The Performance And Learning Algorithm Of Fuzzy Neural Network

Soft computation technology, whose primary members are fuzzy logic (FL), neuro-computing (NC), evolutionary computing (EC), and probabilistic computing (PC) and so on, is an association of computing methodologies and an effective tool to deal with nonlinear complicated systems. Fuzzy neural network (FNN), which is the organic integration of neural network and fuzzy system, is an important hybrid intelligent system of soft computing technique and an active branch of intelligent control theory. FNN can deal with the abstract information, such as the language information and has good self-learning and self-tuning capabilities. Therefore, the research of FNN is significant in soft computation technology and intelligent control.This thesis systematically studies the performances and learning algorithms of two FNN models, monolithic FNN and polygonal FNN, based on the past progress of FNN theory and application. The major issues in the thesis are the perturbation of monolithic FNN, the learning algorithms and universal approximation of polygonal FNN and the achievements obtained here are applied to fuzzy control area. The research in the thesis provides the applications of FNN and soft computing technique with the necessary theoretic basis.The main contributions of the thesis can be enumerated as follows:1. The perturbation of training pattern pairs on a fuzzy neural network is researched. The definition is established for the robustness of a general FNN to perturbation of training pattern pairs. As a typical instance, this kind of robustness of monolithic fuzzy neural network (MFNN) is analyzed, and the theoretical studies in this paper show that the MFNN has good robustness when training pattern pairs come into the y maximum keep-order perturbations with the coefficient h= 5.2. The universal approximation capability of polygonal FNN is deeply studied. Firstly, the universal approximation of two special polygonal FNNs is analyzed, where inputs or weights of the polygonal FNNs are limited to a small class of fuzzy number. Secondly, universal approximation of the general polygonal FNN are deeply analyzed where there is no limit to inputs or weights of the polygonal FNN. The theory research show that the polygonal FNN can be as a universal approximator to the fuzzy continuous function and the equivalent conditions is the fuzzy functions'increase.3. Two fuzzy learning algorithms are proposed for polygonal FNN. Two fuzzy conjugate gradient algorithms based on genetic algorithm (GA) or quantum genetic algorithm (QGA) are designed for the polygonal FNN. In every step of the algorithms, the learning constant is optimized by GA or QGA and the theory study shows the astringency of the algorithms. The simulate experiments in fuzzy control are employed to illustrate the realization of the corresponding learning algorithms.