Dynamic Fuzzy Classifier Design with Point-Prototype Based Clustering Algorithms

Abstract

In this section a dynamic fuzzy clustering algorithm is developed, which provides a possibility to design a dynamic classifier with an adaptive structure. The main property of a dynamic classifier is its ability to recognise temporal changes in the cluster structure caused by new objects and to adapt its structure over time according to the detected changes. The design of a dynamic fuzzy classifier consists of three main components: monitoring procedure, adaptation procedure for the classifier and adaptation procedure for the training data set. The monitoring procedure consists of a number of heuristic algorithms, which allow the recognition of abrupt changes in the cluster structure such as the formation of a new cluster, the merging of several clusters or the splitting of a cluster into several new clusters. The outcome of the monitoring procedure is a new number of clusters and an estimation of the new cluster centres. The adaptation law of the classifier depends on the result of the monitoring process. If abrupt changes were detected, the classifier is re-learned with its initialisation parameters obtained from the monitoring procedure. If only gradual changes were observed by the monitoring procedure, the classifier is incrementally updated with the new objects. The adaptation procedure is controlled by a validity measure, which is used as an indicator of the quality of fuzzy partitioning. This means that an adaptation of the classifier is carried out if this leads to an improvement of the current partitioning. The improvement can be determined by comparing the value of a validity measure for the current partitioning (after re-learning) with its previous value (before relearning). If the validity measure indicates an improvement of the partitioning a new classifier is accepted, otherwise the previous classifier is preserved.