Neural network methods with traffic descriptor compression for call admission control

We present and evaluate new techniques for call admission control (in ATM networks) based on neural networks. The methods are applicable to very general models that allow heterogeneous traffic sources and finite buffers. A feedforward neural network (NN) is used to predict whether or not accepting a requested new call would result in a feasible aggregate scream, i.e., one that satisfies the QOS requirements. The NN input vector is a traffic descriptor for the aggregate stream that has the following beneficial properties: its dimension is independent of the number of traffic classes; and it is additive, allowing it to be updated efficiently by simply adding the traffic descriptor of the new call. A novel asymmetric error function for the NN helps achieve our asymmetric objective in which rejecting an infeasible stream is more important than accepting a feasible one. We present a NN design that provides an optimal linear compression of the NN inputs to a smaller number of traffic parameters. The special case of one compressed parameter corresponds to an NN version of the equivalent bandwidth. Experiments show our methods to be better than methods based on the equivalent bandwidth, with respect to call blocking probability and the percentage of feasible streams that an correctly classified