Article 4419
Title of the article |
REAL-TIME NEURAL NETWORK CONTROL METHOD BASED ON THE SYNTHESIS OF THE COST FUNCTION |
Authors |
Darina Anna Nikolaevna, candidate of physical and mathematical sciences, associate professor, leading researcher, Federal research center «Computer science and control» of RAS (Dorodnitsyn computer center of the Russian Academy of Sciences) (119333, 40 Vavilova street, Moscоw, Russia), Е-mail: daryina@ccas.ru |
Index UDK |
658.62.018.012 |
DOI |
10.21685/2307-4205-2019-4-4 |
Abstract |
When working with an unmanned vehicle, the task is usually to find such a control that carries out its transition from the initial state to the given final state. The developed control laws should provide the required quality indicators (accuracy, speed, etc.) for all controlled coordinates, taking into account the set control restrictions and the condition of the unmanned vehicle. From a practical point of view, it is very important that these control laws are optimal with respect to a given quality functional. However, in some cases this is not enough and it is necessary to synthesize control laws that ensure the achievement of the control goal in a wide class of uncertainty of the dynamics model of an unmanned vehicle. In this case, the control system should provide control taking into account dynamic constraints in real time, the parameters of which are not known in advance. The solution of such problems by classical methods is complicated by the fact that a large number of calculations are required, and they cannot be implemented on board the robot during its operation. This article discusses the solution of the quasi-optimal control problem in real time; for this, the traditional approach to the synthesis of control laws and the approximation of neural networks to these laws are combined taking into account information that the traditional approach does not take into account. As a result, the method consists in sequentially performing three stages: generating trajectories, synthesizing the selection function and training the neural network. The selection function is determined by the evolutionary methods of symbolic regression, and takes into account dynamic constraints in real time. For this, a neural network representation of the algorithms of software, adaptive and intelligent control is used, i.e. an integrated network in which a direct distribution neural network imitates, for example, a proportional integral differential controller, and a convolutional network is used to determine the characteristics of the road and make the right management decisions. An example of applying the method to a real robot is given. |
Key words |
unmanned vehicle, neural network control, trajectory generator, synthesis of the cost function, optimal control |
Download PDF |
Дата обновления: 04.02.2020 10:00