Maximum correlation methods in constructing nonlinear models
Ultimately, the goal of system identification is to determine the effectiveness of methods, such as statistically based fundamental regularity assessments, for extracting knowledge from empirical data on the processes occurring in the systems under study. The need to generate a formal description of the object under study from experimental data dictates the use of system identification tools. At the same time, the inherent incompleteness and approximate nature of any mathematical representation of a real process make probabilistic-statistical approaches appropriate. When solving problems of identifying systems with random components, the need inevitably arises for quantitative characteristics of the statistical relationships between random variables — dependence measures. The most common tools in this field are classical linear covariance and correlation indices. Their use naturally follows from the formulation of the identification problem itself, which relies on a quadratic quality functional. The main advantage of linear correlation measures is their computational accessibility: on the one hand, they allow the derivation of closed-form analytical relationships for determining the desired system parameters, and on the other, they enable the construction of simple statistical estimates for both independent and dependent samples. However, a significant limitation of linear correlation measures is that they may vanish in the presence of a stochastic or even a functional (deterministic) relationship between the analyzed random variables. The need to eliminate this limitation motivates the introduction of more meaningful relationship characteristics into stochastic system identification procedures that account for the nonlinear dependencies among the variables under study. Moreover, the nonlinear nature of interactions is largely determined by the complex structure of the probabilistic relationships between the variables involved in the constructed models. This paper is devoted to the construction of nonlinear stochastic models that satisfy the conditions for the detection and correct quantitative description of stochastic dependencies between the model’s input and output variables. Herein lies the novelty of the proposed approach: the application of a consistent dependence measure — namely, maximal correlation — to problems of nonlinear identification.
Authors
- K. R. Chernyshev
Institutions
- V. A. Trapeznikov Institute of Control Sciences (RU)
Publication Details
- Journal
- Industrial laboratory Diagnostics of materials
- Published
- 2026-11-20
- DOI
- https://doi.org/10.26896/1028-6861-2026-92-11-84-92
- Primary Topic
- Control Systems and Identification
- Type
- article
- Field-Weighted Citation Impact
- 0.00