Congestion Level Prediction Of Self-Similar Traffic Using Reinforcement Learning

In this research the congestion prediction in nodal packet switches such as the routers is concerned. We introduce and verify a novel scheme for the prediction based on the reinforcement learning. It is a dynamic scheme and non sensitive to short term independence of the traffic flow. We assume a self similar behavior for the traffic and its long term correlation is used as the prediction basis. In self similar process high congestion are more likely to be followed and low congestion are more likely to be followed. Therefore, in a sense, it is possible to predict future sample values based on past values. Therefore, the predictor should not be too sensitive so as to respond to short term burstiness. There have been several solutions for the problem in which parametric models are used. The measured traffic data adjusts the parameters in a prediction formula. These methods provide a good prediction and needs a high computational power. Reinforcement learning, because of their nonlinear nature and simple method of learning, can provides a power-full tool to model bursty traffic patterns. In this work, we introduce a new application of reinforcement learning in traffic congestion prediction. The proposed method is simulated by Network Simulator 2 and following results are obtained. 1. The congestion level prediction is correct is %85. 2. The method has first order complexity and therefore the minimum processing time is required. 3. The prediction result can be used for bandwidth requirements of LRD traffic or call admission control or any performance management problem.