This is the first time to write a technical blog post. If there are mistakes, please give pointers.
This blog post summarizes the learning of the LM algorithm through an example, and the programming language is python.
I won't talk about the theory, there are a lot of them on the Internet.
Fits the parameters a, b, c in the function y(x) = exp(a*x^2 + b * x + c) . Without further ado, let’s go directly to the code:
# -*- coding:utf-8 -*- # autor: HuangYuliang import numpy as np from numpy import matrix as mat from matplotlib import pyplot as plt import random n = 100 a1,b1,c1 = 1,3,2 # This is the real parameter of the function y(x) that needs to be fitted h = np.linspace(0,1,n) # generate data with noise y = [np.exp(a1*i**2+b1*i+c1)+random.gauss(0,4) for i in h] y = mat(y) # convert to matrix form def Func(abc,iput): # The function to be fitted, abc is a matrix containing three parameters [[a],[b],[c]] a = abc[0,0] b = abc[1,0] c = abc[2,0] return np.exp(a*iput**2+b*iput+c) def Deriv(abc,iput,n): # Find partial derivatives of functions x1 = abc.copy() x2 = abc.copy() x1[n,0] -= 0.000001 x2[n,0] += 0.000001 p1 = Func (x1, iput) p2 = Func (x2, iput) d = (p2-p1)*1.0/(0.000002) return d J = mat(np.zeros((n,3))) #Jacobian matrix fx = mat(np.zeros((n,1))) # f(x) 100*1 error fx_tmp = mat(np.zeros((n,1))) xk = mat([[0.8],[2.7],[1.5]]) # parameter initialization lase_mse = 0 step = 0 u,v= 1.2 conve = 100 while (conve): mse,mse_tmp = 0,0 step += 1 for i in range(n): fx[i] = Func(xk,h[i]) - y[0,i] # Note that it cannot be written as y - Func , otherwise it will diverge mse += fx[i,0]**2 for j in range(3): J[i,j] = Deriv(xk,h[i],j) # Numerical derivation mse /= n # range constraints H = J.T*J + u*np.eye(3) # 3*3 dx = -HI * JT*fx # Note that there is a negative sign here, which corresponds to the sign of fx = Func - y xk_tmp = xk.copy() xk_tmp += dx for j in range(n): fx_tmp[i] = Func(xk_tmp,h[i]) - y[0,i] mse_tmp += fx_tmp[i,0]**2 mse_tmp /= n q = (mse - mse_tmp)/((0.5*dx.T*(u*dx - J.T*fx))[0,0]) if q > 0: s = 1.0/3.0 v = 2 mse = mse_tmp xk = xk_tmp temp = 1 - pow(2*q-1,3) if s > temp: u = u*s else: u = u*temp else: u = u*v v = 2*v xk = xk_tmp print "step = %d,abs(mse-lase_mse) = %.8f" %(step,abs(mse-lase_mse)) if abs(mse-lase_mse)<0.000001: break last_mse = mse # record the position of the last mse conve -= 1 print xk z = [Func(xk,i) for i in h] #Draw a graph with the fitted parameters plt.figure(0) plt.scatter(h,y,s = 4) plt.plot(h,z,'r') plt.show()
The fitting effect is shown in the following figure:
The fitting results are as follows:
step = 1,abs(mse-lase_mse) = 6427.823881782
step = 2,abs(mse-lase_mse) = 1348.739146779
step = 3,abs(mse-lase_mse) = 7725.330011166
step = 4,abs(mse-lase_mse) = 51.226834643
step = 5,abs(mse-lase_mse) = 0.013664543
step = 6,abs(mse-lase_mse) = 0.000026830
step = 7,abs(mse-lase_mse) = 0.000000194
参数:[[ 0.97256842]
[ 3.03242402]
[ 1.99001528]]
It can be seen that the algorithm converges after seven iterations. And it is very close to the true value [1, 3, 2].
This blog post references the next blog post:
http://blog.csdn.net/jinshengtao/article/details/53310804