This paper presents an adaptive approach to the problem of congestion
control arising at the User-to Network Interface (UNI) of an ATM multi
plexer. We view the ATM multiplexer as a non-linear stochastic system
whose dynamics are ill-defined. Rear-time measurements of the arrival
rate process and the queueing process, are used to identify, and minim
ize congestion episodes. The performance of the system is evaluated us
ing a performance-index function which is a quantitive measure of ''ho
w well'' the system is performing. A three-layers backpropagation neur
al network controller generates a signal that attempts to minimize con
gestion without degrading the quality of the traffic, During periods o
f buffer over-load the control signal, adaptively, modulates the arriv
al process such that its peak-rate is throttled-down. As soon as conge
stion is terminated, the control signal is adjusted such that the codi
ng rates are restored back to their original values. Adaptability is a
chieved by continuously adjusting the weights of the neural network co
ntroller such that the performance of the system, measured by its perf
ormance index function, is maximized over a certain optimization perio
d. The performance index function is defined in terms of two main obje
ctives: (1) to minimize the cell loss rate (CLR), i.e., minimize conge
stion episodes, and (2) to maintain the quality of the video/audio tra
ffic by maintaining its original source coding rate. The neural networ
k learning process can be viewed as a specialized form of reinforcemen
t learning in the sense that the control signal is reinforced if it te
nds to maximize the performance index function. Performance evaluation
results prove that this approach is effective in controlling congesti
on while maintaining the quality of the traffic.