Dear Svetlana, regarding your first question: as I understand it from theoretical viewpoint, since equivalence classes of networks describe the same families of distributions, so the ML parameter configuration for the equivalent networks should (in theory) describe the same distribution and thus probability accessments given the evidence should be the same. Now the question is: "Do you observe this in practice?". If not, I would suppose that it is due to approximation/numerical issues in beleif updating.
Dmitry Rusakov, Technion. Svitlana Bulashevska wrote: > Dear Sirs, > I have a question concerning the equivalence classes of Bayesian > Networks. Since learning of Bayesian Networks can find correct network > structures up to their equivalence classes, so that in a one found > network some arcs can be reversible, how reliable is then the > probability assessement given evidence (belief updating) in the learned > Bayesian Network? > One more question. I know that there exists an algorithm of Chickering > transferring a DAG into PDAG that defines uniquely the equivalence class > of DAG. Is there any software to do this automatically? > Thank you very much for the help, > Svetlana Bulashevska, Ph.D. Student > >
