A new time-variant neural based approach for nonstationary and non-linear system identification

In this paper, the problem of non-stationary and non-linear system modeling is addressed, and an original solution based on time-variant neural networks proposed. The time-variance property is due to the decomposition of the weight parameters into a linear combination of proper time functions, namely basis functions, as already investigated by Grenier for linear models. The neural architecture here addressed is an IIR-buffered MLP, trained through teacher-forced based backpropagation. Experimental results confirmed the effectiveness of the idea, since modeling performances achieved by using these networks are superior to those based on classic (time-invariant) MLP schemes.