A nonlinear estimation network with versatile bump shaping units for fast online adaptation

This paper proposes a new model of nonlinear estimation network suitable to online adaptation. The distinctive characteristics of this network is that the hidden unit can generate versatile bump-shaped functions. The hidden unit is modeled by using a radial basis function and a linear activation function. For the training of this network, we combine self-organizing method and linear fitting technique. As a result, the proposed model can achieve fast online adaptation. The mapping capability of the proposed network is also discussed in connection with other techniques such as multilayer perceptron, radial basis function network, and fuzzy systems. To show the effectiveness, the proposed model is applied to the identification of nonlinear dynamics governed by Mackey-Glass equation and the prediction of its time series