Pattern classification in dynamical environments

Abstract

This paper presents a statistic pattern recognition and machine learning based classification system paradigm. It focuses on solving problems underlying environments where the attributes of the system exhibit strong uncertainties. Configurations of classifiers therefore have to undergo continual changes during the running process. The classifier discussed here is characterized by a tagged feature-class representation, a dimension-wise univariate discrimination scheme, a default hierarchy of classification, as well as a logic-based learning strategy. Concepts of distinguishability and univariately distinguishable classes under classification are studied. Procedures developed for forming an univariately distinguishable class-feature space by learning are described.