Prediction for Non-Gaussian Self-Similar Traffic with Neural Network

The burstiness of self-similar traffic is vital for the network analysis and management. The classical FARIMA processes cannot capture non-Gaussian, namely heavy tailness that is the key factor of the burstiness of self-similar traffic. We present a novel FARIMA predictor with a-stable innovations based on a new non-Gaussian self-similar traffic model. A FARIMA process can be regarded as ARMA process driven by fractional differencing process. ARMA processes with infinite variance can be simulated with recurrent neural network (RNN) instead of conventional least squares methods. To train the weights of RNN, we adopt two methods including the conventional back-propagation algorithm and the hybrid method with genetic algorithm and simulated annealing algorithm. The two weights training approaches can minimize the dispersion. The final predicted values of the self-similar traffic are attained by combining the previous two individual FARIMA predicted values with the different hybrid schemes. Our experimental results for the traffic trace collected from Bellcore Lab and Lawrence Berkeley Lab show that the two FARIMA predictors are efficient, the compound predictors are more accurate