Performance of a neural net used as admission controller in ATM systems

The ability of neural networks to control connection admission in asynchronous transfer mode (ATM) networks is investigated. The general problem of connection admission control (CAC) and its formulation as a functional mapping are discussed, leading to applications of learning algorithms to CAC problems. In particular, the use of the class of feedforward neural net with a backpropagation learning rule is considered, where various architecture alternatives are presented. As an example, a simple neural net structure and its use to control connection acceptance is discussed in detail. The neural net performance is compared with other CAC mechanisms like the peak bit rate, the equivalent bandwidth, and the weighted variance methods. Numerical results for both cases, stationary load and nonstationary pulse-form overload patterns, illustrate the capability of neural nets to act as CACs in ATM environments