LaTeX-examples/source-code/Pseudocode/q-lambda/q-lambda.tex

\documentclass{article}
\usepackage[pdftex,active,tightpage]{preview}
\setlength\PreviewBorder{2mm}

\usepackage[utf8]{inputenc} % this is needed for umlauts
\usepackage[ngerman]{babel} % this is needed for umlauts
\usepackage[T1]{fontenc}    % this is needed for correct output of umlauts in pdf
\usepackage{amssymb,amsmath,amsfonts} % nice math rendering
\usepackage{braket} % needed for \Set
\usepackage{caption}
\usepackage{algorithm}
\usepackage{xcolor}
\usepackage[noend]{algpseudocode}
\usepackage{mathtools,bm}
\DeclareMathOperator*{\argmax}{arg\,max} 

\DeclareCaptionFormat{myformat}{#3}
\captionsetup[algorithm]{format=myformat}

\begin{document}
\begin{preview}
    \begin{algorithm}[H]
        \begin{algorithmic}
        \Require
        \Statex Sates $\mathcal{X} = \{1, \dots, n_x\}$
        \Statex Actions $\mathcal{A} = \{1, \dots, n_a\},\qquad A: \mathcal{X} \Rightarrow \mathcal{A}$
        \Statex Reward function $R: \mathcal{X} \times \mathcal{A} \rightarrow \mathbb{R}$
        \Statex Black-box (probabilistic) transition function $T: \mathcal{X} \times \mathcal{A} \rightarrow \mathcal{X}$
        \Statex Learning rate $\alpha \in [0, 1]$, typically $\alpha = 0.1$
        \Statex Discounting factor $\gamma \in [0, 1]$
        \Statex $\lambda \in [0, 1]$: Trade-off between TD and MC
        \Procedure{QLearning}{$\mathcal{X}$, $A$, $R$, $T$, $\alpha$, $\gamma$, $\lambda$}
            \State Initialize $Q: \mathcal{X} \times \mathcal{A} \rightarrow \mathbb{R}$ arbitrarily
            \State Initialize $e: \mathcal{X} \times \mathcal{A} \rightarrow \mathbb{R}$ with 0. \Comment{eligibility trace}
            % \State Start in state $s \in \mathcal{X}$
            \While{$Q$ is not converged}
                \State Select $(s, a) \in \mathcal{X} \times \mathcal{A}$ arbitrarily
                \While{$s$ is not terminal}
                    \State $r \gets R(s, a)$
                    \State $s' \gets T(s, a)$ \Comment{Receive the new state}
                    \State Calculate $\pi$ based on $Q$ (e.g. epsilon-greedy)
                    \State {$\color{red} a^* \gets \argmax_{\tilde{a}} Q(s', \tilde{a})$}
                    \State $a' \gets \pi(s')$
                    \State $e(s, a) \gets e(s, a) + 1$
                    \State $\delta \gets r + \gamma \cdot Q(s', {\color{red} a^*}) - Q(s, a)$
                    \For{$(\tilde{s}, \tilde{a}) \in \mathcal{X} \times \mathcal{A}$}
                        \State $Q(\tilde{s}, \tilde{a}) \gets Q(\tilde{s}, \tilde{a}) + \alpha \cdot \delta \cdot e(\tilde{s}, \tilde{a})$
                        \State ${\color{red} e(\tilde{s}, \tilde{a}) \gets \begin{cases}\gamma \cdot \lambda \cdot e(\tilde{s}, \tilde{a})&\text{if } a' = a^*\\
                                     0 &\text{otherwise}\end{cases}}$
                    \EndFor
                    \State $s \gets s'$
                    \State $a \gets a'$
                \EndWhile
            \EndWhile
            \Return $Q$
        \EndProcedure
        \end{algorithmic}
    \caption{SARSA($\lambda$): Learn function $Q: \mathcal{X} \times \mathcal{A} \rightarrow \mathbb{R}$}
    \label{alg:sarsa-lambda}
    \end{algorithm}
\end{preview}
\end{document}