The causal Markov condition, fact or artifact?

This paper provides a priori cirteria for determing when a causal model is sufficiently complete to be considered a Bayesian Network, and a new representation for Bayesian Networks shown to be more computationally efficient in a wide range of circumstances than current representations.Expert Systems for domains in which uncertainty plays a major role are often built from causal models. These models are usually implemented using Baysesian Network technology under the often tacit assumption that a causal model is a satisfactory Bayesian Network of the domain. If the system produces unsatisfactory results, the causal model is usually deemed inadequate, probably containing insufficient detail.The assumption that a causal model is an appropriate Bayesian Network model is justified by invoking the so called "Causal Markov Condition". In this paper we argue that in many cases it is not the inadequacy of the causal model which produces unsatisfactory results, but rather the inappropriateness of the Causal Markov Condition itself.In this paper we introduce a new functional model of causality, the Communicating Causal Process model, and analyze the appropriatenss of the Causal Markov Condition in light of this model. This analysis yelds domain based a priori criteria for judging when the Causal Markov Condition does or does not hold, and when a Causal Model is sufficiently detailed that it can be considered a Bayesian Network.The Communicating Causal Process model also provides the basis for a new representation of Bayes Networks which shown to be more computationally efficient than current representations.