An Analysis of Convergence in Generalized LVQ

Abstract

We have proposed a new formulation of learning vectorquantization (LVQ) called "generalized LVQ" based onminimum classification error (MCE). In this paper, weattempt to clarify the convergence property of referencevectors in our formulation. We discuss the equilibrium in adynamical system for two-class classification, and provethat equilibrium states exist in our formulation, whilethey do not exist in LVQ2.1 or Juang and Katagiri'sformulation (1992) based on MCE.