Analysis of Learning Algorithms for Multilayer Neural Networks

Abstract

The modern stage of development of science and technology is characterized by a rapid increase in the complexity of the created technical systems. The management of such systems requires the development of new management methods, since the modification and improvement of traditional management techniques does not always ensure the fulfillment of stringent requirements for management quality indicators. Classical control methods are mainly based on the theory of linear systems, while most real objects are non-linear. The problem of the synthesis of control systems under conditions of uncertainty is currently one of the central problems in the modern theory of automatic control. The complexity of the control object itself, structural, parametric and information uncertainties in the description of the control object, and the complexity of control problems, the multi criteria of optimization problems, the lack of possible analytical solutions, the need to take into account all the properties of disturbances, etc. The solution to this problem requires a search for alternative approaches to the design of control systems, one of which involves the introduction of neural network systems. Neural network control systems are a high-tech direction of control theory and belong to the class of nonlinear dynamic systems. High performance due to parallelization of input information in combination with the ability to train neural networks makes this technology very attractive for creating control devices in automatic systems. Neural networks can be used to build regulating and switching devices, reference, adaptive, nominal and inverse-dynamic models of objects, on the basis of which objects are studied, analysis of the influence of disturbances acting on an object, determination of the optimal control law, search or calculating the optimal program for changing the impact when changing the values of the parameters of the object and the characteristics of the input data. In addition, neural networks can be used to identify objects, predict the state of objects, recognize, cluster, classify, analyze a large amount of data arriving at high speed from a large number of devices and sensors, and the like. The ability to learn according to a given principle of functioning allows creating automated control systems that are optimal in terms of speed, energy consumption, etc. Naturally, in this case, it is possible to implement several principles of functioning and the transition from one to another. They are a universal tool for modeling multidimensional nonlinear objects and finding solutions to ill-posed problems.