Adaptive fuzzy clustering of noisy data

Clustering of real and noisy data sets is a problem encountered in many pattern recognition applications. Several existing unsupervised clustering algorithms require the number of clusters in the data set as an input to the algorithm. While classifying unknown data sets, it is not always possible to provide this information as an input Certain ART-type self-organizing networks solve this problem by determining the number of clusters, making use of the similarities and differences within the data set However, in the presence of noise outliers these algorithms are severely biased and could lead to erroneous classification. In this work, an existing adaptive fuzzy leader clustering (AFLC) algorithm has been further developed and modified to make the system more robust and to incorporate a method of identifying and separating noise outliers from the data set. The flexibility of AFLC makes it possible to apply it to various real life pattern recognition and control problems. This algorithm has an ART like neural network architecture integrated with fuzzy learning rules, thereby providing accurate classification even while dealing with imprecise data. It learns on-line in a stable and efficient manner and requires very little a priori information.