Neural-network connection-admission control for ATM networks

ATM connection-admission control (CAC) using neural networks offers improvement over conventional CAC but creates some difficulties in real operations, such as complicated training processes. This is because ATM traffic characteristics are quite diverse, and quality of service (QoS) and bandwidth requirements vary considerably. A neural-network connection-admission control (NNCAC) method which can overcome these difficulties by preprocessing neural-network input parameters is proposed. The NNCAC method introduces a unified metric for input-traffic parameters by utilising robust analytical results of the equivalent-capacity method. It diminishes the estimation error of the equivalent-capacity method, due to modelling, approximation and unpredictable statistical fluctuations of the system, by employing the learning capability of a neural network. The method further considers the congestion status parameter and the cell loss probability, which provides insight information about the system. Simulation results revealed that the proposed NNCAC method provided a 20% system-utilisation improvement over Hiramatsu's (1990) neural-network CAC scheme and a 10%, system-utilisation improvement over the fuzzy-logic-based CAC scheme, while maintaining QoS contracts. It was also found that the NNCAC method provided utilisation comparable with that of the NFCAC scheme but possessed a lower cell loss probability. NNCAC is suitable for designers who are not familiar with fuzzy-logic control schemes or have no ideas about the requisite knowledge of CAC.