| The objective of this dissertation is to develop different algorithms for neural network training and investigate the impacts of performances for these algorithms. The resulting networks are then used in the decision surface models in textile manufacturing and management. Five different neural network models, namely the "back propagation neural network" (BPNET), "BPNET with fuzzy control" (BPNET-FC), "BPNET with curved search method" (BPNET-CS), "BPNET-CS with a fuzzy controller" (BPNET-CSFC) and, "BPNET-CS with the fuzzy neuron controller" (BPNET-CSFNC) are investigated. First, the traditional back propagation neural network with delta learning is introduced. Later, to improve the speed of network training, a fuzzy controller for the learning rate is applied in the back propagation neural network training. Then, a new learning algorithm using a curved search method, which incorporates second-order information, is investigated. A fuzzy controller for choosing step size in the curve search algorithm is added to replace the commonly used line search method. Results obtained using the curved search method and the fuzzy controller indicate a great potential for saving computational time in the network training. Finally a fuzzy neuron controller is incorporated in the BPNET-CS to simplify the design process for the fuzzy controller. Three small-scale and two larger textile real-life examples are illustrated and discussed. |
