EM算法笔记

1、目标函数

      L(\theta ) =\prod_{i=1}^{n} p(x_{i}|\theta ) = \prod_{i=1}^{n} \sum_{j=1}^{m} p(x_{i};z_{j}|\theta)

      l(\theta ) = log L(\theta)

       l(\theta ) = \sum_{i=1}^{n} log\sum_{j=1}^{m} p(x_{i};z_{j}|\theta) \\= \sum_{i=1}^{n} log\sum_{j=1}^{m}Q(z_j) \frac{p(x_{i};z_{j}|\theta)}{Q(z_j)} \\\geq \sum_{i=1}^{n} \sum_{j=1}^{m}Q(z_j)log \frac{p(x_{i};z_{j}|\theta)}{Q(z_j)}

       \theta = arg(max(l(\theta ) ))

2、取等号

        \frac{p(x_{i};z_{j}|\theta)}{Q(z_j)} = C ; \sum_{j=1}^{m}Q(z_j)=1 \\ \Rightarrow p(x_{i};z_{j}|\theta) = C*Q(z_j) \\ \Rightarrow \sum_{j=1}^{m}p(x_{i};z_{j}|\theta) = \sum_{j=1}^{m}(C*Q(z_j))=C

       Q(z_j) = \frac{p(x_{i};z_{j}|\theta)}{\sum_{j=1}^{m}p(x_{i};z_{j}|\theta)} = \frac{p(x_{i};z_{j}|\theta)}{p(x_{i}|\theta)} =p(z_{j}|x_{i},\theta)           

3、迭代

       E-step:Q(z_j) =p(z_{j}|x_{i},\theta)      

       M-step:\theta = arg(max(l(\theta ) ))

4、用途

       模型参数估计

猜你喜欢

转载自blog.csdn.net/ziyue246/article/details/81711966