Varying cooperation in SOM for improved function approximation

This paper presents a data-driven principle for the adaptation of the neighbourhood range in self-organizing maps (SOM). The objective is a reduction of the approximation error in a counterpropagation-like architecture. Therefore the neighbourhood cooperation range of neurons in different regions of the training data space is adapted. In this framework a neuron specific approximation error serves as a criterion. Different examples show the influence of this modification onto the training of the map and on the function approximation quality of the output layer. This principle may also be used for other methods such as the interpolated self-organizing map, the local linear maps and as input quantizer for the radial basis function nets