Investigation of using variant differential evolutions on optimizing 2-level self-organizing map

Self-organizing map (SOM) is a very powerful tool for automatic detection of relevant clusters. The extended version of SOM, two-level self-organizing map (2LSOM) was introduced for improving SOM clustering in explorative manner. However, structural methods for efficiently confirming the competent optimization of 2LSOM initialization are lacking. Due to the important advantages over other optimization algorithms belonging to differential evolution (DE) approach, this paper investigates the utilization of the original DE as well as the variations, here called VarDE1 and VarDE2 as tools for optimizing the initial cluster weights of 2LSOM. Such investigated approaches are respectively so called DE+2LSOM, VarDE1+2LSOM and VarDE2+2LSOM. With respect to the different choices of mutation process, both variant DEs would get better accuracy than the original one. More elitism on mutation process is involved with VarDE2+2LSOM rather than with VarDE1+2LSOM; whilst the most random mutation is applied by DE+2LSOM. 10-fold cross validation experiments are taken on real-world and artificial data sets with an identified number of clusters. Within the scope of this paper, the investigation results point out the better clustering performance of the variant DEs, VarDE2+2LSOM over the related approaches.