A New Method for Adaptive Selection of Self-Organizing Map Self-Training Endpoint

The paper presents a new method for adaptive selection of Self-Organizing Map (SOM) self-training endpoint. A method is based on the estimation of the newly introduced parameter ∆init and the learning depth parameter κ. In order to propose an optimal range of κ values, the influence of the selected learning depth parameter to the performance of SOM was tested experimentally using input data with uniform distribution. Additionally, four endpoint selection approaches were tested in spectrum sensing application where the SOM based detector was used to detect primary user emissions in 25 MHz wide spectrum band. Three alternative SOM self-training endpoint selection methods were tested on the same topology based SOM. In comparison to SOM self-training endpoint selection algorithm, based on the cluster quality estimation, the proposed method required from 2.6% (for the SOM with small number of neurons) to 44.6% (for the SOM with higher number of neurons) less iterations to reach the endpoint and preserve the similar sensitivity of the spectrum sensor based on SOM.