Normal-Wishart Clustering for Novelty Detection
Christian Gruhl, Jörn Schmeißing, Sven Tomforde, Bernhard Sick · 2020
Novelty detection is an important building block of self-improving system integration since it establishes a kind of self-awareness in the perception. This can be used, for instance, to detect novel processes or subsystems that have an impact on the efficiency of the integration state. CANDIES is a particular instance from the field of novelty detection that has been shown to fulfil the specific requirements of self-integration. In this article, we introduce a novel approach to online clustering of samples based on Bayesian methods that replace the spatial-density clustering approach which is currently used in CANDIES. This novel approach is called Normal-Wishart-Clustering. We analyse the accuracy of the approach in comparison to the standard setup of CANDIES in an example scenario and demonstrate the advantages.