О выборе архитектуры нейрорегулятора

Abstract

In practice, the choice of the type of neural network is carried out empirically based on an experience of an investigator and many training attempts. At the same time, the excessive complexity of the neural network leads to an increase in its training time, and in some cases, to the impossibility of learning at all. Thus, the justification of the choice of an artificial neural network structure and/or its preliminary calculation based on other models is an urgent task. An equally important task is the choice of an initial weighting coefficients of an neural network, the choice of which determines the speed of convergence of search algorithms. This paper demonstrates several approaches to solving the problem of choosing an architecture and initializing a weighting coefficients of a neural network. One of them is carried out on the basis of a previously calculated function using Petri nets. This approach is demonstrated for solving various tasks, which include the implementation of functions using previously defined neural network models of the simplest logical operations "and", "or", etc. An approach is given that allows optimizing an architecture of a neural network that solves the problem of approximating functions of one and several variables. The principle of determining an architecture and initial weight coefficients is also used in the tasks of training neural networks with reinforcement. A separate section is devoted to the formation of a methodology for determining an architecture and initialization of a weighting coefficients of a neural network of the controller based on information about the controller obtained by a modal method using a polynomial matrix decomposition of a system. The problem of synthesis of a neural network controller for an object model containing nonlinearities and nonparametric uncertainties in the control channel is solved.