Quality of service prediction using neural networks

Network bandwidth (for instance, on the internet) is in great demand at the present time, and efficient allocation of network resources to maximize throughput without compromising quality of service (QoS) is a major problem. ATM switches are one critical component of this resource allocation scheme. They decide, based on the users present on the system and their required QoS and a new call request and its QoS, whether to add the new call. This requires prediction of the QoS parameters if the new user is added. Since networks operate at such high speeds, call admission decisions must be made quickly, yet the prediction function may be complicated. These conditions are ripe for application of a neural network. QoS requirements are typically specified in term of cell delay and cell loss. Previous attempts to estimate the cell delay have either used training methods which produce a single percentile value or have estimated unconditional distributions. We propose a more natural method which admits greater flexibility in specification of user´s QoS requirements and accounts for conditioned variables. We present a conditional cumulative distribution estimating neural network. We then estimate the cumulative distribution of cell delay conditioned on the number of users in a typical ATM switch model and discuss the merits of our approach