Determination of the appropriate node function of NNs by using the cascade-correlation algorithms

How to determine the appropriate or optimal activation function in the neural networks for a specific learning samples remains open. In this paper the cascade-correlation algorithm which is an efficient constructive algorithm is used after implementation of some kinds of clustering algorithms to produce a modular network structure as a surrogate of activation node functions in the radial basis function (RBF) networks. In this way great improvement on the convergence rate of training algorithms and better approximation are achieved. Simulations with the two-spiral data sets proved the above assertion.