Generic Database for Hybrid Bayesian Pattern Recognition

Abstract

A Bayesian pattern recognition system is proposed, that processes information encoded by four types of features: discrete, pseudo-discrete, multi-normal continuous and independent continuous. This hybrid system utilizes the combined frequentist-subjective approach to probabilities, uses parametric and nonparametric techniques for the conditional likelihood estimation, and relies heavily on the fuzzy theory for data presentation, learning, and information fusion. The information for training, recognition, and prediction of the system is organized in a database, which is logically structured into three interconnected hierarchical sub-databases. A software tool is created under MATLAB that assures consistency, integrity, and maintenance of the database information. Three application examples from the fields of technical and medical diagnostics are presented, which illustrate the types of problems and levels of complexity that the database tool can handle.