Constructing stochastic networks via β-RBF networks

Without considering spatial, stochastic, and temporal features inherent in natural neural systems, the computational power of conventional artificial neural networks (ANNs) is limited. In the present paper, we look at the stochastic complexity and construct a stochastic ANN by modeling stochastic fluctuations in the environmental stimuli such that all stimuli are prone to be corrupted by noise or even outliers and to break networks down; therefore, a positive-breakdown network is required. We investigate the stochasticity in the domain of function approximation (estimation) in the framework of radial basis function networks (RBFNs) and propose a robust RBFN, β-RBFN, by applying the breakdown point approach in robust regression. Experimental results demonstrate the advantages of the proposed networks in robustness and simplicity over the plain RBFNs.