ddpg.tex

\begin{algorithm}[t!]
\caption{Deep Deterministic Policy Gradient (DDPG)}
\begin{algorithmic}[1]
\label{alg:ddpg}
\WHILE{true}
    \STATE Take some action $a_i$ and observe $(s_i,a_i,s'_i,r_i)$ and add it to $\mathcal{B}$
    \STATE Sample mini-batch $\{s_j,a_j,s'_j,r_j\}$
    \STATE Compute $y_j = r_j + \gamma Q_{\phi'}(s'_j,a'_j)$ using target networks $Q_{\phi'}$ and $\mu_{\theta'}$
    \STATE $\phi \leftarrow \phi-\alpha\Sigma_j\frac{dQ_\phi}{d\phi}(s_j,a_j)(Q_\phi(s_j,a_j) - y_J)$
    \STATE $\theta \leftarrow \theta + \beta\sum_j\frac{d\mu}{d\theta}(s_j)\frac{dQ_\phi}{da}(s_j,a)$
    \STATE update $\phi'$ and $\theta'$
\ENDWHILE
\end{algorithmic}
\end{algorithm}