Initial Free Self-Organizing Map by Aggregation

Abstract

Initialization parameters of Self-organizing map networks (SOM) have a great influence in <br><br>their final result. This is the only known problem of SOM networks. The uncertainty of <br><br>solution is resolved with the fusion methods in many machine learning algorithms. Motivated <br><br>by this application, we aim to reduce the uncertainty in SOM final result and provide a more <br><br>robust solution. This is done by aggregating the results obtained from several run of SOM <br><br>network. The objective in uniting several SOM run is to reduce the effect of two above factors <br><br>and reach a more appropriate solution. To achieve this, a number of methods for integration <br><br>of SOM solutions have been proposed and their characteristics are discussed in detail. <br><br>Finally, these approaches are experimentally compared with some recently proposed SOM <br><br>networks and their effectiveness are investigated.