An evolutionary method of training topography-preserving maps

In this paper, we introduce an evolutionary training method that can be used either to replace standard self-organizing map (SOM) training or to post-process a trained SOM. The approach is motivated by a desire to improve the way a map preserves relationships in the data beyond the presentation of continuity. There are two stages in the algorithm, prompted by the observation that there is conflict in standard SOM training between the goal of representing the probability distribution of the data and the goal of preserving topology between input and output (a conflict between competition and cooperation). By pursuing these two goals in two separate stages of training, we are able to focus on each goal individually and prevent each from impeding the other. The use of a genetic algorithm in the second stage determines the adjacencies of neurons in the output map grid and allows greater control over the way relationships between output neurons preserve the relationships found in the input data. This is important because it enhances the ability of the map to more accurately represent the structure of the input data. It may prove especially valuable when dealing with high-dimensional data, when one cannot visually inspect the map plotted in the data space to verify the quality of the mapping