Using regression analysis methods for building an optimal model of dependency between the queue size and Hurst exponent when transforming a self-similar input packet flow into a flow with exponential distribution

Using machine learning methods, the model has been obtained for predicting the queue size of an input self-similar packet fow distributed according to the Pareto law when it is transformed into a fow with exponential distribution. Since the amount of losses in general case does not provide any information about the efciency of using bufer space in the process of transforming a self-similar packet fow, a complex quality metric (penalty) was introduced to assess the quality of investigated models. This metric takes into account both packet loss during functional transformations and inefcient use of bufer space of switching nodes. It was shown that the models using isotonic regression and support vectors methods are the best by the considered metric.