Radius Parallel Self-Organizing Map (RPSOM)

概要

論文の詳細を見る
Self-Organizing Map (SOM) proposed by Kohonen is a kind of the Neural Networks and using for the competitive learning. It is difficult to specify the learning parameters for SOM, because the neighborhood radius, which is a one of learning parameter, depends on the scale of the problem or map space, and the users of SOM code usually find the proper value through a trial and error process. In this circumstance, the authors propose a SOM that has several neighborhood radii, which is similar to the temperature parallel Simulated Annealing (SA) method. In this method, several neighborhood radii conduct competitive learning in parallel, and a highly evaluated neighborhood radius is dynamically selected. Furthermore, the authors propose the new evaluation function of the map, in order to evaluate the proposed method quantitatively. The results of the map evaluation with the proposed evaluation function were consistent with those of visual evaluation. The results are compared with those of the conventional SOM, to demonstrate the effectiveness of the proposed method.