simulation_model_complexity.tex

\documentclass[margin=1cm]{standalone}
\usepackage{tikz}

\definecolor{ACMYellow}{RGB}{255, 214, 0}
\definecolor{ACMOrange}{RGB}{252, 146, 0}
\definecolor{ACMRed}{RGB}{253, 27, 20}
\definecolor{ACMLightBlue}{RGB}{131, 206, 226}
\definecolor{ACMGreen}{RGB}{166, 188, 9}
\definecolor{ACMPurple}{RGB}{101, 1, 107}
\definecolor{ACMDarkBlue}{RGB}{9, 53, 122}

\begin{document}

\begin{tikzpicture}[scale=2, domain=-2:2, y=0.5cm]

\draw[->, thick] (-2.25,-0.25) -- (2.25,-0.25) node [below] {Model complexity};
\draw[->, thick] (-2.25,-0.25) -- (-2.25,5) node [midway, left] {Model error};

\draw[dashed, thick] (0,-0.25) node [below, text width=3cm, align=center] {Optimal model complexity} -- (0,5) ;

\draw[->, thick] (-0.5,4) -- (-1.5,4) node [midway, above] {Underfitting};
\draw[->, thick] (0.5,4) -- (1.5,4)node [midway, above] {Overfitting};

\coordinate (O) at (0,0);

\draw[ACMRed, thick] plot[samples=100] (\x, {2^\x + 2^(-\x)}) node[above right] {Overall model error};
\draw[ACMPurple, thick] plot[samples=100] (\x, {2^(\x)}) node[right] {Modelled noise (variance)};
\draw[ACMOrange, thick] plot[samples=100] (\x, {2^(-\x)}) node[right] {Unmodelled information (bias)};

\node[text width=12cm, anchor=north, yshift=-2cm] {\textbf{Caption:} Model error resulting from the superposition of two error curves: (i) the bias error which results from too simple models and can be mitigated by increasing the model complexity and (ii) the variance error which reduces the uncertainty with growing model complexity. };
\end{tikzpicture}


\end{document}