Generalizing a generic elliptic RBF learning by bootstrap

The problems of reducing the RBF learning time and the generalization of an RBF network are simultaneously considered. Radial basis function (RBF) based on Gaussian distribution function has been shown as an efficient classification function. The function can be used to cover a set data vectors of various sizes and arbitrary eigen-directions by adjusting its center, variance, and mean accordingly to the data vectors. This flexibility is compensated by a high computational cost of covariance matrix generation and inversion. To reduce this computational complexity, we propose a generic elliptic radial basis function which can arbitrarily change it shape to fit the data vectors by adjusting its center and size. The generalization of each GERBF neuron is achieved by applying the Bootstrap estimation to relocate the center and to widen the size of the GERBF. The proposed technique is successfully experimented with some 2-dimensional benchmarks.