Method of Adjusting Linear and Nonlinear Classifiers . Einstellverfahren fuer Lineare und Nichtlineare Klassifikatoren

摘要

The problem of adjusting the parameters of a classifier so that a given pattern set is recognized correctly, is treated as that of transducing a closed-loop adaptive system from an arbitrary initial state into a solution state, which is defined by the maximum of a gain function. The adaptive system consists of an adjustable threshold classifier preceded by a Boolean network, an adjustment unit, and an evaluation unit which calculates the gain function. Training algorithms are suitable for solving a great variety of problems which are described and interpreted as interactions between these units. Examples from threshold logic demonstrate both the linear and the nonlinear methods; results of practical pattern recognition problems are also discussed. (Author)