GASOM: Genetic Algorithm Assisted Architecture Learning in Self Organizing Maps

Abstract

Self Organizing Map (SOM) is a special kind of neuron architecture that partially simulates the visual cortex of the animal brain and has been proven to be exceptionally successful in data visualization and clustering applications. Generally, these applications start with a predefined and fixed representation architecture of SOM without considering the underlying characteristics of the data in the original input space. In such a scenario, the performance so obtained might not be considered to be optimal. In order to enhance the quality and performance of SOM, we propose to use an evolutionary computation approach, the Genetic Algorithm (GA) to learn the optimal architecture of SOM given any data with adverse characteristics and complexity. The developed package named GASOM has been extensively evaluated with 6 synthetic datasets and 6 real-world datasets. The quality of mapping in terms of error measures have been noted carefully for each evaluation. The recorded quantitative outcomes of GASOM for each dataset demonstrate promising performance with regard to quality of mapping from the input space to the representation space.

A. Saboo, A. Sharma and T. Dash have contributed equally to this work.