Accurate estimation of large-scale IP traffic matrix

Traffic matrix (TM) estimation, which is an interesting and important research topic at present, is used to conduct network management, traffic detecting, provisioning and so on. However, because of inherent characteristics in the IP network, especially in the large-scale IP network, TM estimation itself is highly under-constrained, and so it is an very ill-posed problem. how fast and accurately to attain large-scale IP TM estimation is a challenge. Based on back-propagation neural network (BPNN), this paper proposes a novel method for large-scale IP TM estimation, called BPNN TM estimation (BPTME). In contrast to previous methods, BPTME can easily avoid the complex mathematical computation so that we can quickly estimate the TM. The model of large-scale IP TM estimation built on top of BPNN, whose outputs can sufficiently represent TM's spatial-temporal correlations, ensures that we can attain an accurate estimation result. Finally, we use the real data from the Abilene Network to validate and evaluate BPTME. Simulation results show that BPTME not only improves remarkably and holds better robustness, but it can also make more accurate estimation of large-scale IP TM and track quickly its dynamics.