Adaptive access control of ATM traffic using neural networks

We report a new adaptive mechanism for the access control of VBR sources in an ATM network environment. The general requirement for ATM in the context of B-ISDN is to accommodate on the same network, the varying bandwidth needs of future video, voice and data services. In establishing a call or connection, the quality of service on the source-to-destination path of already existing connections within the network, has to be evaluated and found not to violate an agreed contract. This evaluation of the end-to-end performance along a complete virtual path (VP), before deciding whether to reject or accept the call is both slow and difficult to achieve, especially for inhomogeneous variable bit rate sources. If the statistically time independent characteristics of the new and existing connections are known, then a more direct and preventive access control can be exerted on individual cell streams at the nodal switching points. An efficient access control has to be adaptive and fast (short delay), in order to cope well with the high bit rate traffic expected in ATM networks. We have trained the probabilistic random access memory neural network to learn the stochastic arrival pattern of cells at an ATM switching node during a given time slot. This knowledge is then applied to achieve adaptive access control of both homogeneous and inhomogeneous sources.