GPS方法中监督相推导

GPS方法中监督相推导

GPS方法中监督相优化问题
$${\pi }_{\theta }\leftarrow \arg \underset{\theta }{\min }\sum _{t,i,j}{D}_{\mathrm{K}\mathrm{L}}\left({\pi }_{\theta }\left({\mathbf{u}}_t\mid {\mathbf{x}}_{t,i,j}\right)\Vert p_i\left({\mathbf{u}}_t\mid {\mathbf{x}}_{t,i,j}\right)\right)$$
其中${\pi }_{\theta }\left({\mathbf{u}}_t\mid {\mathbf{x}}_t\right)=\mathcal{N}\left({\mu }^{\pi }\left({\mathbf{x}}_t\right),{\Sigma }^{\pi }\left({\mathbf{x}}_t\right)\right)$，$p_i\left({\mathbf{u}}_t\mid {\mathbf{x}}_t\right)=\mathcal{N}\left({\mathbf{K}}_{ti}{\mathbf{x}}_t+{\mathbf{k}}_{ti},{\mathbf{C}}_{ti}\right)$，$i$为condition的数量，$j$为采样数量。

展开$p_i\left({\mathbf{u}}_t\mid {\mathbf{x}}_t\right)$可得：
$$\begin{array}{rl}{p}_i\left({\mathbf{u}}_t\mid {\mathbf{x}}_t\right)& =\mathcal{N}\left({\mathbf{K}}_{ti}{\mathbf{x}}_t+{\mathbf{k}}_{ti},{\mathbf{C}}_{ti}\right)\\ & =\frac{1}{\sqrt{(2\pi {)}^m}|{\mathbf{C}}_{ti}|}\exp (-\frac{1}{2}({\mathbf{u}}_t-({\mathbf{K}}_{ti}{\mathbf{x}}_t+{\mathbf{k}}_{ti}){)}^T{\mathbf{C}}_{ti}^{-1}({\mathbf{u}}_t-({\mathbf{K}}_{ti}{\mathbf{x}}_t+{\mathbf{k}}_{ti})))\end{array}$$

接下来有：
$$\begin{array}{rl}& D_{\mathrm{K}\mathrm{L}}\left({\pi }_{\theta }\left({\mathbf{u}}_t\mid {\mathbf{x}}_{t,i,j}\right)\Vert p_i\left({\mathbf{u}}_t\mid {\mathbf{x}}_{t,i,j}\right)\right)\\ & =\int {\pi }_{\theta }\ln \frac{{\pi }_{\theta }}{p}\\ & =-\int {\pi }_{\theta }\ln p-(-\int {\pi }_{\theta }\ln {\pi }_{\theta })\\ & =-{\mathbb{E}}_{{\pi }_{\theta }}\left[\ln p\right]-\mathcal{H}({\pi }_{\theta })\\ & ={\mathbb{E}}_{{\pi }_{\theta }}\left[\frac{1}{2}\ln ((2\pi {)}^m|{\mathbf{C}}_{ti}|)+\frac{1}{2}({\mathbf{u}}_t-({\mathbf{K}}_{ti}{\mathbf{x}}_t+{\mathbf{k}}_{ti}){)}^T{\mathbf{C}}_{ti}^{-1}({\mathbf{u}}_t-({\mathbf{K}}_{ti}{\mathbf{x}}_t+{\mathbf{k}}_{ti}))\right]-\mathcal{H}({\pi }_{\theta })\end{array}$$

由多变量高斯分布之间的KL散度（KL Divergence）知：

$$\begin{array}{rl}& D_{\mathrm{K}\mathrm{L}}\left({\pi }_{\theta }\left({\mathbf{u}}_t\mid {\mathbf{x}}_{t,i,j}\right)\Vert p_i\left({\mathbf{u}}_t\mid {\mathbf{x}}_{t,i,j}\right)\right)\\ & =\frac{1}{2}\ln ((2\pi {)}^m|{\mathbf{C}}_{ti}|)+\frac{1}{2}(\text{tr}({\mathbf{C}}^{-1}{\Sigma }^{\pi }({\mathbf{x}}_{t,i,j}))+({\mu }^{\pi }({\mathbf{x}}_{t,i,j})-{\mu }_{ti}^p({\mathbf{x}}_{t,i,j}){)}^T{\mathbf{C}}_{ti}^{-1}({\mu }^{\pi }({\mathbf{x}}_{t,i,j})-{\mu }_{ti}^p({\mathbf{x}}_{t,i,j}))-\frac{1}{2}\ln |{\Sigma }^{\pi }({\mathbf{x}}_{t,i,j})|-\text{const}\end{array}$$

所以：
$$\begin{array}{r}{\pi }_{\theta }\leftarrow \arg \underset{\theta }{\min }\sum _{t,i,j}(\text{tr}({\mathbf{C}}^{-1}{\Sigma }^{\pi }({\mathbf{x}}_{t,i,j}))+({\mu }^{\pi }({\mathbf{x}}_{t,i,j})-{\mu }_{ti}^p({\mathbf{x}}_{t,i,j}){)}^T{\mathbf{C}}_{ti}^{-1}({\mu }^{\pi }({\mathbf{x}}_{t,i,j})-{\mu }_{ti}^p({\mathbf{x}}_{t,i,j}))-\ln |{\Sigma }^{\pi }({\mathbf{x}}_{t,i,j})|\end{array}$$