Adaptive measurement-based admission control using neural network predictor

In this paper, we use a neural network based system to predict multi-scale high-speed network traffic and apply it to build a measurement-based admission control (MBAC) scheme. An adaptive two-stage predictor is proposed with the first stage containing a multilayer perceptron (MLP) neural network to forecast multi-step-ahead traffic values at a certain timescale. The second stage combines the forecasts to produce the predicted traffic at higher timescales. Four combination schemes, averaging, RLS, MLP, and fuzzy neural network, are investigated. The performance is tested on four different types of traffic data, MPEG and JPEG video, Ethernet and Internet. The results show that the two-stage predictor outperforms the autoregressive (AR) model. The adaptive MBAC scheme uses a neural network controller to take the decision based on the congestion status and the available bandwidth computed from the multi-scale predicted traffic and the user´s peak rate. The proposed scheme achieves higher network utilization compared to the peak rate and Chernoff bound schemes.