EM学习-思想和代码

EM算法的简明实现

当然是教学用的简明实现了,这份实现是针对双硬币模型的。

双硬币模型

假设有两枚硬币A、B,以相同的概率随机选择一个硬币,进行如下的抛硬币实验:共做5次实验,每次实验独立的抛十次,结果如图中a所示,例如某次实验产生了H、T、T、T、H、H、T、H、T、H,H代表正面朝上。

假设试验数据记录员可能是实习生,业务不一定熟悉,造成a和b两种情况

a表示实习生记录了详细的试验数据,我们可以观测到试验数据中每次选择的是A还是B

b表示实习生忘了记录每次试验选择的是A还是B,我们无法观测实验数据中选择的硬币是哪个

问在两种情况下分别如何估计两个硬币正面出现的概率?

EM算法.png

a情况相信大家都很熟悉,既然能观测到试验数据是哪枚硬币产生的,就可以统计正反面的出现次数,直接利用最大似然估计即可。

b情况就无法直接进行最大似然估计了,只能用EM算法,接下来引用nipunbatra博主的简明EM算法Python实现。

 

 1 # -*- coding: utf-8 -*-
 2 """
 3 Created on Tue Jul  4 18:23:28 2017
 4 
 5 @author: Administrator
 6 """
 7 
 8 import numpy as np
 9 from scipy import stats
10 
11 priors = [0.6, 0.5]
12 observations = np.array([[1,0,0,0,1,1,0,1,0,1],
13                          [1,1,1,1,0,1,1,1,1,1],
14                          [1,0,1,1,1,1,1,0,1,1],
15                          [1,0,1,0,0,0,1,1,0,0],
16                          [0,1,1,1,0,1,1,1,0,1]])
17 
18 def em_single(priors, observations):
19     """
20     input:
21         priors:[theta_A, theta_B]
22         obvervations:m*n matrix
23         
24     output:
25         
26     """
27     theta_A = priors[0]
28     theta_B = priors[1]
29     counts = {'A':{'H':0,'T':0}, 'B':{'H':0,'T':0}}
30     
31     # e-step
32     for observation in observations:
33         len_observation = len(observation)
34         num_heads = observation.sum()   # 正面个数
35         num_tails = len_observation - num_heads     # 反面个数
36         # 两个二项分布
37         contribution_A = stats.binom.pmf(num_heads, len_observation, theta_A)
38         contribution_B = stats.binom.pmf(num_heads, len_observation, theta_B)
39         # 采用各自硬币的权重
40         weight_A = contribution_A/(contribution_A+contribution_B)
41         weight_B = contribution_B/(contribution_A+contribution_B)
42         
43         # 更新在当前参数下,硬币A和B产生正反面的次数
44         counts['A']['H'] += weight_A * num_heads
45         counts['A']['T'] += weight_A * num_tails
46         counts['B']['H'] += weight_B * num_heads
47         counts['B']['T'] += weight_B * num_tails
48         
49     # M-step
50     new_theta_A = counts['A']['H']/(counts['A']['H'] + counts['A']['T'])
51     new_theta_B = counts['B']['H']/(counts['B']['H'] + counts['B']['T'])
52     
53     return [new_theta_A, new_theta_B]
54 
55     
56 
57 def em(observations, prior, tol=1e-6, iterations=10000):
58     """
59     EM算法
60     param observations: 观察数据
61     param prior: 模型初值
62     param tol: 迭代结束阈值
63     param iteration: 最大迭代数
64     return: 局部最优的模型参数
65     """
66     import math
67     iter = 0
68     while iter < iterations:
69         new_prior = em_single(prior, observations)
70         delta_change = np.abs(new_prior[0]-prior[0])
71         if delta_change < tol:
72             break
73         else:
74             prior = new_prior
75             iter += 1
76         print (iter)
77             
78     return [new_prior, iter]
79 
80 y = em(observations, priors)

 

 

参考自:http://www.hankcs.com/ml/em-algorithm-and-its-generalization.html

posted @ 2017-07-04 19:21  今夜无风  阅读(754)  评论(0编辑  收藏  举报