NUMERICAL METHODS OF TEACHING OF ARTIFICIAL NEURAL NETWORKS

The problem of modeling and training of artificial neural network of enough general topology, which dynamics is described by the system of the integro-differential equations is solved. The training task of artificial neural network is examined as an optimal control problem with unfixed time. The requisite optimum condition of weight numbers of artificial neural network and the external control action taking into account the defined constraints and a type of functionality are formulated. The initial task is reduced to a discrete problem of optimum control, which is solved on the basis of methodology of fast automatic differentiation. The analysis of impact of the task parameters on the optimal resolution is carried out.