GENERALIZED REGULARIZATION NETWORKS WITH A PARTICULAR CLASS OF BELL-SHAPED BASIS FUNCTION

A generalized radial basis function network consisting of (1 + cosh x) (-1) as the basis function of the same class as Gaussian functions is investigated in terms of the feasibility of analog-hardware implementa tion. A simple way of hardware-implementing (1 + cosh x)(-1) is propos ed to generate the exact input-output response curve on an analog circ uit constructed with bipolar transistors. To demonstrate that networks consisting of the basis function proposed actually work, the networks are applied to numerical experiments of forecasting chaotic time seri es contaminated with observational random noise. Stochastic gradient d escent is used as learning rule. The networks are capable of learning and making short-term forecasts about the dynamic behavior of the time series with comparable performance to Gaussian radial basis function networks.