A Fuzzy Reinforcement Learning Approach for Pre-Congestion Notification Based Admission Control.

Abstract

Admission control aims to compensate for the inability of slow-changing network configurations to react rapidly enough to load fluctuations. Even though many admission control approaches exist, most of them suffer from the fact that they are based on some very rigid assumptions about the per-flow and aggregate underlying traffic models, requiring manual reconfiguration of their parameters in a "trial and error" fashion when these original assumptions stop being valid. In this paper we present a fuzzy reinforcement learning admission control approach based on the increasingly popular Pre-Congestion Notification framework that requires no a priori knowledge about traffic flow characteristics, traffic models and flow dynamics. By means of simulations we show that the scheme can perform well under a variety of traffic and load conditions and adapt its behavior accordingly without requiring any overly complicated operations and with no need for manual and frequent reconfigurations. © 2012 IFIP International Federation for Information Processing.

Document Downloads