Statistical parameter estimation with neural networks of average flows in ATM networks

Summary form only given. The problem of flow control in ATM networks is a difficult one due to the uncertainties about the flow rates of the data streams coming from various sources of traffic at different nodes of the network. The flow control problem consists of multiplexing the ATM traffic and it is a challenging statistical management strategy where different patterns are directed into the communication channels. During the multiplexing stage, some prescribed performance constraints on the communication parameters should be strictly satisfied. The real motivation for considering an unconventional control technique for flow control in the ATM networks is because with the growing demand for the traffic of the integrated services (including voice and data communications), reliable decisions should be taken in real time. We consider the problem of estimating the statistical parameters of an intelligent control scheme for smoothing the ATM networks based on a neural network architecture. The neural network considered for carrying out the task of flow control runs the statistical estimation as an iterated map. The reason for this architectural design is the need to have a robust estimation scheme, where the convergence to an attractor of the neural dynamics can be enforced to hold in a small number of iterations of the statistical mapping. The neural network is symmetrically connected and this choice reflects a kind of tradeoff between the speed of convergence and the stability of the states during the successive states of the dynamical relaxation. A numerical experiment describing the dynamic performances of the neural controller for the ATM flow control is presented