Efficient inferencing for sigmoid Bayesian networks by reducing sampling space

Abstract

A sigmoid Bayesian network is a Bayesian network in which a conditional probability is a sigmoid function of the weights of relevant arcs. Its application domain includes that of Boltzmann machine as well as traditional decision problems. In this paper we show that the node reduction method that is an inferencing algorithm for general Bayesian networks can also be used on sigmoid Bayesian networks, and we propose a hybrid inferencing method combining the node reduction and Gibbs sampling. The time efficiency of sampling after node reduction is demonstrated through experiments. The results of this paper bring sigmoid Bayesian networks closer to large scale applications.