main.tex

\documentclass[titlepage]{elsarticle}
\usepackage[utf8]{inputenc}
\usepackage{natbib}

\title{Machine Learning Controllers for Autonomous Surface Vessels}
\author{Haakon Robinson}
\date{November 2018}

\begin{document}
\maketitle

% \pagenumbering{roman}

\frontmatter
\begin{abstract}
  \input{abstract}
\end{abstract}


\pagebreak
\tableofcontents
\pagebreak

%\pagenumbering{arabic}
%\setcounter{page}{1}

\mainmatter
\section{Introduction}
\label{sec:intro}
Study the possibilities of applying reinforcement learning to controls
problems involving marine surface vessels.



\section{Background}
\label{sec:bg}

\subsection{Modeling Marine Surface Vessels}
\label{sec:bg-marine}
% The 6 DOF model
% The 3 DOF model
% Some control laws

\subsection{Deep Learning and Decision Making}
\label{sec:bg-ml}
% Good old fashioned AI (symbolic artificial intelligence)
% Neural network
% Deep neural network

\subsection{Reinforcement Learning}
\label{sec:bg-rl}
% Markov Decision processes
% Bellman equation
% Q learning
% Deep learning
\input{background/reinforcement-learning}



\section{Literature Review}
\label{sec:lit}

\subsection{Deep Reinforcement Learning}
\label{sec:lit-dl}
(Human level control through deep reinforcement learning)

\subsection{Set-Based Tasks for Controls}
\label{sec:lit-set-tasks}
Framework for implementing several modes of operation based on the
current active tasks. Allows for the use of several different control
systems at once, each one specialising in different combinations of
tasks. This is a good way to combine the strengths of rule-based and
ML-based controllers.



\section{Problem Formulation}
\label{sec:formulation}
In a navigation problem, one generally has a destination that you wish
to get your system to. If the environment is deterministic, using
some kind of path planning method like A* works well, as long as the
``obstacles'' are known. When the environment is dynamic and unknown,
this is not as easy. The ``global'' path may still be there, but there
might be many small local obstacles that need to be avoided. A local
avoidance system is therefore needed.

Potential fields might lead to oscillatory behavior.

Dynamic window assume no sideways velocity, which is difficult in the
context of ocean currents. Also computationally heavy, but works well
with COLREGS.



\section{Simulation and Testing Environment}
\label{sec:sim+test}



\section{Deep Learning Models}
\label{sec:learning-models}



\section{Training, Testing, and Validation}

\label{sec:training}


\section{Thoughts}
\input{main-thoughts}


\end{document}

%%% Local Variables:
%%% mode: latex
%%% TeX-master: t
%%% End: