[概率图模型]最大后验概率(MAP)变量消除实例

以学生网络为例 ：

.

求 MAP

解

最大-积 变量消除算法的顺序为：

联合概率

联合概率为：$P(S,I,D,L,G)={\varphi }_s(I,S)\ast {\varphi }_I(I)\ast {\varphi }_D(D)\ast {\varphi }_G(G,D,I)\ast {\varphi }_L(L)$
每个$\varphi$对应于图中节点的条件概率表。

本文的两个约定

• a. 因子表示为矩阵形式，因子间相乘对应于矩阵相乘，这里的矩阵乘法，是指待消除变量取值相同的元素相乘，如
  ${\varphi }_I(I)\ast {\tau }_1(I)=[0.7,0.3]\ast [0.95,0.8]=[0.7\ast 0.95,0.3\ast 0.8]=[0.665,0.024]$
• b. $ma{x}_x$。对于其它变量的各种组合，尝试消除变量x的可能取值，得到的各种组合最大值。
  如对于条件概率表：
  
  要消除S，当$S=s^0$ 时 $i^0$ 取最大值 0.95，当$S=s^1$ 时 $i^1$ 取最大值 0.8
  表示为：
  $ma{x}_S(\left[\begin{array}{c}0.95,0.05\\ 0.2,0.8\end{array}\right])=[0.95,0.8]$

消除步骤

• 1. 消除 S
     ${\tau }_1(I)=ma{x}_S({\varphi }_S(I,S))=ma{x}_S(\left[\begin{array}{c}0.95,0.05\\ 0.2,0.8\end{array}\right])=[0.95,0.8]$
• 2. 消除 I
     ${\tau }_2(G,D)=ma{x}_I({\varphi }_I(I)\ast {\tau }_1(I)\ast {\varphi }_G(G,I,D))=[0.7,0.3]\ast [0.95,0.8]\ast ma{x}_{GD}(\left[\begin{array}{c}0.3,0.4,0.3\\ 0.05,0.25,0.7\\ 0.9,0.08,0.02\\ 0.5,0.3,0.2\end{array}\right])=[0.665,0.024]\ast ma{x}_{GD}(\left[\begin{array}{c}0.3,0.4,0.3\\ 0.05,0.25,0.7\\ 0.9,0.08,0.02\\ 0.5,0.3,0.2\end{array}\right])=ma{x}_{GD}(\left[\begin{array}{c}0.3\ast 0.665,0.4\ast 0.665,0.3\ast 0.665\\ 0.05\ast 0.665,0.25\ast 0.665,0.7\ast 0.665\\ 0.9\ast 0.024,0.08\ast 0.024,0.02\ast 0.024\\ 0.5\ast 0.024,0.3\ast 0.024,0.2\ast 0.024\end{array}\right])=\left[\begin{array}{c}0.3\ast 0.665,0.4\ast 0.665,0.3\ast 0.665\\ 0.05\ast 0.665,0.25\ast 0.665,0.7\ast 0.665\end{array}\right]=\left[\begin{array}{c}0.3,0.4,0.3\\ 0.05,0.25,0.7\end{array}\right]\ast 0.665$
     这里正好巧了，都是取$I=i^0$时 G和D的各种联合概率最大，所以又把 0.665提出来，减少运算，因为常数不影响比较。
• 3. 消除 D
     ${\tau }_3(G)=ma{x}_D({\varphi }_D(D)\ast {\tau }_2(G,D))=ma{x}_D([0.6,0.4]\ast \left[\begin{array}{c}0.3,0.4,0.3\\ 0.05,0.25,0.7\end{array}\right]\ast 0.665)=ma{x}_D(\left[\begin{array}{c}0.18,0.24,0.18\\ 0.02,0.1,0.28\end{array}\right]\ast 0.665)=[0.18,0.24,0.28]\ast 0.665$
• 4. 消除 L
     ${\tau }_4(G)=ma{x}_L({\varphi }_L(L,G))=ma{x}_L(\left[\begin{array}{c}0.1,0.9\\ 0.4,0.6\\ 0.99,0.01\end{array}\right])=[0.9,0.6,0.99]$
• 5. 消除 G
     ${\tau }_5()=ma{x}_G(\tau 3(G)\ast \tau 4(G))=ma{x}_G([0.18,0.24,0.28]\ast 0.665\ast [0.9,0.6,0.99])=ma{x}_G([0.162,0.144,0.277]\ast 0.665)=0.277\ast 0.665=0.184$

MAP 即 $ma{x}_{SIDLG}P(S,I,D,L,G)=0.184$