This paper presents a distributed learning method using the random neural network (RNN) as a baseline for call admission control and more generally service quality control in broadband ATM networks. An efficient controller not only requires a precise knowledge of source characteristics to achieve admission control but it should also anticipate new connection impact on the network state to ensure optimal performance. Therefore, we first address the problem of modelling each queueing system in the network with three random neurons representing respectively the average number of waiting cells, the queue utilization factor and the traffic burstiness, and show how the connection scheme and the weight initialization benefit from the stationary queueing theory results. Then, we model the overall network with several sub-RNNs located on the switching facilities and virtually connected via the real communication links. The method's efficiency is illustrated by an experiment in which the neural networks achieve call admission control on the basis of user requested service quality. This simple controller is shown to maximize the system throughput while maintaining quality of service.<<ETX>>
[1]
Erol Gelenbe,et al.
Random Neural Networks with Negative and Positive Signals and Product Form Solution
,
1989,
Neural Computation.
[2]
Wolfgang E. Denzel,et al.
Performance of packet switches with input and output queueing
,
1990,
IEEE International Conference on Communications, Including Supercomm Technical Sessions.
[3]
Masayuki Murata,et al.
Performance analysis of nonblocking packet switch with input and output buffers
,
1992,
IEEE Trans. Commun..
[4]
David M. Lucantoni,et al.
A Markov Modulated Characterization of Packetized Voice and Data Traffic and Related Statistical Multiplexer Performance
,
1986,
IEEE J. Sel. Areas Commun..
[5]
Erol Gelenbe,et al.
Learning in the Recurrent Random Neural Network
,
1992,
Neural Computation.