Learning the Probability Distribution for Probabilistic Expert Systems

Baker, Michelle; Roberts, Kenneth S.

This paper describes a method for learning the joint probability distribution of a set of variables from a sample of instances from the domain. The method is based on a straightforward application of Bayes Law to the problem of estimating individual probabilities from a probability distribution. We use a maximum entropy distribution as an initial estimate and show how this estimate can be easily updated each time an additional example is observed. Although developed for the purpose of estimating the conditional probabilities required for Bayesian inference networks, this method can be adopted to simplify knowledge acquisition in any expert system that uses knowledge in the form of probabilities.



More About This Work

Academic Units
Computer Science
Department of Computer Science, Columbia University
Columbia University Computer Science Technical Reports, CUCS-490-89
Published Here
January 17, 2012