Adaptive Neural Network Controller for ATM Traffic

摘要

Broadband-Integrated Services Digital Networks (B-ISDN), along with AsynchronousTransfer Mode (ATM), were designed to meet the requirements of modern communication networks to handle multiple users and a wide variety of diverse traffic including voice, data and video. ATM responds to requests for admission to the network by analyzing whether or not the grade of service (GOS) requirement, specified in the admission request, can be guaranteed without violating the GOS guaranteed to traffic already accepted into the network. The GOS is typically a parameter such as cell loss rate (CLR), average delay, or some other measurement associated with network performance. In order to develop a tractable mathematical algorithm for controlling admission, an accurate model of the communication network and traffic in question is necessary. The complex and dynamic nature of these communication networks make them very difficult to model. Even when such a model can be developed, often with unrealistic simplifications or unsupportable assumptions, the associated mathematical algorithm is frequently excessively cumbersome and timely processing of an admission request is lost. An alternative to conventional mathematical algorithms for cases like these is the use of neural networks (NN). NNs can learn complicated functions relating the inputs and outputs of a system without prior knowledge about the system itself. For ATM B-ISDN networks, NNs can learn the function relating input traffic parameters and resulting network performance by training on an appropriate set of traffic parameter inputs and resulting GOS outputs. In this work three neural network admission controller schemes are examined. The Bayes error rate, as bounded by the Parzen window technique, is also introduced as a benchmark for measuring the performance of these admission controllers.