A neural network controller using reinforcement learning method for ATM traffic policing

A neural network (NN) controller using the reinforcement learning method for ATM traffic policing is presented. The mechanism is based upon an accurate estimation of the probability density function (pdf) of the traffic via its count process and the design of a NN critic function capable of evaluating the system performance in terms of the pdf violation. The pdf-based policing is made possible only by NNs. This is due to the fact that pdf policing requires complex calculations, in real-time, at very high rates which is not feasible via conventional mathematical approaches. The architecture of the NN policing mechanism is composed of critic function and control function. The critic function uses two inter-connected NNs. The first NN is trained to learn the pdf of "ideal non-violating" traffic, whereas the second NN is trained to capture the "actual" characteristics of the "actual" offered traffic during the progress of the call. The output of both NNs is compared. Consequently, an error signal is generated whenever the pdf of the offered traffic violates its "ideal" case. The error signal is, then, used by the critic function to produce an evaluation signal based upon a certain performance measure. A reinforcement learning method uses the evaluation signal to adjust the weights of a third NN, in the control function that generates a control signal capable of maximizing the system performance. The reported results prove that the present policing mechanism is very effective in detecting (and controlling) all possible kinds of traffic violations.<>