A. L. Tulupyev, A. A. Filchenkov, N. A. Valtman
Algebraic Bayesian networks (ABN) are graphic probabilistic models of knowledge with uncertainty patterns bases. The ABN local and global automated learning algorithm system development is an actual scientific problem. The term of automated learning consists of processes of 1) system (model, graphic model) structure design and 2) such a system (model) parameters values adjustment. The local learning is a knowledge pattern learning on selected data. The learning samples might have absence of one or more elements, in that case it’s impossible to obtain probability scalar estimates without subsidiary assumptions. When there is no possibility to use a sample without gaps, work with probability interval estimates gets indispensable. There are algorithms for processing a learning sample with gaps that return knowledge patterns conjuncts probability interval estimations. The global learning is the network learning by means of designing knowledge pattern set on a learning sample and designing ABN secondary structure that is a graph on knowledge patterns. There are almost no results for the first part, while the algorithms of synthesizing a particular minimal join graph and synthesizing min-imal join graph set that contains “best” secondary structure are known for the second part of global learning, but the algorithm looking for only the “best” secondary structure is needed.