STUDY OF SCENARIOS OF COOPERATIVE LEARNING OF DISTRIBUTED SYSTEMS COORDINATORS

Distributed systems are becoming more and more common in energy, telecommunications, transportation, etc. Centralized management of such systems is unreliable in war, climatic disasters, cyber attacks and other critical circumstances. Decentralization allows to increase the stability and reliability of distributed systems, but it requires management coordination. Coordinators determine the optimal setting for local control systems, for which they provide forecasting, clustering, estimation of uncertain parameters, system modeling, and optimization modules. For the implementation of coordination methods, learning systems are preferred, as they are able to flexibly adjust to the peculiarities of managing objects in different conditions. However, the training of coordinators is complicated by the need to train many coordinators. This paper examines five neural coordinator training scenarios. The scenarios differ depending on the structure of the neural coordinators: a segmented network according to the structure of the simulation model of the coordinator or an integrated network, as well as on the learning strategy: train all the coordinators of the decentralized system in parallel or only one and clone the results; cooperative training of a group of coordinators. Experimental studies of the proposed method of training neural network coordinators, implemented on Python TensorFlow, were carried out. The distributed system of linear objects and the connection of the heat conduction type were investigated. RMSE over a given number of training epochs was used as a criterion for learning efficiency. The study showed greater effectiveness of cooperative learning of segmented coordinators, however, the effectiveness depends on the connectivity index of the objects of the distributed system.