A new neural network architecture, called lateral feedback neural network, has been introduced in this dissertation, which introduces intra-layer connections to the hidden layer of backpropagation. The learning algorithms are developed, which adapt all inter-layer and intra-layer connections and bias terms by using Generalized Delta Rule. The benchmark tests show that the lateral feedback network has advantages in the speed of learning, convergence, and stability over the original backpropagation.;The sensitivity analysis has been developed for both backpropagation and lateral feedback networks. The first derivative of an output variable with respect to an input variable has been developed for sensitivity analysis through the network learning algorithms and architectures.;Genetic algorithm is a very efficient and robust search methodology. It is applied in this dissertation to guide the search for optimal combination of input variables for neural networks to reach the goal of small size, fast training, and accurate recall.;The application of the sensitivity analysis to TVA's Sequoyah nuclear power plant's thermal performance study shows that the methodology can be used to identify important variables for the plant system to help the plant personnel control the heat rate deviation. The sensitivity analysis has also been applied to diagnostics of accidents simulated by TVA's Watts Bar Nuclear Power Plant training simulator. The information obtained from sensitivity analysis guides the selection of input variables for small modular networks to improve the performance of the neural network diagnostic systems. Genetic algorithms have also been applied to the same application to search for optimal combination of input variables for the modular networks.
展开▼
