A fuzzy ART-based modular neuro-fuzzy architecture for learning hierarchical clusterings

This paper introduces a modular neuro-fuzzy architecture (FUZZY HART-S) that is capable of incremental learning of stable hierarchical clusterings of arbitrary sequences of analogue and binary input patterns by self-organisation. The architecture is a cascade of fuzzy adaptive resonance theory (ART) network modules, in which each module learns to categorise the difference between the current input to the preceding module and the prototype of its selected category. The developed hierarchy is represented via explicit links between category nodes in neighbouring modules. The experimental results demonstrate the representational capabilities of FUZZY HART-S as well as provide comparisons to two other clustering methods. Due to its simple and fast operation, modularity and hierarchical representation, the FUZZY HART-S architecture lends itself to fast and effective hardware implementation, which makes it particularly suitable for real-world applications.