\documentstyle[12pt]{article}
\input{psfig.sty}
\input epsf
\oddsidemargin 0cm % this is margin in addition to the usual 1 inch
\evensidemargin 0cm
\textwidth 16.5cm
%\oddsidemargin -0.5cm % this is margin in addition to the usual 1 inch
%\evensidemargin-0.5cm
%\topmargin -1.9cm
%\topmargin 0.0cm
%\baselineskip 0.0cm
%\textwidth 16.5cm
%\textwidth 17.0cm
%\textheight 24.7cm
\newcommand{\bi}{\begin{itemize}}
\newcommand{\ei}{\end{itemize}}
\newcommand{\be}{\begin{enumerate}}
\newcommand{\ee}{\end{enumerate}}
\newcommand{\tand}{\wedge}
\newcommand{\tor}{\vee}
\newcommand{\union}{\cup}
\newcommand{\intersection}{\cap}
\newcommand{\ch}{$\surd$}
\newcommand{\pgm}[1]{{\sc #1}}
\newcommand{\la}{\leftarrow}
\newcommand{\ra}{\rightarrow}
\begin{document}
\begin{center}
\large
{\bf Machine Learning 15-681/781}\\
Tom Mitchell, Fall 1998
\end{center}
\begin{center}
\large Homework Assignment 4\\
Due at beginning of class on Nov 17, 1998 \\
\ \\
\end{center}
\be
\item Exercise 5.2 from the textbook.
\item Exercise 5.4
\item Exercise 7.8
\item Exercise 6.5
\item This question asks you to consider the complexity of several
probabilistic approaches to learning the $Cup$ target concept summarized in
Table 12.3 (page 342) of the textbook. Note all the attributes in this
problem are boolean valued.
\bi
\item First consider using a Naive Bayes classifier to learn this target
concept. Exactly how many distinct probability terms must be estimated from
the training data to learn a Naive Bayes classifier for this problem?
\item
Now consider learning a Bayesian belief network to recognize $Cup$s, instead
of a Naive Bayes Classifier. In particular, use the network suggested by the
domain theory in Table 12.3, and interpret each rule in the domain theory as a
specification of the parents of the corresponding network variable. (e.g.,
the first rule in the domain theory specifies that $Parents(Cup) = \{ Stable,
Liftable, OpenVessel \}$).
Draw the Bayes network as a graph with directed edges. You may omit variables
that are not mentioned by the domain theory.
Exactly how many distinct probability terms must be estimated from the
training data to learn the conditional probability tables for your network?
\item
Both of the above approaches make conditional independence assumptions to
reduce the complexity of estimating $P(Cup | BottomIsFlat, ConcavityPointsUp,
Expensive, \ldots)$ from the training data. If no such assumptions are made,
how many distinct probability terms must be estimated from the training data?
\ei
\ee
\end{document}