Lazy propagation and independence of causal influence

Abstract

The effciency of algorithms for probabilistic inference in Bayesian networks can be improved by exploiting independence of causal in uence. In this paper we propose a method to exploit independence of causal in uence based on online construction of decomposition trees. The effiency of inference is improved by exploiting independence relations induced by evidence during decomposition tree construction. We also show how a factorized representation of independence of causal influence can be derived from a local expression language. The factorized representation is shown to fit ideally with the lazy propagation framework. Empirical results indicate that considerable effciency improvements can be expected if either the decomposition trees are constructed online or the factorized representation is used.