Neuro-Fuzzy Processing of Packet Dispersion Traces for Highly Variable Cross-Traffic Estimation Marco A. Alzate 1,2 , Néstor M. Peña 1 , and Miguel A. Labrador 3 1 Universidad de los Andes, Bogotá, Colombia 2 Universidad Distrital, Bogotá, Colombia 3 University of South Florida, Tampa, FL, USA {m-alzate,npena}@uniandes.edu.co, labrador@cse.usf.edu Abstract. Cross-traffic data rate over the tight link of a path can be estimated using different active probing packet dispersion techniques. Many of these techniques send large amounts of probing traffic but use just a tiny fraction of the measurements to estimate the long-run cross-traffic average. In this paper, we are interested in short-term cross-traffic estimation using bandwidth efficient techniques when the cross-traffic exhibits high variability. High variability increases the cross-correlation coefficient between cross-traffic and dispersion measurements on a wide range of utilization factors and over a long range of measurement time scales. This correlation is exploited with an appropriate statistical inference procedure based on a simple heuristically modified neuro-fuzzy estimator that achieves high accuracy, low computational cost, and very low transmission overhead. The design process led to a very simple architecture, ensuring good generalization properties. Simulation experiments show that, if the variability comes from a complex correlation structure, a single estimator can be used over a long range of utilization factors and measurement periods with no additional training. Keywords: Traffic estimation; Packet pair dispersion; Neuro-fuzzy systems. 1 Introduction Several network parameters and traffic conditions can be inferred from packet dispersion measurements, when a sender transmits probing packets of given length at given instants of time, and a receiver collects them taking note of their inter-arrival times [1]. For example, if the tight link is 100% busy between a pair of probing packets, the correlation coefficient between the dispersion measurements and the tight-link cross-traffic will be 1, i.e., the dispersion measurement reveals the average cross-traffic rate over that link [2]. Otherwise, this correlation will be less than one, increasing directly with the utilization factor and inversely with the probing packets inter-departure times. Most available bandwidth estimation techniques send large amounts of probing traffic (overhead) in order to select the tiny fraction of measurements that satisfies the high correlation condition [3][4][5][6]. However, several simulation experiments with different synthetic traces exhibiting different degrees of variability (not shown here) reveal that this correlation can still be high over a wide range of link utilizations and over a long range of measurement time scales if the traffic’s coefficient of variation is high and the traffic exhibits long range dependence. In this work, we consider a computational intelligence approach to estimate the competing traffic rate in the tight link that, instead of ignoring those measurements during which the tight-link becomes idle, it exploits the correlation that still exists between those dispersion measurements and the bursty cross-traffic under high variability conditions.