rl_0727.tex

\documentclass{beamer}
\title{RL with initial quess}
\author{Yoon-gu Hwang}
\institute{LG CNS}
\date{\today}

\begin{document}
\begin{frame}
\frametitle{Sample frame title}
Initial guess from Open loop control results
\begin{itemize}
	\item DQN: Action-Value Function, $Q(s,a;\theta_0)$
	\begin{equation}
		\theta_0 = \arg\min_\theta L(Q(s,a;\theta), V(s,a))
	\end{equation}where $V(s,a)$ is the cost function produced by an open loop control.
	\item PPO: Policy Network, $\pi_0(s;\theta_0)$
	\begin{equation}
		\theta_0 = \arg\min_\theta L(\pi_0(s;\theta_0), u(s))
	\end{equation}where $u(s)$ is an action produced by an open loop control.
\end{itemize}
\end{frame}

\begin{frame}\frametitle{DQN; Value-based Iteration}
Assume that we have $(S(t), I(t), u^*(t))$ after applying a successful Pontraygin's principle algorithm.
We want an initial parameter $\theta_0$ such that
\begin{equation}
	V(S(t_i), I(t_i);\theta_0) \approx J(S(t), I(t), u(t))
\end{equation}

\end{frame}

\begin{frame}\frametitle{PPO; Policy Iteration}
Assume that we have $(S(t), I(t), u^*(t))$ after applying a successful Pontraygin's principle algorithm.
We want an initial parameter $\theta_0$ such that
\begin{equation}
	\pi(S(t_i), I(t_i);\theta_0) \approx u(t_i)
\end{equation}

\end{frame}
\end{document}