The Parameter-Less SOM algorithm

One of the biggest problems facing practical applications of Self-Organising Maps (SOM) is their dependence on the learning rate, the size of the neighbourhood function and the decrease of these parameters as training progresses, all of which have to be selected without firm theoretical guidance. This paper introduces a simple modification to the SOM that completely eliminates the learning rate, the decrease of the learning rate and the decrease of the neighbourhood size. This is done by making the learning rate and neighbourhood size dependant on a variable calculated from the internal state of the SOM, rather than on externally applied variables.