An adaptive-topology neural architecture and algorithm for nonlinear system identification

Neural-based nonlinear system identification and control suffer from the problem of slow convergence. The selection of a suitable architecture for a problem is made through trial and error. There is a need for an algorithm that would provide an efficient solution to these problems. One possible solution is presented. The network is built slowly in a step-by-step fashion. This evolving architecture methodology allocates a certain number of nodes that avoid training on outliers and, at the same time, provide sufficient complexity for the approximation of a data set. Through simulation examples it is shown that this algorithm also exhibits faster convergence properties than the usual multi-layered neural network algorithms