Measurement based traffic prediction using fuzzy logic

We investigate algorithms, based on the theory of fuzzy logic systems, which use on-line traffic measurements to learn packet arrival patterns adaptively, leading to "model-free" traffic prediction. The main predictor and several novel variants are developed and tested on diverse data streams. For autoregressive-type traffic sources, our algorithms are on a par with standard time-series predictors. More significantly, our algorithms provide reliable predictions for highly variable, non-stationary MPEG data, a situation where standard methods are poorly suited.