A statistical approach to determine the neighborhood function and learning rate in Self-Organizing Maps

Abstract

The problem of choosing the neighborhood function and learning rate in Kohonen's Self-Organizing Map (SOM) is generally application dependent, empirical and difficult to analyze. In this paper, a new approach based on the statistical characteristic of the behavior of the SOM during training is introduced to determine the neighborhood function and the learning rate. Keeping statistics related to training and network adaptation characteristics and figuring out the SOM parameters according to those, allows problem independent solutions and a fast map convergence rate.