To search, Click below search items.


All Published Papers Search Service


Recoverable Lost Node to Overcome a Problem of the Single Winner Node by ASOM


Napatsarun Chatchawalanonth, Chom Kimpan, Sanon Chimmanee


Vol. 10  No. 7  pp. 229-236


The Self-Organizing Map (SOM) is a neural network algorithm based on unsupervised learning. The weight initialization is very important for training data. Because the appropriate weight initializations enable a learning convergence to global or local minima to be more correctly. However, traditional weight initialization is randomized in a range of 0 and 1. This results in a problem of a single winner node, which leads to miss of clustering because the initial value is not related to the input datasets. Additionally, the learning convergence is very slow when the initialize value is long distance from the dataset. This paper presents a novel method called as Ant Self Organize Map (ASOM) to alleviate such problems. A proposed original idea is that lost node is recoverable to overcome the problem of a single winner node by exchanging from new datasets. The proposed ASOM is based on the Ant Colony Optimization (ACO) and Fuzzy Ants Clustering as following. Firstly, weight initialization is done from random input data which is selected by a probability of loading. This technique of selecting is applied from Fuzzy Ant Clustering. Secondly, exchange connection weights for lose nodes are made by the methodology of Pheromone of Ant which will keep records of winner and lost nodes. Lost nodes which have pheromone values less than a given threshold are re-initialized by randomizing from new datasets. From the experimental result, it is found that the average of accuracy rates increases up to nearly 16 percent when compared with traditional SOM on standard datasets e.g. iris datasets.


The Self-Organizing Map, Data Clustering