Kussul A. Software tools for modeling of modular neural networks and their application.

The thesis is devoted to the development and a concrete definition of the mathematical model for а graph of modular neural networks. The representation intended for the general analysis of architecture of modular networks and, first of all, for the analysis of architecture by automatic systems is offered. The general properties of cycles in modular networks are considered. Algorithms of determination of run's order of the modular networks containing cycles are suggested. Analysis algorithms of the modular networks, allowing defining presence of cycles, allocating belonging vertices and defining types of cycles are offered. The developed models of modular neural networks and methods are realized within the system of automatic designing multimodular neural networks CAD MNN. For the first time methods of cooperative training of modular neural networks are experimentally investigated. It is shown, that use of cooperative training reduces time expenses in 1.5 times in comparison with individual at close results of classification. Advantages of modular networks on an example of the decision of a problem of classification of the passenger of the car are experimentally proved.