Prediction of MPEG video traffic over ATM networks using dynamic bilinear recurrent neural network

Dynamic bilinear recurrent neural network (D-BLRNN) with a dynamic learning gain and optimization layer by layer procedure is proposed and applied to MPEG video traffic prediction over ATM networks in this paper. Since the BLRNN is based on the bilinear polynomial, it has been more effectively used in modeling highly non-linear systems with time-series characteristics than the conventional multi-layered perceptron type neural network (MLPNN). D-BLRNN is proposed to enhance the prediction capability of the BLRNN further by introducing dynamic learning control and optimization layer by layer procedure. Since ATM networks have a bursty nature because of their variable bit rate (VBR) and MPEG video traffic has a typical VBR characteristics, the proposed D-BLRNN can be an optimal candidate for the MPEG traffic prediction problem. In the experiment, the proposed D-BLRNN is applied to a real world MPEG traffic problem (‘MPEG.data’ from the movie ‘Star Wars’) as an on-line learning and a prediction problem. When compared with other prediction schemes including MLPNN and pipelined recursive neural network, the D-BLRNN improves the prediction error by 60% or more in most cases. The results may imply that the D-BLRNN based predictor proposed in this paper is promising and practically reliable tool for predicting bursty sources including real-time MPEG video traffic.