Sample selection and training of self-organizing map neural network in multiple models approximation

The self-organizing map (SOM) neural network has been used widely in multiple models approximation (MMA). However, the clustering property of SOM may not be fit for MMA. This paper introduces the idea of active learning into the training of SOM, especially for MMA. The neural network selects actively the training samples according to the approximation error of local models. As a result, the distribution of the neural nodes is changed so that the performance of MMA is improved. The process of this training method and the performance improvement are illustrated by a simulation example.