An approach to problem-oriented learning of a multi-agent system is proposed. The corresponding methods apply a kind of such a notion as knowledge and the introduced notion of φ-linear models that are a specific class of general nonlinear ones. Specifically, the values of transfer coefficients were shown to be adapted to the solved identification problem and are in sequel used by the multi-agent system to form the problem-oriented knowledge about the object under study, and updating the knowledge base.