Robust radial basis function neural networks

Function approximation has been found in many applications, The radial basi s function (RBF) network is one approach which has shown a great promise in this sort of problems because of its faster learning capacity, A tradition al REF network takes Gaussian functions as its basis functions and adopts t he least-squares criterion as the objective function, How ever, it still su ffers from two major problems. First, it is difficult to use Gaussian funct ions to approximate constant values. If a function has nearly constant valu es in some intervals, the RBF network will be found inefficient in approxim ating these values, Second, when the training patterns incur a targe error, the network mill interpolate these training patterns incorrectly, In order to cope with these problems, an RBF network is proposed in this paper whic h is based on sequences of sigmoidal functions and a robust objective funct ion, The former replaces the Gaussian functions as the basis function of th e network so that constant-valued functions can be approximated accurately by an RBF network, while the latter is used to restrain the influence of la rge errors. Compared with traditional RBF networks, the proposed network de monstrates the following advantages: 1) better capability of approximation to underlying functions; 2) faster learning speed; 3) better size of network; 4) high robustness to outliers.