Skip to content

Commit

Permalink
resolve szcf-weiya#125
Browse files Browse the repository at this point in the history
  • Loading branch information
szcf-weiya committed Apr 29, 2018
1 parent c06f511 commit 3c42f01
Show file tree
Hide file tree
Showing 2 changed files with 8 additions and 2 deletions.
1 change: 1 addition & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,7 @@ The Elements of Statistical Learning (ESL) 的中文翻译、代码实现及其
- [![](https://img.shields.io/badge/solution-chapter%205-blue.svg)](https://github.com/szcf-weiya/ESL-CN/milestone/6)
- [![](https://img.shields.io/badge/solution-chapter%206-blue.svg)](https://github.com/szcf-weiya/ESL-CN/milestone/14)
- [![](https://img.shields.io/badge/solution-chapter%207-blue.svg)](https://github.com/szcf-weiya/ESL-CN/milestone/4)
- [![](https://img.shields.io/badge/solution-chapter%208-blue.svg)](https://github.com/szcf-weiya/ESL-CN/milestone/17)
- [![](https://img.shields.io/badge/solution-chapter%209-blue.svg)](https://github.com/szcf-weiya/ESL-CN/milestone/15)
- [![](https://img.shields.io/badge/solution-chapter%2010-blue.svg)](https://github.com/szcf-weiya/ESL-CN/milestone/1)
- [![](https://img.shields.io/badge/solution-chapter%2012-blue.svg)](https://github.com/szcf-weiya/ESL-CN/milestone/3)
Expand Down
9 changes: 7 additions & 2 deletions docs/08-Model-Inference-and-Averaging/8.5-The-EM-Algorithm.md
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,8 @@
| ---- | ---------------------------------------- |
| 翻译 | szcf-weiya |
| 时间 | 2016-12-20 & 2017-02-01:2017-02-03 |
|更新|2018-04-29|
|状态|Done|

EM算法是简化复杂极大似然问题的一种很受欢迎的工具。我们首先在一个简单的混合模型中讨论它。

Expand Down Expand Up @@ -132,10 +134,10 @@ $$
&\equiv Q(\theta',\theta)-R(\theta',\theta)\qquad\qquad (8.46)
\end{align}
$$
在最大化那一步,EM算法最大化关于$\theta'$的$Q(\theta',\theta)$,而不是实际的目标函数$\ell(\theta';\mathbf Z)$。为什么这样能成功地最大化$\ell(\theta';\mathbf Z)$?注意到$R(\theta^\*,\theta)$是由$\theta^\*$指示的密度的期望,这个密度同时由$\theta$指示,因此(由琴生不等式)当$\theta^\*=\theta$时(见练习8.1)最大化关于$\theta^\*$的函数。
在最大化那一步,EM算法最大化关于$\theta'$的$Q(\theta',\theta)$,而不是实际的目标函数$\ell(\theta';\mathbf Z)$。为什么这样能成功地最大化$\ell(\theta';\mathbf Z)$?注意到$R(\theta^\*,\theta)$是由$\theta^\*$指示的密度的期望,这个密度同时由$\theta$指示,因此(由琴生不等式)当$\theta^\*=\theta$时([练习 8.1](https://github.com/szcf-weiya/ESL-CN/issues/125))最大化关于$\theta^\*$的函数。

!!! note "weiya注"
练习8.1如下
练习 8.1 如下
![](../img/08/ex8.1.png)
证明想法如下:
在给定$Y=y$的情况下,记$x=r(y), c=q(y)$
Expand All @@ -154,6 +156,9 @@ $$
$$
注意到$\ell_1$其实是对数似然,分母为是与$\theta$有关的常数,所以满足题目中结论的形式,于是$R(\theta,\theta)\ge R(\theta',\theta)$

!!! info "Ex. 8.1"
详细解答过程也可以参见 [Issue 125: Ex. 8.1](https://github.com/szcf-weiya/ESL-CN/issues/125)


所以如果$\theta'$最大化$Q(\theta',\theta)$,我们可以看到
$$
Expand Down

0 comments on commit 3c42f01

Please sign in to comment.