Neural network software technology trainable on the random search and gradient descent principles

Abstract

The paper considers an applied problem related to the construction of efficient neural network technologies implemented in the traditional frameworks' standards. It is shown that the increase in efficiency is achieved due to the additional inclusion in the framework's structure of training algorithms based on the ideas of random search. Original implementations of such algorithms are proposed, with experimental confirmation of their effectiveness. It is shown that in this case not only the solutions' obtained quality increases, but it is also possible to extend the range of applied problems to be solved.