Neural network software technology trainable on the random search principles

Abstract

The paper deals with a state-of-art neural technology programmed implementation problem in which the training process is based on random search algorithms. Training neural networks is a typical optimization problem. At the initial stage of neural network technologies development, various variants of gradient methods were traditionally used to solve such problems. Such methods, as a rule, met the requirements for the problem in terms of quality and speed of training. However, with the appearing of a new class of applied problems, the situation has changed. The traditional approach to training using gradient methods did not always meet the requirements of the applied problem in terms of the resulting solution quality.