Improvement of Learning Performance of Self-Organizing Relationship Network by Topology Representing Network with Weighted Input Data

Abstract

In this paper, we proposed Evaluation based Topology Representing Network (E-TRN) is proposed to improve learning accuracy of self-organizing relationship (SOR) network. In case of that a dimension of inputoutput space is high, an approximation of a target system sometimes includes large error, because the topology of the network is limited by the topology of competitive layer, usually 1- or 2-D. TRN can extract a desired topology precisely by using an idea of Competitive Hebbian Learning rule (CHL). By hybridding SOR network and TRN, both of an evaluation based learning and a topology extraction can be realized.