April 2000
Abstract:This paper presents and evaluates a gain scheduling approach to solving the admission control and routing problems for self-similar call arrival processes. The control problem is decomposed into two sub-problems: prediction of near-future call arrival rates and computation of control policies for Poisson arrival processes. At decision time, the predicted arrival rates are used to select one of the control policies. The rate predictions are made by neural networks, trained on-line, and the control policies are computed using standard techniques for Markov decision processes. In simulations, this method achieves higher link utilization than methods which do not exploit the memory of the arrival process. It also adapts to the network traffic considerably faster than a previously presented controller employing reinforcement learning without decomposition of the problem.
Available as compressed Postscript (667 kB, no cover)
Download BibTeX entry.