Cell-loss ratio evaluation of ATM call admission control using fast convergent neural network

The Asynchronous Transfer Mode (ATM) provides significant switching capability for the request of fast growing integrated communication services. Meanwhile, the efficiency and quality of ATM communication can be well sustained by deploying traffic policing and monitoring mechanism to keep bandwidth allocation and Quality of Service (QoS) well balanced. Call Admission Control (CAC) is one of these mechanisms and has been applied to both the User-Network Interface (UNI) and the Network-Network Interface (NNI). Due to the requirement of the high speed switching in ATM network, the CAC algorithm should be as simple as possible. However, because most of the CAC schemes make their decision based on the traffic descriptor at the user side, simple scheme will lead to the poor estimation of the user's traffic and degrade the efficiency of the resource utilization. The success of the broadband multimedia communication in the future will largely depend on the degree to which the efficiency can be achieved in the communication system. Cell Loss Ratio, an important QoS parameter and network traffic indicator, can be used in combination with the resource utilization to give an optimal solution for the CAC.; This thesis proposes a novel Cell Loss Ratio estimation scheme for the real time CAC application in ATM network based on Neural Network methodology. A high performance neural network model is employed to improve the speed and accuracy of the CLR estimation. And a fast convergent neural network training algorithm using Kalman Filter is adopted to speed up the training process. Also, a non parametric Cell-Loss Ratio evaluation method using Allan Variance is used as the basis of the estimation. Based on Peak Cell Rate, Sustainable Cell Rate, Maximum Burst Size and Buffer Size of a specific VC (Virtual Circuit), the CLR of the VP (virtual path) can be quickly obtained by using the Neural Network based method, and so can the CAC function. Several performance issues regarding the training process of the Neural Network are discussed. Test results are provided. Then, a conclusion is made at the end of the thesis regarding the overall performance of this new methodology.