Hierarchical Fuzzy Clustering in Conjunction with Particle Swarm Optimization to Efficiently Design RBF Neural Networks

A new method that combines hierarchical fuzzy clustering and particle swarm optimization is proposed to elaborate on an effective design of radial basis functions neural networks. As a first step, we pre-process the available data using ordinary fuzzy partitions defined on the input-output space and generate an aggregate of fuzzy subspaces that uniformly cover that space. We, then, put hierarchical fuzzy clustering in place to reform the aforementioned subspaces in terms of a weighted version of the fuzzy c-means algorithm. The network's kernel centers are elicited by projecting the centers of the resulting fuzzy clusters in the input space. The widths and the connection weights are estimated via the implementation of the particle swarm optimization. To this end, the novelty of our contribution relies on the way we manipulate the available information in order to investigate the network's input-output relationships in the context of fuzzy clustering considering the particle swarm optimizer as the major parameter estimation platform. Finally, the modelling capabilities and the effectiveness of the proposed network are demonstrated through several experiments using 10-fold cross-validation analysis.