Recent studies of high quality, high resolution traffic measurements have revealed that network traffic appears to be statistically self similar. Contrary to the common belief, aggregating self-similar traffic streams can actually intensify rather than diminish burstiness. Thus, traffic prediction plays an important role in network management. In this paper, Least Mean Kurtosis (LMK), which uses the negated kurtosis of the error signal as the cost function, is proposed to predict the self similar traffic. Simulation results show that the prediction performance is improved greatly over the Least Mean Square (LMS) algorithm.

The Quality of Service (QoS) requirements such as delay and cell loss ratio (CLR) are very stringent for video transmission. These constraints are difficult to meet if high network utilization is desired. Dynamic bandwidth allocation in which video traffic prediction can play an important role is thus needed. In this paper, we suggest to predict the variation of I frames instead of the actual size of I frames, and propose an algorithm that can achieve fast convergence and small prediction error, thus imposing QoS and attaining high network utilization. The performance of the scheme is studied using the renegotiated constant bit rate (RCBR) service model. The overall dynamic bandwidth allocation scheme based on our fast convergent algorithm is shown to be promising, and practically feasible for efficient transmission of real time videos.