Research On Pi-Sigma Neural Networks Learning Algorithms

Pi-Sigma neural network,a kind of higher-order neural network with abilities of fast learning and dealing with nonlinear problems,has been applied to various intelligent information processing fields, such as pattern recognition, signal processing and image processing. As other neural networks, the performance of Pi-Sigma neural network mainly depends on network learning algorithms and network architecture optimization methods which converge faster and learn better.In this thesis, the further research on the hybrid genetic learning algorithm for learning and optimizing of Pi-Sigma neural network is done. The results obtained are as follows:(1) The usual algorithms of learning Pi-Sigma neural network are reviewed. Then the advantages and disadvantages of learning algorithms are summarized.(2) A hybrid genetic learning algorithm is proposed in this thesis. The algorithm mainly incorporates genetic algorithm with the ability of powerful global exploration and simplex method with the ability of powerful local exploration. Then the convergence of the proposed algorithm is analyzed.(3) An architecture optimization method of Pi-Sigma neural network based on Taguchi Method is presented. It optimizes Pi-Sigma neural network architecture through the presented hybrid genetic and enhances the performance of genetic operator, reduces the number of experiments, thereby, reduces the time of algorithm optimization. The proposed method can efficiently prune the redundant weights, reduce complexity training time for Pi-Sigma neural network while it does not increase the training errorFinally, the research work involved in the thesis is summarized and the future developments in optimization of neural network architecture and learning algorithm are forecast.