Next: Experiments with Casebase Up: Results Previous: Construction of Casebase

Construction of Bayesian Network

From Rosetta, the range of each feature was exported to a separate file. This file was edited to make declarations for the Bayesian Network to be constructed by Bayesian Knowledge Discoverer. The database output file was edited to replace Rosetta-notation ``Undefined'' by BKD-notation ``NA'' for missing values.

This produced a network definition file, which is shown in Fig 5.1, and a database file. An excerpt is given in Fig 5.2. The network definition file is the background knowledge that exists before examining the database. The nodes in the network must be defined, with the range of each feature. It is also possible to define known edges with respective probabilities. In the network in Fig 5.1 no background knowledge is stated. The database file is a tab-separated file with a first line not shown in the figure which states the names of the features. These names must correspond with the feature names given in the network definition file.

**Figure 5.1:** The Network Definition File rcomp.bbn
$\begin{figure} \begin{center} \footnotesize \begin{verbatim} (in-package :l... ...`(CLASSI CLASSII CLASSIII))\end{verbatim} \normalsize \end{center} \end{figure}$

**Figure 5.2:** Excerpt From Data File.
$\begin{figure} \begin{center} \scriptsize \begin{verbatim} 1 NOT MECHFAIL G... ...LEAK MINOR LU LOW CE CLASSI\end{verbatim} \normalsize \end{center} \end{figure}$

The network definition file was then loaded in Bayesian Knowledge Discoverer which executed the following script:


logfile "rcomp.log"
load "rcomp.bbn"
model "rcomp.db"
nodes
save "utcomp.bbn"
exit

**Figure 5.3:** The Bayesian Network Produced.
$\begin{figure} \begin{center} \scalebox {0.5}{\includegraphics{nettverk.eps}} \end{center}\end{figure}$

BKD was run on a UltraSparc running Sun OS 5.5.1. It used approximately 12:05- hours to construct the network. The network is shown in Fig 5.3, and in the Bayesian information exchange format in appendix D. This produced the Bayesian network used in the experiments to follow.

Next: Experiments with Casebase Up: Results Previous: Construction of Casebase

Torgeir Dingsoyr
2/26/1998