Identification of the Internet end-to-end delay dynamics using multi-step neuro-predictors

Accurate end-to-end delay and roundtrip time estimates are crucial for implementation of performance management strategies in heterogeneous networks, such as the Internet. In particular, accurate predictions of these delay variables could be effectively used for improvements in the quality of service (QoS) of real-time flows over best-effort networks, and for implementing delay-based congestion control and bandwidth allocation strategies, in general. In this study an empirical approach is proposed for the identification of the end-to-end delay and round-trip time dynamics for a source-destination pair on the Internet using recurrent neural networks. The predictors are designed for multi-step-ahead prediction accuracy within a finite horizon. Measured values of packet source departure, destination arrival and source acknowledgment times are used to investigate the accuracy of the proposed approach.