9.6:定理9.5的证明
首先,由
\[v_\pi(s)=\sum_{a\in\mathcal{A}}\pi(a|s,\theta)q_\pi(s,a)\]
可得
\[\begin{aligned}
\nabla_\theta v_\pi(s)
&=\nabla_\theta\left[\sum_{a\in\mathcal{A}}\pi(a|s,\theta)q_\pi(s,a)\right]\\
&=\sum_{a\in\mathcal{A}}\left[\nabla_\theta\pi(a|s,\theta)q_\pi(s,a)+\pi(a|s,\theta)\nabla_\theta q_\pi(s,a)\right].
\end{aligned}\tag{9.29}\]
其中行动值满足
\[\begin{aligned}
q_\pi(s,a)
&=\sum_r p(r|s,a)(r-\bar{r}_\pi)+\sum_{s'}p(s'|s,a)v_\pi(s')\\
&=r(s,a)-\bar{r}_\pi+\sum_{s'}p(s'|s,a)v_\pi(s').
\end{aligned}\]
由于\(r(s,a)=\sum_r rp(r|s,a)\)与\(\theta\)无关,因此
\[\nabla_\theta q_\pi(s,a)
=-\nabla_\theta\bar{r}_\pi+\sum_{s'\in\mathcal{S}}p(s'|s,a)\nabla_\theta v_\pi(s').\]
代入式\((9.29)\)可得
\[\begin{aligned}
\nabla_\theta v_\pi(s)
&=\sum_{a\in\mathcal{A}}\nabla_\theta\pi(a|s,\theta)q_\pi(s,a)
-\nabla_\theta\bar{r}_\pi\\
&\quad+\sum_{a\in\mathcal{A}}\pi(a|s,\theta)\sum_{s'\in\mathcal{S}}p(s'|s,a)\nabla_\theta v_\pi(s').
\end{aligned}\tag{9.30}\]
令
\[u(s)\doteq \sum_{a\in\mathcal{A}}\nabla_\theta\pi(a|s,\theta)q_\pi(s,a).\]
由于
\[\sum_{a\in\mathcal{A}}\pi(a|s,\theta)\sum_{s'\in\mathcal{S}}p(s'|s,a)\nabla_\theta v_\pi(s')
=\sum_{s'\in\mathcal{S}}p(s'|s)\nabla_\theta v_\pi(s'),\]
式\((9.30)\)可写成矩阵-向量形式:
\[\nabla_\theta v_\pi=u-\mathbf{1}_n\otimes\nabla_\theta\bar{r}_\pi+(P_\pi\otimes I_m)\nabla_\theta v_\pi.\]
因此
\[\mathbf{1}_n\otimes\nabla_\theta\bar{r}_\pi
=u+(P_\pi\otimes I_m)\nabla_\theta v_\pi-\nabla_\theta v_\pi.\]
在上式两边左乘\(d_\pi^T\otimes I_m\),可得
\[\begin{aligned}
(d_\pi^T\mathbf{1}_n)\otimes\nabla_\theta\bar{r}_\pi
&=(d_\pi^T\otimes I_m)u+(d_\pi^TP_\pi\otimes I_m)\nabla_\theta v_\pi-(d_\pi^T\otimes I_m)\nabla_\theta v_\pi\\
&=(d_\pi^T\otimes I_m)u,
\end{aligned}\]
其中使用了\(d_\pi^T\mathbf{1}_n=1\)和\(d_\pi^TP_\pi=d_\pi^T\)。于是
\[\begin{aligned}
\nabla_\theta\bar{r}_\pi
&=(d_\pi^T\otimes I_m)u\\
&=\sum_{s\in\mathcal{S}}d_\pi(s)u(s)\\
&=\sum_{s\in\mathcal{S}}d_\pi(s)\sum_{a\in\mathcal{A}}\nabla_\theta\pi(a|s,\theta)q_\pi(s,a).
\end{aligned}\]
最后使用
\[\nabla_\theta\pi(a|s,\theta)=\pi(a|s,\theta)\nabla_\theta\ln\pi(a|s,\theta),\]
即可得到
\[\nabla_\theta\bar{r}_\pi
=\mathbb{E}[\nabla_\theta\ln\pi(A|S,\theta)q_\pi(S,A)],\]
其中\(S\sim d_\pi\),\(A\sim\pi(S,\theta)\)。