Inductive Learning of Probabilistic Model and Its Application to Diagnostic Systems

Abstract

This paper describes an inductive learning method in probabilistic domain. It acquires an appropriate probabilistic model from a small amount of observation data. In order to derive an appropriate probabilistic model, a presumption tree with least description length is constructed. The description length of a presumption tree is defined as sum of the code length and the log-likelihood. Using the derived presumption tree, the probabilistic distribution of future events can be presumed appropriately. This capability enables improving the efficiency of certain kinds of performance systems, such as diagnostic systems, that deal with probabilistic problems.