Introduction to Bayesian Networks

Abstract

Bayesian networks are probabilistic causal models. The graph represents the structure of a domain knowledge, and probabilities represent the uncertain part of this domain. This chapter explains this idea with a very simple example in industrial safety. The first use of Bayesian networks is that of a “conditional probability calculator". Given a model, assumed to be constructed by an expert, the use of a Bayesian network amounts to calculating the probability of a nonobserved variable conditionally to the observed variables. A second domain of research for Bayesian networks is the automatic construction of models. This is a fascinating subject. Indeed, if we think about it, the notion of conditional probability also applies to models. From this point of view, Bayesian theory offers an answer to one of the most critical aspects of empirical modeling, which is the dialectic between observations and the model.