Robustness in Statistical Pattern Recognition

This book is concerned with important problems of robust (stable) statistical pat­ tern recognition when hypothetical model assumptions about experimental data are violated (disturbed). Pattern recognition theory is the field of applied mathematics in which prin­ ciples and methods are constructed for classification and identification of objects, phenomena, processes, situations, and signals, i. e. , of objects that can be specified by a finite set of features, or properties characterizing the objects (Mathematical Encyclopedia (1984)). Two stages in development of the mathematical theory of pattern recognition may be observed. At the first stage, until the middle of the 1970s, pattern recogni­ tion theory was replenished mainly from adjacent mathematical disciplines: mathe­ matical statistics, functional analysis, discrete mathematics, and information theory. This development stage is characterized by successful solution of pattern recognition problems of different physical nature, but of the simplest form in the sense of used mathematical models. One of the main approaches to solve pattern recognition problems is the statisti­ cal approach, which uses stochastic models of feature variables. Under the statistical approach, the first stage of pattern recognition theory development is characterized by the assumption that the probability data model is known exactly or it is esti­ mated from a representative sample of large size with negligible estimation errors (Das Gupta, 1973, 1977), (Rey, 1978), (Vasiljev, 1983)).

This monograph is devoted to problems of robust (stable) statistical pattern recognition. Experimental data to be classified usually deviate from assumed hypothetical probability models of the data. In such cases traditional decision rules constructed by means of the classical pattern recognition theory based on a fixed hypothetical model of the data often become non-stable, and the classification risk increases non-controllably. The book concentrates on three main problems: robustness evaluation for classical decision rules in the presence of distortion; estimation of critical levels of distortions for given values of the robustness factor; and the construction of robust decision rules with stable classification risk regarding certain types of distortions. Theoretical results are illustrated by computer modelling and by application to medical diagnostics. Audience: This volume is primarily intended for mathematicians, statisticians, and engineers in applied mathematics, computer science and cybernetics. It is also recommended as a textbook for a one-semester course for advanced undergraduate and graduate students training in the indicated fields.