Research of Recurrent Dynamic Neural Networks for Adaptive Control of Complex Dynamic Systems
Abstract
This report results from a contract tasking Institute of Mathematical Machines and Systems as follows: Main target of research are dynamic recurrent neural networks and their application for adaptive control of complicated dynamical systems. Recurrent neural networks, unlike classical methods, don't require full a priori information about properties of controlled object. So their use allows to achieve high precision and reliability for control of complicated dynamical objects in conditions of lack of information about internal state of the object. Distinctive feature of researched in this project recurrent neural networks is the use of dynamical neurons, that are able to ?forget╗ outdated information, and multimodular architecture, that allows to flexibly combine adaptation for changing environment properties and features of associative memory that keeps different controlled object's models. Neural associative memory is able to recognize large-scale changes of environment conditions, that allows to reconfigure network's structure in real time for more efficient control. Project includes the critical review of publications in the given area, theoretical research of training methods of dynamical recurrent neural networks, methods of reconfiguration of such networks, decomposition of neurocontrolling task, and experimental test of achieved theoretical results, that will be performed on MNN CAD with the use of specially developed models of RNN and controlled objects.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 08, 2010
- Accession Number
- ADA525703
Entities
People
- Alexander M. Reznik
Organizations
- National Academy of Sciences