Multiscale analysis and prediction of network traffic

Traffic prediction plays an important role in network management especially for the current networks that do not comply to the Poisson model. Wavelet transform is an emerging technique that has a significant advantage in analyzing time domain signals. When combined with LMS (Least Mean Square), wavelet based predictor can achieve better performance than time domain predictor for self similar traffic which are revealed as the current network traffic. However, the computational complexity in predicting each wavelet coefficient is high. In this paper, first, the Least Mean Kurtosis (LMK) which uses the negated kurtosis of the error signal as the cost function, is proposed to estimate wavelet coefficients; then by analyzing the wavelet coefficients of two consecutive data sets, a fast WLMK is proposed to reduce the computational complexity. Simulation results show that the fast WLMK not only incurs smaller prediction error but also reduces the computational complexity greatly.