proposal.tex

\documentclass[11pt]{article}

%<<< packages
\usepackage[margin=1in]{geometry}
\usepackage{titlesec}
\usepackage{titling}
%>>>

\setlength{\droptitle}{-1.9cm}

\title{Final Project Proposal}
\author{Teo Asinari, Yukun Lin}
\date{April 1, 2014}
\begin{document}

\maketitle

\section{Introduction}

The task of a controller is to determine and manipulate the inputs of a system in a way
such that the output of that system matches with some desired outputs. For the purpose of
maintaining a stable desired output, PID (proportional, integral, derivative) controllers
are typically used.

\section{Problem Statements}
Our proposed project is to build and train a neural networks to serve the purpose of a
controller. The network will attempt to stabilize an inverted pendulum on a moving
platform. Below are some tentative problem statements.

In all three problems, we will start with a simple ANN e.g., Adaline, and move on in
complexity (next step is MLP) until we stop seeing improvements.  We plan to use Python
for the implementation (Theano or Scikitlearn).

\subsection{Problem 1}
Construct, train network to imitate the behaviour of a PID controller. For example, given
the desired error (of the system) as a function of time, train a network to output control
signals that will be able to achieve that. If we take the error of a PID controller as
that desired error, then training the network to achieve that may give it the behaviour of
a PID controller.

\subsection{Problem 2}
For a PID controller to be stable, the constants it uses for its proportional, integral
and derivative term must be properly chosen. Thus, we propose training a network that can
determine the optimal constants for a PID controller, where the optimal constants are
those that achieve system stability in the shortest amount of time.

\subsection{Problem 3}
Extend the works of problem 1 and 2 by using a genetic algorithm to train the network.



\section{Reference}

http://deeplearning.net/tutorial/DBN.html\\
http://www.iro.umontreal.ca/\textasciitilde bengioy/papers/ftml\_book.pdf\\
http://www.cs.toronto.edu/\textasciitilde ranzato/publications/DistBeliefNIPS2012\_withAppendix.pdf\\
http://www.cs.toronto.edu/\textasciitilde hinton/absps/fastnc.pdf\\
http://www.iro.umontreal.ca/\textasciitilde lisa/pointeurs/theano\_scipy2010.pdf\\
https://www.stanford.edu/class/ee373b/NNselflearningcontrolsystems.pdf\\
http://research.cs.wisc.edu/machine-learning/shavlik-group/scott.NeuralComp92.pdf\\
http://archive.ics.uci.edu/ml/index.html\\
\end{document}