Adaptable neural networks for modeling recursive non-linear systems

In this paper, an efficient algorithm for recursive estimation of a Non-linear Autoregression (NAR) model is proposed. In particular, the model parameters are dynamically adapted through time so that a) the model response, after the parameter updating, satisfies the current conditions and b) a minimal modification of the model parameters is accomplished. The first condition is expressed by applying a first-order Taylor series to the non-linear function, which models the NAR system. The second condition implies the solution to be as much as close to the previous model state. The proposed recursive scheme is evaluated for the traffic prediction of real-life MPEG coded video sources.