Research And Application Of The Co-operative Modular Neural Network

A modular neural network (MNN) is a learning system combined of several sub-networks in a co-operative or competitive way. It attempts to explore the different behaviors of the sub-learners by using multiple neural networks, and so as to improve the precision and reliability of the entire system. A single neural network restricted by its intrinsic limitations can hardly help to handle the complex problems with huge sizes. However, as a high effective computing approach, MNN can sometimes complete this important task. Various researchers have swarmed into this field since they realize the great potential of the MNN. Many theories and application achievements have continually emerged, which made the MNN a hot topic in both machine learning and neural network fields.In this dissertation, the methods of training and combination of the sub-network have been studied with the help of some regression problems. A new sub-network combining method, namely "cooperatively adaptive combination" of MNN, is proposed by fully considering the relationships and interactions between the sub-networks, which vary in different problems, and utilizing the relevant theories of genetic algorithm combined with the existing learning methods of MNN. In this approach, the set of samples is divided into several sub-sets in a particular similarity measure, and the fitness of each sub -network to different sub-set is then evaluated. Therefore, suitable sub-set can be assigned to the sub-networks for training. During the training process, the structure and parameters of a sub-net may be adjusted according to its training effect. So the individual neural networks generated in this way obtain both the relative accuracy and the adaptability to other sorts of samples. MNN consisting of such sub-networks is improved not only in robustness but also in generalization ability. 8 regression problems are provided to test the proposed method and the results are analyzed comparatively.This dissertation also presents a MNN simulation system called "NeuralCraft", and describes the architecture and the functions of each modules as well as man-machine interface in detail. Based on the integrated developing environment, Visual Studio.NET, the system utilizes abundant libraries and encapsulated displaying controls provided by the .NET framework, and acquires improvements and innovations on the base of the original algorithms and builds up a platform for experiments and researches. As the core module of the system is flexible, reusable and extendable, it provides advantaged supports for future practice of teaching and research.