Identification Based on the Optimal Information Structure of Non-Stationary Objects

Scientific and methodological foundations have been developed for optimizing the identification of non-stationary objects based on the optimal information structure with mechanisms for selecting an adequate model, reducing excess space, highlighting, dividing random time series, clustering, and forecasting, subject to a priori limitations and parametric uncertainty of processes. The effectiveness of using identifiers based on dynamic models combined with a neural network has been studied. Algorithms for combined statistical and dynamic models and neural networks of various architectures have been implemented. Tools for modeling processes and presenting statistical, dynamic and specific characteristics were obtained. The computational schemes of network components with mechanisms for setting variables, optimization, and adaptive network training have been improved.v A cognitive analysis of variables has been implemented taking into account structural complexity, stochasticity of element connections based on a stochastic regression model, by tracking the ambiguity of dynamics. Optimization problems based on the probabilistic risk function and sliding control were solved. A generalized algorithm for identifying non-stationary objects is constructed based on tools for identifying boundaries, the general range of element values, selecting informative elements, and generating training sets. An identification software package was implemented in C++ in the parallel computing environment “CUDA”.