Adaptive LSTAR Model for Long-Range Variable Bit Rate Video Traffic Prediction

Abstract

Static bandwidth allocation for variable bit rate (VBR) video traffic forfeits the available bandwidth. Prediction of the next frame size is thus useful in dynamic bandwidth allocation. It has been shown that VBR video traces are long-range dependent, which makes one-frame-ahead prediction insufficient for dynamic bandwidth allocation. Several studies have been conducted based on the linear autoregressive (AR) model to address VBR traffic prediction. In this paper, we propose the use of a nonlinear model from the AR family called logistic smooth transition autoregressive (LSTAR) to predict VBR video traffic. Furthermore, we introduce adaptive algorithms, including least mean square (LMS), normalized LMS (NLMS), kernel LMS (KLMS), and normalized KLMS (NKLMS), to obtain the parameters of the LSTAR model used in long-range VBR traffic prediction. In the proposed model, we do not separate traffic of different frame types and use only one predictor, which results in lower computational complexity. The performance of the proposed predictor for different prediction steps was evaluated and compared with recently introduced predictors. The results indicate that the proposed nonlinear LSTAR-based predictor yields better results than the optimum linear AR predictor, i.e., Wiener–Hopf and others.