EM算法
简介
EM算法的核心分为两步
- E步(Expection-Step)
- M步(Maximization-Step)
因为在最大化过程中存在两个参量 r , θ r,\theta r,θ,其中若知道 r r r,则知道 θ \theta θ;若知道 θ \theta θ,则知道 r r r。且两个量未存在明显的关系,但又互相依存可以采用EM算法
其中主要思想为:
- 首先随机初始化参数 r r r
- 然后求的在参数 r r r下按照极大似然估计求得参数 θ \theta θ
- 然后根据参数 θ \theta θ按照极大似然估计求得参数 r r r
- 循环至收敛
算法示例
如下图所示存在A,B两种硬币,其中抛出正反面的概率未知,其中H
表示正面,F
表示反面
根据统计可得
可得
P ( H ∣ X = A ) = 24 24 + 6 = 0.8 P ( H ∣ X = B ) = 9 9 + 11 = 0.45 P(H|X=A)=\frac{24}{24+6}=0.8\\ P(H|X=B)=\frac{9}{9+11}=0.45 P(H∣X=A)=24+624=0.8P(H∣X=B)=9+119=0.45
若更改条件,不知道此时抛出是哪一枚硬币,只知道抛出的结果,即
首先初始化,设
P ( H ∣ X = A ) = 0.6 P ( H ∣ X = B ) = 0.5 P(H|X=A)=0.6\\ P(H|X=B)=0.5 P(H∣X=A)=0.6P(H∣X=B)=0.5
若当抛出的第一枚硬币为A时
此时的出现该情况的概率为 P 1 ( A ) = 0. 6 5 ∗ ( 1 − 0.6 ) 5 = 0.0007962624 P_1(A)=0.6^5*(1-0.6)^5=0.0007962624 P1(A)=0.65∗(1−0.6)5=0.0007962624
若当抛出的第一枚硬币为B时
此时的出现该情况的概率为 P 1 ( B ) = 0. 5 5 ∗ ( 1 − 0.5 ) 5 = 0.0009765625 P_1(B)=0.5^5*(1-0.5)^5=0.0009765625 P1(B)=0.55∗(1−0.5)5=0.0009765625
其中
P 1 ( A ) = P 1 ( A ) P 1 ( A ) + P 1 ( B ) ≈ 0.45 P 1 ( B ) = P 1 ( B ) P 1 ( A ) + P 1 ( B ) ≈ 0.55 P^1(A)=\frac{P_1(A)}{P_1(A)+P_1(B)}\approx0.45\\ P^1(B)=\frac{P_1(B)}{P_1(A)+P_1(B)}\approx0.55 P1(A)=P1(A)+P1(B)P1(A)≈0.45P1(B)=P1(A)+P1(B)P1(B)≈0.55
同理可得
P 2 ( A ) ≈ 0.80 , P 2 ( B ) ≈ 0.20 P 3 ( A ) ≈ 0.73 , P 3 ( B ) ≈ 0.27 P 4 ( A ) ≈ 0.35 , P 4 ( B ) ≈ 0.65 P 5 ( A ) ≈ 0.65 , P 5 ( B ) ≈ 0.35 P^2(A)\approx0.80,P^2(B)\approx0.20\\ P^3(A)\approx0.73,P^3(B)\approx0.27\\ P^4(A)\approx0.35,P^4(B)\approx0.65\\ P^5(A)\approx0.65,P^5(B)\approx0.35 P2(A)≈0.80,P2(B)≈0.20P3(A)≈0.73,P3(B)≈0.27P4(A)≈0.35,P4(B)≈0.65P5(A)≈0.65,P5(B)≈0.35
计算其数学期望
E 1 ( H ∣ X = A ) = 0.45 ∗ 5 = 2.2 E 1 ( T ∣ X = A ) = 0.45 ∗ 5 = 2.2 E 1 ( H ∣ X = B ) = 0.55 ∗ 5 = 2.75 E 1 ( T ∣ X = B ) = 0.55 ∗ 5 = 2.75 E 2 ( H ∣ X = A ) = 0.80 ∗ 9 = 7.2 E 2 ( T ∣ X = A ) = 0.45 ∗ 5 = 2.2 E 2 ( H ∣ X = B ) = 0.20 ∗ 9 = 1.8 E 2 ( T ∣ X = B ) = 0.20 ∗ 1 = 0.2 E 3 ( H ∣ X = A ) = 0.73 ∗ 8 = 5.8 E 3 ( T ∣ X = A ) = 0.73 ∗ 2 = 1.46 E 3 ( H ∣ X = B ) = 0.27 ∗ 8 = 2.16 E 3 ( T ∣ X = B ) = 0.27 ∗ 2 = 0.54 E 4 ( H ∣ X = A ) = 0.35 ∗ 4 = 1.4 E 4 ( T ∣ X = A ) = 0.35 ∗ 6 = 2.1 E 4 ( H ∣ X = B ) = 0.65 ∗ 4 = 2.6 E 4 ( T ∣ X = B ) = 0.65 ∗ 6 = 3.9 E 5 ( H ∣ X = A ) = 0.65 ∗ 7 = 4.55 E 5 ( T ∣ X = A ) = 0.65 ∗ 3 = 1.95 E 5 ( H ∣ X = B ) = 0.35 ∗ 7 = 2.45 E 5 ( T ∣ X = B ) = 0.35 ∗ 3 = 1.05 \begin{aligned} &E_1(H|X=A)=0.45*5=2.2\\ &E_1(T|X=A)=0.45*5=2.2\\ &E_1(H|X=B)=0.55*5=2.75\\ &E_1(T|X=B)=0.55*5=2.75\\ &E_2(H|X=A)=0.80*9=7.2\\ &E_2(T|X=A)=0.45*5=2.2\\ &E_2(H|X=B)=0.20*9=1.8\\ &E_2(T|X=B)=0.20*1=0.2\\ &E_3(H|X=A)=0.73*8=5.8\\ &E_3(T|X=A)=0.73*2=1.46\\ &E_3(H|X=B)=0.27*8=2.16\\ &E_3(T|X=B)=0.27*2=0.54\\ &E_4(H|X=A)=0.35*4=1.4\\ &E_4(T|X=A)=0.35*6=2.1\\ &E_4(H|X=B)=0.65*4=2.6\\ &E_4(T|X=B)=0.65*6=3.9\\ &E_5(H|X=A)=0.65*7=4.55\\ &E_5(T|X=A)=0.65*3=1.95\\ &E_5(H|X=B)=0.35*7=2.45\\ &E_5(T|X=B)=0.35*3=1.05\\ \end{aligned} E1(H∣X=A)=0.45∗5=2.2E1(T∣X=A)=0.45∗5=2.2E1(H∣X=B)=0.55∗5=2.75E1(T∣X=B)=0.55∗5=2.75E2(H∣X=A)=0.80∗9=7.2E2(T∣X=A)=0.45∗5=2.2E2(H∣X=B)=0.20∗9=1.8E2(T∣X=B)=0.20∗1=0.2E3(H∣X=A)=0.73∗8=5.8E3(T∣X=A)=0.73∗2=1.46E3(H∣X=B)=0.27∗8=2.16E3(T∣X=B)=0.27∗2=0.54E4(H∣X=A)=0.35∗4=1.4E4(T∣X=A)=0.35∗6=2.1E4(H∣X=B)=0.65∗4=2.6E4(T∣X=B)=0.65∗6=3.9E5(H∣X=A)=0.65∗7=4.55E5(T∣X=A)=0.65∗3=1.95E5(H∣X=B)=0.35∗7=2.45E5(T∣X=B)=0.35∗3=1.05
并计算其总共的期望
E ( H ∣ X = A ) = ∑ i = 1 5 E i ( H ∣ X = A ) ≈ 21.3 E ( H ∣ X = A ) = ∑ i = 1 5 E i ( T ∣ X = A ) ≈ 8.6 E ( H ∣ X = B ) = ∑ i = 1 5 E i ( H ∣ X = B ) ≈ 11.7 E ( H ∣ X = B ) = ∑ i = 1 5 E i ( T ∣ X = B ) ≈ 8.4 E(H|X=A)=\sum_{i=1}^5E_i(H|X=A)\approx21.3\\ E(H|X=A)=\sum_{i=1}^5E_i(T|X=A)\approx8.6\\ E(H|X=B)=\sum_{i=1}^5E_i(H|X=B)\approx11.7\\ E(H|X=B)=\sum_{i=1}^5E_i(T|X=B)\approx8.4\\ E(H∣X=A)=i=1∑5Ei(H∣X=A)≈21.3E(H∣X=A)=i=1∑5Ei(T∣X=A)≈8.6E(H∣X=B)=i=1∑5Ei(H∣X=B)≈11.7E(H∣X=B)=i=1∑5Ei(T∣X=B)≈8.4
可得
P ( H ∣ X = A ) = 21.3 21.3 + 8.6 = 0.71 P ( H ∣ X = B ) = 11.7 11.7 + 8.4 = 0.58 P(H|X=A)=\frac{21.3}{21.3+8.6}=0.71\\ P(H|X=B)=\frac{11.7}{11.7+8.4}=0.58 P(H∣X=A)=21.3+8.621.3=0.71P(H∣X=B)=11.7+8.411.7=0.58
由此循环直至收敛
得到最终
P ( H ∣ X = A ) = 0.80 P ( H ∣ X = B ) = 0.58 P(H|X=A)=0.80\\ P(H|X=B)=0.58 P(H∣X=A)=0.80P(H∣X=B)=0.58