Improving the performance of self-organizing map

Abstract

Clustering data with self organizing maps is a widely used method. In this thesis, self-organizing map and learning vector quantization methods are considered for solving clustering problems. Firstly, self organizing map is modified to use class information like learning vector quantization method. Also, inspired by the hybrid methods proposed recently which combine reinforcement learning with self organizing map, a novel approach is proposed where a simple reinforcement learning algorithm is used both in self organizing map and self organizing map with class information. Self organizing map and all the proposed approaches are realized in MATLAB and applied to clustering set of two-dimensional points and Iris data set. The discussion on the simulation results are based on the results obtained for training and test sets and ideas are given to improve these results. The main purpose of this thesis is to propose a new learning method and explain how to implement the new method, so the problems choosen to be easily manufactured and widely used for educational purposes. So clustering points on a two-dimensional surface and Iris data set are choosen as benchmark problems and in order to have statistically significant conclusion, the evaluation measures are used as sensitivity, specificity, accuracy, precision and similarity.