Decentralized routing, teams and neural networks in communications

A communication network with stochastic input flows is considered. The nodes which route the traffic are required: (i) to react instantaneously to the variations of their incoming flows so as to minimize an aggregate transmission cost, and (ii) to compute or adapt their routing strategies online on the basis of the measured values of the incoming flows and of some local information. Owing to the first requirement, the routing nodes must be considered as the cooperating decision makers of a team organization. The second requirement calls for a computationally distributed algorithm. This fact and the intractability, under general conditions, of team functional optimization problems were the reasons to assign each routing node a multilayer feedforward neural network, which generates the routing variables. For these neural networks the stochastic input flows play the role of training patterns. The weights of the routing neural networks are then adjusted by means of an efficient algorithm based on backpropagation and stochastic approximation