Rival penalized Self-Organizing Map

Kohonen's Self-Organizing Map (SOM) is one of the most commonly used competitive learning algorithms that provide a topological mapping from the input space to the output space. In the conventional SOM, it needs to choose an appropriate learning rate as well as a monotonically decreasing function that lowers the learning rate with time to ensure the convergence of the map. Otherwise, its performance may seriously deteriorate. In this paper, we therefore propose a novel Rival Penalized Self-Organizing Map (RPSOM) learning algorithm, which dynamically penalizes a set of rivals towards driving far away from the input data set during the learning. Compared to the existing methods, this new one need not select the monotonically decreasing function of the learning rate, but still gives a robust result. The experiments have shown its outstanding performance in comparison with the existing algorithms.