ABSTRACT

A model of conditional independence is studied as well as the problem of its identifiability when a definite subset of its parameter is hidden and not observable. The identification problem for such type of models is formulated as a special case of the well-known problem of non-supervised learning or self-learning. Commonly used procedures of self-learning are based on the maximum likelihood estimation of statistical parameters of the model. These procedures perform a certain type of hill-climbing and converge to the local rather than global maximum of the likelihood function. The main new result of the investigation is that for the model under consideration there exists an algorithm of self-learning, that finds a global maximum of likelihood function though the likelihood function is not unimodal.