Using neural networks for underwater target ranging

Underwater weapons often have to range their targets from complex and highly variable signals, referred to as signatures. If artificial neural networks are applied to this task, the complexity of the ranging problem demands the use of networks with relatively large internal structures. The application of layered, feedforward networks learning by backpropagation (Rumelhart networks) to this problem is discussed. It is demonstrated that these networks may have improved generalization by using noise added to the training set. A large network successfully learns to distinguish between two subsets of a large set of complex and similar patterns, even though in learning the very much smaller training set of patterns the network has uniquely used all units. This indicates that a network can be capable of improved performance, and so might be considered to have `unused capacity´, when it does not have unused units