Data fitting problem under interval uncertainty in data

Abstract

We consider the data fitting problem under uncertainty, which is not described by probabilistic laws, but is limited in magnitude and has an interval character, i.e., is expressed by the intervals of possible data values. The most general case is considered when the intervals represent the measurement results both in independent (predictor) variables and in the dependent (criterial) variables. The concepts of weak and strong compatibility of data and parameters of functional dependence are introduced. It is shown that the resulting formulations of problems are reduced to the study and estimation of various solution sets for an interval system of equations constructed from the processed data. We discuss in detail the strong compatibility of the parameters and data, as more practical, more adequate to the reality and possessing better theoretical properties. The estimates of the function parameters, obtained in view of the strong compatibility, have a polynomial computational complexity, are robust, almost always have finite variability, and are also only partially affected by the so-called Demidenko paradox. We also propose a computational technology for solving the problem of constructing a linear functional dependence under interval data uncertainty and take into account the requirement of strong compatibility. It is based on the application of the so-called recognizing functional of the problem solution set — a special mapping, which recognizes, by the sign of the values, whether a point belongs to the solution set and simultaneously provides a quantitative measure of this membership. The properties of the recognizing functional are discussed. The maximum point of this functional is taken as an estimate of the parameters of the functional dependency under construction, which ensures the best compatibility between the parameters and data (or their least discrepancy). Accordingly, the practical implementation of this approach, named «maximum compatibility method», is reduced to the computation of the unconditional maximum of the recognizing functional — a concave non-smooth function. A specific example of solving the data fitting problem for a linear function from measurement data with interval uncertainty is presented.