Rethinking the Learning of Belief Network Probabilities
Ron Musick
Advanced Information Technology Program
Lawrence Livermore National Laboratory
P.O. Box 808, L-419, Livermore, CA 94551
rmusick@llnl.gov
Abstract
Belief networks are a powerful tool for knowledge discovery that
provide concise, understandable probabilistic models of data. There
are methods grounded in probability theory to incrementally update the
relationships described by the belief network when new information is
seen, to perform complex
inferences over any set of
variables in the data, to incorporate domain expertise and prior
knowledge into the model, and to automatically learn the model
from data.
This paper concentrates on part of
the belief network induction problem, that of learning the
quantitative structure (the conditional probabilities), given the
qualitative structure. In particular, the current
practice of rote learning the probabilities in belief networks can
be significantly improved upon. We advance the idea of applying any
learning algorithm to the task of conditional probability
learning in belief networks, discuss potential benefits, and show
results of applying neural networks and other algorithms to a medium
sized car insurance belief network. The results demonstrate from 10
to 100% improvements in model error rates over the current approaches.
Appeared in
KDD 1996