Mike Dixon

Neural networks have shown promise as new computation tools for solving constrained optimization problems. This thesis examines the application of neural networks to solve the routing problem in communication networks. Communications applications require efficient and robust algorithms to reduce delay and avoid congestion. The algorithms used to determine these routes are usually iterative and often require the use of a shortest path algorithm to find a number of minimum cost routes through the network at each iteration. Due to the possibly large number of shortest path computations that it may be necessary to perform, one should select a shortest path algorithm that is relatively efficient. It is this need for efficiency that has motivated much of the research into shortest path algorithms, which are implemented with neural networks. The parallel nature of neural networks provides for the possibility of significant improvements in processing time with the use of specialized hardware designed to take advantage of this parallel structure. This thesis develops modified Hopfield neural network implementations of the shortest path algorithm. The properties of these neural networks are investigated and their performances are evaluated through simulations. Despite achieving substantial improvements in performance over existing Hopfield neural networks the research in the thesis shows that there are some major shortcomings with the use of neural networks to solve the routing problem in communication networks.