Циклы в байесовских сетях: вероятностная семантика и отношения с соседними узлами

We consider directed (feedback) cycles in Bayesian belief networks, taking into consideration their interaction with successors and predecessors in the network. We also consider the case of multivalued variables. Our discussion leads to the conclusion that a cycle in a BBN may by itself influence its parents, making some of their assignments inconsistent. We also consider a sufficient condition of a cycle’s consistency in an algebraic Bayesian network. If it holds, the cycle may be transformed into a linear chain of knowledge patterns of constant size. We also make numerical examples, including an example of a cycle, some assignments of whose parents turn out to be inconsistent.