CAC FOR ATM NETWORKS USING NEURAL NETWORKS

Abstract

The asynchronous transfer mode (ATM) principle has been recommended by the ITU as the transfer mode for future broadband integrated service digital networks (B-ISDN). ATM is considered capable of supporting virtually all communication services expected in the future, by asynchronous multiplexing of fixed sized packets called "cells". High speed transmission of short cells makes the cell transmission delay between terminals short enough to support voice and video services. However, traffic control is needed to maintain the quality of service (QoS) of network connections. Traffic control is divided into reactive and preventive traffic control. . In a broadband network, preventive traffic control is believed to be most important. Call admission control is the preventive traffic control procedure which decides whether new connection request should be accepted or rejected based upon availability of the capacity required to support its quality-of-service (QoS). At connection setup, a route through the network is selected. Then, the QoS of each affected link is estimated, taking into account the effect of the new connection. The connection request is accepted if each link can offer sufficient QoS to all connections.

The admission control problem is the selective admission of a set of calls from a number of heterogeneous call classes, having widely different characteristics. It is not known in advance which combination of calls can be simultaneously accepted so as to ensure satisfactory performance. In this thesis we describe a new statistical preprocessing technique for a supervised neural network for quality of service estimation. The preprocessing yields a good description of aggregate link traffic. The link statistics are computationally easy to obtain and comply with ATM real time requirements. The neural network target QoS is derived through simulating an ATM network to allow high utilization of the network resources. Experimental results verify the feature detecting ability of the link statistics. The proposed neural network based call admission controller can efficiently estimate the cell loss rate of the aggregate traffic for the heterogeneous traffic classes with priority constraints.