Topology based Fuzzy Clustering for robust ANFIS creation

This paper describes how the clustering topology of an input space data distribution is utilized to properly initialize an adaptive neuro-fuzzy inference system (ANFIS). We used a new unsupervised clustering algorithm called topology based fuzzy clustering (TFC) that performs better than growing neural gas (GNG) in extracting the input-space topology. The topology information in the form of number of nodes, node positions and node connectivity is used for the initialization of the ANFIS. Using two robotic modeling tasks as benchmarks, we demonstrate the improved performance of TFC-derived ANFIS when compared to the subclustering method found in the fuzzy logic toolbox of Matlab.