Congestion control of ATM networks using a learning methodology
S. Jagannathan, A. Tohmaz · 2002
Proposes an adaptive methodology using neural networks (NN) for the available bit rate (ABR) service class in a high-speed network. The transmission rates of the sources are controlled in response to the feedback information from the network nodes to prevent congestion. Specifically, the switch and the buffer dynamics are modeled as a nonlinear discrete-time system and a two-layer neural network controller is designed to prevent congestion. Tuning methods are provided for the NN to estimate the unknown traffic. Mathematical analysis is given to demonstrate the stability of the closed-loop error in the buffer occupancy system so that a desired quality of service (QoS) can be guaranteed. The QoS is defined in terms of cell loss ratio, and transmission or transfer delay (latency). No learning phase is required for the NN and initialization of the network weights is straightforward. However, by adding an initial learning phase, the QoS is shown to improve due to low cell losses during transient conditions. Simulation results are provided to justify the theoretical conclusions. Finally, comparison studies are also included to show the effectiveness of the proposed method over adaptive ARMAX, one layer NN and thresholding techniques during simulated congestion.