Stability Study of Learning Vector Quantization

In this paper the stability of Learning Vector Quantization is studied. The analysis is based on theoretical considerations and simulations. The Learning Vector Quantization (LVQ) is found to be quite insensitive to disturbances. The localization of the closest reference vector on the map is the most sensitive part LVQ method. Systematic errors that influence on a feature vector, the expected values of stochastic processes, have an effect on classification. The method is not sensitive to error in labelling during the supervised learning phase.