Abstract:
The article focuses on design of experiment for conducting Sensitivity Analysis of mathematical models used in forecasting and controlling complex systems. Special attention is given to cases where factors are unevenly distributed in space, which is typical for problems with nonlinear dependencies, local features, and high computational complexity. In such situations, the application of mathematical remodeling is justified, whereby models with a complex structure are replaced (remodeled) by objects of a selected remodeling class that have a predefined structure, which allows unificating system research. The purpose of the study is the development and comparison of the design of experiment strategies aimed at improving the efficiency of Sensitivity Analysis. Methods adapted to uneven data distribution are considered. The foundation of sensitivity research is the analysis of finite fluctuations, built upon the application of the Lagrange mean value theorem. Numerical experiments on a test function and its neural network approximation confirmed that the proposed algorithms (central composite design and an adaptive Latin hypercube sampling-based method) enable highly accurate identification of significant factors, aligning with classical methods (Sobol indices, Morris method), while significantly reducing computational costs. It is shown that the remodeling approach refines sensitivity estimates and ensures a unified analysis procedure for models of complex structure.
Keywords:sensitivity analysis, design of experiment, remodeling.
UDC:519.7 BBK:
22.18
Received: August 1, 2025 Published: September 30, 2025
Fulltext:
PDF file (546 kB)