Bayesian Knowledge Discoverer

Bayesian Knowledge Discoverer (BKD) is implemented in Common Lisp, and is running on most operating systems, including Power Macintosh and Sun OS which were used for experiments in this report. It uses a method called Bound and Collapse to learn Bayesian networks from possibly incomplete data. There are three different search strategies for finding the network:

• greedy - bottom-up greedy search strategy.
• arc-reversal - a greedy search that for each node tries to add all the nodes which are not one of its ancestors in the graph.
• exhaustive - a greedy search that tries all the possible combinations of parents in the search space before deciding which one is best.

The input consists of a database file (.db) and a network description file (.bbn). The database is organized as a simple tab-separated database with missing values stated as NA. The feature names are given in the first line. The description file must state all nodes in the Bayesian network as well as the range of all the features. Background knowledge can be stated in the file as edges that are known to exist or known probabilities, which will then be updated in the learning process.

The network can be displayed as a graph on the Power Macintosh version, and all versions can save the network in Bayesian networks interchange format. In experiments we found that we had to reduce the range for variables, or the search for the network would be too time-consuming. A description of Bayesian Knowledge Discoverer can be found in [27]. The version used was 0.1b.