ATM call admission control using neural networks

We propose a novel call admission control (CAC) algorithm for ATM networks using neural networks (NNs). The proposed algorithm employs neural networks to calculate the bandwidth required for heterogeneous multimedia traffic with multiple quality of service (QOS) requirements. The NN controller calculates the required bandwidth per call from online measurement of the traffic via its count process, instead of relying on simple parameters such as the peak and average bit rate and burst length. In order to simplify the design and obtain a small reaction time, the controller was realized using a hierarchical structure of a bank of small size, parallel NN units. Each unit is a feed forward backpropagation NN that has been trained to learn the complex nonlinear function that relates different traffic patterns and their required QOS with the corresponding bandwidth. A large set of training data that represents different traffic patterns with different QOS requirement has been used to ensure that the NN can generalize and produce accurate results when confronted with new test data. The reported results prove that the NN approach is extremely effective in achieving more accurate results than other traditional methods that are based upon mathematical or simulation approximations