EM算法中三硬币模型M步的推导

最新推荐文章于 2024-04-15 16:42:53 发布

Jun~o~

最新推荐文章于 2024-04-15 16:42:53 发布

阅读量762

点赞数 2

分类专栏：机器学习

本文链接：https://blog.csdn.net/reboomjun/article/details/84310710

版权

机器学习专栏收录该内容

2 篇文章 0 订阅

订阅专栏

观测数据来自掷硬币B的概率为 $u_j^{(i + 1)}$ 来自掷硬币C的概率为 $1 - u_j^{(i + 1)}$ 则：

${\bf{Q}}({\theta ^{(i + 1)}},{\theta ^{(i)}}) = \sum\limits_{j = 1}^n {\sum\limits_z {\log } } P({y_j},z|{\theta ^{(i + 1)}})P(z|{y_j},{\theta ^{(i)}})\\ {\rm{ = }}\sum\limits_{j = 1}^n {u_j^{(i + 1)}\log \pi {p^{{y_j}}}{{(1 - p)}^{1 - {y_j}}} + (1 - u_j^{(i + 1)})\log (1 - \pi ){q^{{y_j}}}{{(1 - q)}^{1 - {y_j}}}} \\ = \sum\limits_{j = 1}^n {u_j^{(i + 1)}[\log \pi + {y_j}\log p + (1 - {y_j})\log (1 - p)] + (1 - u_j^{(i + 1)})[\log (1 - \pi ) + } {{\rm{y}}_j}\log q + (1 - {y_j})\log (1 - q)]{\rm{ }}\]$

分别对 $\pi ,p,q$ 求偏导：

$\frac{{\partial {\bf{Q}}}}{{\partial \pi }} = \frac{{\sum\limits_{j = 1}^n {{u_j}} }}{\pi } - \frac{{\sum\limits_{j = 1}^n {1 - {u_j}} }}{{(1 - \pi )}} = 0 \Rightarrow {\pi ^{(i + 1)}} = \frac{1}{n}\sum\limits_{j = 1}^n {u_j^{(i + 1)}}$

$\frac{{\partial {\bf{Q}}}}{{\partial q}} = \frac{{\sum\limits_{j = 1}^n {(1 - u_j^{(i + 1)})} {y_j}}}{q} - \frac{{\sum\limits_{j = 1}^n {(1 - u_j^{(i + 1)})(1 - {y_j})} }}{{(1 - q)}} = 0 \Rightarrow {q^{(i + 1)}} = \frac{{\sum\limits_{j = 1}^n {(1 - u_j^{(i + 1)}){y_j}} }}{{\sum\limits_{j = 1}^n {(1 - u_j^{(i + 1)})} }}$