Patchwork neuro-fuzzy system with hierarchical domain partition

Abstract

The paper presents the patchwork hierarchical domain partition in the neuro-fuzzy system with parameterized consequences. The hierarchical domain partition has the advantages of grid partition and clustering. It avoids the curse of dimensionality and reduces the occurrence of areas with low membership to all regions. The paper depicts the iterative hybrid procedure of hierarchical split. The splitting procedure estimates the best way of creating of the new region: (1) based on finding and splitting the region with the highest contribution to the error of the system or (2) creation of patch region for the highest error area. The paper presents the results of experiments on real life and synthetic datasets. This approach can produce neuro-fuzzy inference systems with better generalisation ability and subsequently lower error rate.