Theoretic Research And Optimization Of FNN

Incremental training and batch training based on gradient learning are discussed in this thesis from different angles, and two experiments are presented to verify views of this thesis. Then a new combined RBF neural network is developed based on wavelet transforming and fuzzy clustering. It gets better results than traditional methods when it is used to classify phytoplankton.The application background and research state of FNN are summarized in chapter one.Gradient descent algorithm is an efficient method to train FNN, and it can be realized in batch or incremental manner.Some claim that batch manner is better, while others share the opposite belief.In chapter two pros and cons between these two algorithms in accuracy, training time, computation complexity, learning rate and momentum coefficient are discussed, also problems waiting to be solved are found. The gradient decent learning with momentum is introduced, and analysis is described why it can speed up learning speed. Further more, two experiments are presented to verify previous views.As one of FNN, RBF network is also widely used in many fields. In chapter three, the basic theory, primary learning algorithm and an improved learning method of RBF network are presented. Also the approximation and classification abilities of RBF network are analyzed.A new combined RBF network for classification is developed based on wavelet transform and fuzzy clustering,which is an innovation in the application field of RBF network. Also it's the keystone of the thesis. It decomposes the one difficult problem into multi-simple problems at first. Then it uses multiple RBF networks to solve every simple problem, which makes the ultimate objective come true.A spectra of phytoplankton classification system is established based on combined RBF network in chapter four. Simulation results show that combined RBF network is prior to traditional networks. Contrast analysis between generalized RBF network and normal RBF network are also made.Finally, simulation systems for data pretreatment and spectra of phytoplankton classification are designed. The result of each step and the principle of combined RBF network can be clearly seen in the simulation systems.