8.3:引理8.1的证明
利用全期望公式,有
\[\begin{aligned}
&\mathbb{E}\left[\left(r_{t+1}+\gamma\phi^T(s_{t+1})w_t-\phi^T(s_t)w_t\right)\phi(s_t)\right]\\
&=\sum_{s\in\mathcal{S}}d_\pi(s)\mathbb{E}\left[r_{t+1}\phi(s_t)|s_t=s\right]\\
&\quad+\sum_{s\in\mathcal{S}}d_\pi(s)\mathbb{E}\left[\phi(s_t)\left(\gamma\phi^T(s_{t+1})-\phi^T(s_t)\right)w_t|s_t=s\right].
\end{aligned}\tag{8.24}\]
这里假设\(s_t\)服从稳态分布\(d_\pi\)。
先看式\((8.24)\)中的第一项。注意
\[\mathbb{E}[r_{t+1}\phi(s_t)|s_t=s]=\phi(s)\mathbb{E}[r_{t+1}|s_t=s]=\phi(s)r_\pi(s),\]
其中
\[r_\pi(s)=\sum_a\pi(a|s)\sum_r rp(r|s,a).\]
因此第一项可写为
\[\sum_{s\in\mathcal{S}}d_\pi(s)\phi(s)r_\pi(s)=\Phi^TDr_\pi.\tag{8.25}\]
再看第二项。由于
\[\begin{aligned}
&\mathbb{E}\left[\phi(s_t)(\gamma\phi^T(s_{t+1})-\phi^T(s_t))w_t|s_t=s\right]\\
&=-\phi(s)\phi^T(s)w_t+\gamma\phi(s)\mathbb{E}[\phi^T(s_{t+1})|s_t=s]w_t\\
&=-\phi(s)\phi^T(s)w_t+\gamma\phi(s)\sum_{s'\in\mathcal{S}}p(s'|s)\phi^T(s')w_t,
\end{aligned}\]
所以第二项变为
\[\begin{aligned}
&\sum_{s\in\mathcal{S}}d_\pi(s)\mathbb{E}\left[\phi(s_t)(\gamma\phi^T(s_{t+1})-\phi^T(s_t))w_t|s_t=s\right]\\
&=\sum_{s\in\mathcal{S}}d_\pi(s)\phi(s)\left[-\phi(s)+\gamma\sum_{s'\in\mathcal{S}}p(s'|s)\phi(s')\right]^Tw_t\\
&=\Phi^TD(-\Phi+\gamma P_\pi\Phi)w_t\\
&=-\Phi^TD(I-\gamma P_\pi)\Phi w_t.
\end{aligned}\tag{8.26}\]
结合式\((8.25)\)和式\((8.26)\),可得
\[\mathbb{E}\left[\left(r_{t+1}+\gamma\phi^T(s_{t+1})w_t-\phi^T(s_t)w_t\right)\phi(s_t)\right]
=\Phi^TDr_\pi-\Phi^TD(I-\gamma P_\pi)\Phi w_t.\]
令
\[b\doteq\Phi^TDr_\pi,\qquad A\doteq\Phi^TD(I-\gamma P_\pi)\Phi,\]
于是期望可写为
\[b-Aw_t.\]