\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[noend]{algpseudocode} \DeclareCaptionFormat{myformat}{#3} \captionsetup[algorithm]{format=myformat} \DeclareMathOperator*{\argmax}{arg\,max} \begin{document} \begin{preview} \begin{algorithm}[H] \begin{algorithmic} \Require \Statex States $\mathcal{X} = \{1, \dots, n_x\}$ \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 $\gamma \in [0, 1]$ \Procedure{QLearning}{$\mathcal{X}$, $A$, $R$, $T$, $\alpha$, $\gamma$} \State Initialize $Q: \mathcal{X} \times \mathcal{A} \rightarrow \mathbb{R}$ arbitrarily \While{$Q$ is not converged} \State Start in state $s \in \mathcal{X}$ \While{$s$ is not terminal} \State Calculate $\pi$ according to Q and exploration strategy (e.g. $\pi(x) \gets \argmax_{a} Q(x, a)$) \State $a \gets \pi(s)$ \State $r \gets R(s, a)$ \Comment{Receive the reward} \State $s' \gets T(s, a)$ \Comment{Receive the new state} \State $Q(s', a) \gets (1 - \alpha) \cdot Q(s, a) + \alpha \cdot (r + \gamma \cdot \max_{a'} Q(s', a'))$ \State $s \gets s'$ \EndWhile \EndWhile \Return $Q$ \EndProcedure \end{algorithmic} \caption{$Q$-learning: Learn function $Q: \mathcal{X} \times \mathcal{A} \rightarrow \mathbb{R}$} \label{alg:q-learning} \end{algorithm} \end{preview} \end{document}