Engineering reliable neural network systems

This thesis presents a study of neural network representation and behaviour. The study places neural networks in the context of designing reliable systems. Several new results on network size and topology are presented. Knowledge based training of neural networks is examined. This is essential for designing reliable neural systems in which the subsymbolic reasoning processes are well defined. Sandwich nodes are introduced and studied as atomic knowledge elements in neural networks. Two new network architectures are introduced, the Loughborough Net and the Loughborough Control Net. These make use of the parallelism inherent in sandwich node representations. The interpretation of neural network representations as logical transformations and rule systems are presented. An equivalence of the rule systems and neural network representation is proposed and discussed. This equivalence is required in order that the total behaviour of the neural network can be understood. A new methodology for designing reliable neural network systems making use of knowledge based training is proposed. This is used to present a general design methodology for the construction of. reliable neural network control systems using the Loughborough Control Net architecture. A case study is discussed where the methodology was applied to the design of an adhesive dispensing controller