Improved Experimental Results Using Fuzzy Lattice Neurocomputing (FLN) Classifiers.pdfVIP

Intl. Conf. on Machine Learning; Models, Technologies and Applications (MLMTA’03) Las Vegas, Nevada, USA, 23-26 June 2003 Improved Experimental Results Using Fuzzy Lattice Neurocomputing (FLN) Classifiers Al Crippsa, V.G. Kaburlasosc, N. Nguyenb, and S.E. Papadakisc a Department of Computer Science acripps@ b Department of Economics and Finance Middle Tennessee State University, Murfreesboro, TN 37132, USA c Department of Industrial Informatics, Division of Computing Systems Technological Educational Institution of Kavala, GR 65404, Greece Abstract — This work shows comparatively the capacity of five Fuzzy Lattice Neurocomputing (FLN) classifiers. The mechanics of the five classifiers are illustrated geometrically on the plane. Both learning and generalization are based on the computation of hyperboxes in space RN. Learning is memory-based, and polynomial O(n3) where n is the number of the training data. The problem of overfitting is ruled out by construction. In addition, a FLN classifier both induces rules from the training data and it is applicable beyond RN, in particular a FLN classifier is applicable in a mathematical lattice data domain hence disparate types of data can be dealt with in principle. Experimental results in three benchmark classification problems involving data sets of various sizes and various types, i.e. numerical/nominal data, compare favorably with the results by alternative classification methods from the literature. Various theoretical advantages are discussed. I. INTRODUCTION This work demonstrates experimentally the effectiveness of five Fuzzy Lattice Neurocomputing (FLN) classifiers in three benchmark classification problems. The aforementioned FLN classifiers include 1) FLN tightest fits (FLNtf), 2) FLN first fit (FLNff), 3) FLN ordered tightest fit (FLNotf), 4) FLN selective fit (FLNsf), and 5) FLN max tightest fit (FLNmtf). Note that FLNtf has been introduced in [10], FLNff, FLNotf,