Application of neural networks to the adaptive routing control and traffic estimation of survivable wireless communication networks

Problems of estimating and optimizing the behavior of wireless networks, based on the structure of a general stochastic model of the network's discrete-event dynamics, lead to mathematically correct, yet computationally intractable, backward recursive conditions defining the stochastic filter of network state and the optimal routing controls. The structure of the stochastic model of network dynamics, reflected in these recursive conditions, strongly parallels the recursive structure found in backpropagation (BP) neural networks. This structural resemblance has suggested the use of variations of the BP approach to compute solutions to the recursive mathematical conditions. Since the structure of the network model is not, in general, feedforward, three variations of recurrent BP algorithms are proposed to solve a partitioned version of the defining filter and optimality conditions. Foreknowledge of the random characteristics of a given network model further suggest which BP technique is appropriate to the application.