Reducing Computational Complexity in k-NN based Adaptive Classifiers

Integrating new information in intelligent measurement systems during their operational life is always profitable from the accuracy point of view but it generally induces an increment in the complexity of the classifier. Adaptive classifiers, which provide adaptive mechanisms to update their knowledge base over time, are able to exploit fresh information to improve accuracy but, traditionally, do not consider complexity issues. In this paper we propose a design solution for adaptive classifiers able to reduce the computational complexity and the memory requirements of k-NN classifiers by including condensing editing techniques. Moreover, we propose a novel approach for estimating the incoming innovation content which allows us for not including redundant or superfluous information (thus minimizing the knowledge base size).