Traffic matrix estimation: A neural network approach with extended input and expectation maximization iteration

Abstract

Accurately estimating of IP Traffic matrix (TM) is still a challenging task and it has wide applications in network management, load-balancing, traffic detecting and so on. In this paper, we propose an accurate method, i.e., the Moore–Penrose inverse based neural network approach for the estimation of IP network traffic matrix with extended input and expectation maximization iteration, which is termed as MNETME for short. Firstly, MNETME adopts the extended input component, i.e., the product of routing matrix׳s Moore–Penrose inverse and the link load vector, as the input to the neural network. Secondly, the EM algorithm is incorporated into its architecture to deal with the output data of the neural network. Therefore, MNETME manifests itself with the advantages that it needs less input data, but has better accuracy of estimation. We theoretically analyze the algorithm and then study its performance using the real data from the Abilene Network. The simulation results show that MNETME leads to a more accurate estimation in contrast to the previous methods, meanwhile it holds better robustness and can well track the traffic fluctuations. We finally extend MNETME to random routing networks by proposing a new model of random routing which overcomes three fatal deficiencies of the existing model and it is easier, more practical and more precise.