Classifying loss rates in broadband networks

Tasks such as admission control in ATM and predicting overload conditions in telephone networks require a function that specifies what conditions will result in loss rates exceeding a threshold, p^*. This paper considers the formal task of deriving such a classification function based on samples at different conditions. When the size of these samples is small relative to 1/p^*, previously proposed methods incorrectly classify conditions that surpass the threshold by orders of magnitude. This paper derives general conditions for consistent and robust classifiers and presents specific methods that meet these conditions. The paper analyzes the methods with respect to asymptotic and finite sample behaviour and the results are confirmed using simulated data