Neural network regressions with fuzzy clustering

A hybrid neural network regression models with unsupervised fuzzy clustering is proposed for clustering nonparametric regression models for datasets. In the new formulation, (i) the performance function of the neural network regression models is modified such that the fuzzy clustering weightings can be introduced in these network models; (ii) the errors of these network models are feed-backed into the fuzzy clustering process. This hybrid approach leads to an iterative procedure to formulate neural network regression models with optimal fuzzy membership values for each object such that the overall error of the neural network regression models can e minimized. Our testing results show that this hybrid algorithm NN-FC can handle cases that the K-means and Fuzzy C-means perform poorly. The overall training errors drop down rapidly and converge with only a few iterations. The clustering accuracy in testing period is consistent with these drops of errors and can reach up to about 100% for some problems that the other classical fuzzy clustering algorithms perform poorly with about accuracy of 60% only. Our algorithm can also build regression models, which has the advantage of the NN component, being non-parametric and thus more flexible than the fuzzy c-regression.