Summary: The Acyclic Bayesian Net Generator
Pankaj B. Gupta1
and Vicki H. Allan2
Microsoft Corporation, One Microsoft Way, Redmond, WA 98052, USA,
Computer Science Department, Utah State University, Logan, UT 84322,
Abstract. We present the Acyclic Bayesian Net Generator, a new ap-
proach to learn the structure of a Bayesian network using genetic algo-
rithms. Due to the encoding mechanism, acyclicity is preserved through
mutation and crossover. We present a detailed description of how our
method works and explain why it is better than previous approaches.
We can efficiently perform crossover on chromosomes with different node
orders without the danger of cycle formation. The approach is capable of
learning over all variable node orderings and structures. We also present a
proof that our technique of choosing the initial population semi-randomly
ensures that the genetic algorithm searches over the whole solution space.