 0for i in range(m):diff = (X[i]-x)*(X[i]-x).Tw = np.exp(-diff/(2*c*c))predictDiff = np.power(y[i] - X[i]*theta,2)summerize = summerize + w*predictDiffreturn summerize@exeTime
def lwr(rate, maxLoop, epsilon, X, y, x, c=1):"""局部加权线性回归Args:rate: 学习率maxLoop: 最大迭代次数epsilon: 预测精度X: 输入样本y: 标签向量x: 待预测向量c: tau"""m,n = X.shape# 初始化thetatheta = np.zeros((n,1))count = 0converged = Falseerror = float('inf')errors = []thetas = {}for j in range(n):thetas[j] = [theta[j,0]]# 执行批量梯度下降while count<=maxLoop:if(converged):breakcount = count + 1for j in range(n):deriv = (y-X*theta).T*X[:, j]/mtheta[j,0] = theta[j,0]+rate*derivthetas[j].append(theta[j,0])error = JLwr(theta, X, y, x, c)errors.append(error[0,0])# 如果已经收敛if(error < epsilon):converged = Truereturn theta,errors,thetas
结果
# coding: utf-8
# linear_regression/test_lwr.py
import regression
import matplotlib.pyplot as plt
import matplotlib.ticker as mtick
import numpy as npif __name__ == "__main__":srcX, y = regression.loadDataSet('data/lwr.txt');m,n = srcX.shapesrcX = np.concatenate((srcX[:, 0], np.power(srcX[:, 0],2)), axis=1)# 特征缩放X = regression.standardize(srcX.copy())X = np.concatenate((np.ones((m,1)), X), axis=1)rate = 0.1maxLoop = 1000epsilon = 0.01predicateX = regression.standardize(np.matrix([[8, 64]]))predicateX = np.concatenate((np.ones((1,1)), predicateX), axis=1)result, t = regression.lwr(rate, maxLoop, epsilon, X, y, predicateX, 1)theta, errors, thetas = resultresult2, t = regression.lwr(rate, maxLoop, epsilon, X, y, predicateX, 0.1)theta2, errors2, thetas2 = result2# 打印特征点fittingFig = plt.figure()title = 'polynomial with bgd: rate=%.2f, maxLoop=%d, epsilon=%.3f'%(rate,maxLoop,epsilon)ax = fittingFig.add_subplot(111, title=title)trainingSet = ax.scatter(srcX[:, 0].flatten().A[0], y[:,0].flatten().A[0])print thetaprint theta2# 打印拟合曲线xx = np.linspace(1, 7, 50)xx2 = np.power(xx,2)yHat1 = []yHat2 = []for i in range(50):normalizedSize = (xx[i]-xx.mean())/xx.std(0)normalizedSize2 = (xx2[i]-xx2.mean())/xx2.std(0)x = np.matrix([[1,normalizedSize, normalizedSize2]])yHat1.append(regression.h(theta, x.T))yHat2.append(regression.h(theta2, x.T))fittingLine1, = ax.plot(xx, yHat1, color='g')fittingLine2, = ax.plot(xx, yHat2, color='r')ax.set_xlabel('temperature')ax.set_ylabel('yield')plt.legend([trainingSet, fittingLine1, fittingLine2], ['Training Set', r'LWR with $\tau$=1', r'LWR with $\tau$=0.1'])plt.show()# 打印误差曲线errorsFig = plt.figure()ax = errorsFig.add_subplot(111)ax.yaxis.set_major_formatter(mtick.FormatStrFormatter('%.2e'))ax.plot(range(len(errors)), errors)ax.set_xlabel('Number of iterations')ax.set_ylabel('Cost J')plt.show()
最后,我们分别对ττ取值 0.10.1 和 11 ,得到了不同的拟合曲线: