A neural network control for effective admission control in ATM networks

We propose and analyze a new call admission controller for ATM networks using neural networks (NN). The proposed model is based upon real time measurements of the traffic via a simple parameter, which is the number of cells arriving during the measurement interval. The length of the measurement interval and the number of traffic samples within, are selected to capture the variability properties of the traffic. A neural network controller is then trained to learn the long term correlation properties of the traffic which is essential for effective statistical multiplexing and bandwidth allocation. A large set of training data representing multiservice traffic patterns with multiple QOS requirements is used to ensure that the controller can generalize and produce accurate results when confronted with new test data. The reported results prove that the neural network approach is effective in estimating the bandwidth requirements, when compared to other traditional methods that are based upon an algorithmic approach. This is, primarily, due to the unique learning and adaptive capabilities of neural networks enable them to approximate any non-linear function from previous experience. Evidently, such unique capabilities are the reasons for proposing the use of neural networks to solve many of the problems encountered in the design of ATM networks