ABSTRACT
This chapter presents the call admission control (CAC) problem and various ways of using neural networks as the key component of a CAC algorithm. The call admission control problem is one of deciding whether or not to admit a particular call into the network. The CAC problem involves the prediction of the Quality of Service that would be received by the new call and the existing calls, if the new call were admitted. Neural networks are attractive for solving CAC problems because they are a class of approximators that are well suited for learning nonlinear functions. Modular networks offer several advantages over a single neural network. A common architecture for all the models described is a 3-layer feedforward neural network, with a layer of inputs, a single hidden layer of neurons, and an output layer. The entropy of traffic streams is attractive as a descriptor because it can capture the behavior of correlations over many time scales.
