Innokentiy V. Igumnov, Nicolai N. Kucyi

National research Irkutsk state technical university,

The purpose of this article is to eliminate the shortcomings in the neural network training algorithm, which include insufficiently accurate determination of the direction of movement [4], slow convergence to an extremum, and the need to use a sufficiently large number of initial simplexes. It is proposed to introduce an additional search direction into the neural network training algorithm, in relation to solving the problem of parametric optimization of artificial neural networks (ANN) contained in links with pulse width modulation (PWM) of automatic control systems. Due to the fact that ANNs are used in PWM, the tasks of training and parametric optimization are equivalent and ultimately come down to determining the weighting coefficients of the ANN.To achieve this goal, the following tasks were set and solved: 1) existing approaches used in direct search methods to improve their characteristics are analyzed; 2) conducting experiments on the use of the most common approaches, in the context of the problem of parametric optimization of systems with PWM controllers; 3) development of recommendations for their use. Ultimately, the above makes it possible to resolve the problems of speed and the number of initial simplexes that arise when solving the problem of parametric optimization of automatic control systems with a device that performs PWM using ANN. Based on the above, we can talk about the relevance of the presented article.

artificial neural network, pulse-width modulation, neural network training, Nelder-Mead method, integral criterion, quasi-gradient