Research On Fuzzy Neural Network And ATM Switching Fabric

Fuzzy systems are good at expressing human heuristic knowledge and processing fuzzy information. But the parameters of fuzzy systems, such as rule sets and membership functions, are difficult to automatically design and regulate. It can realized self-learning and adaptive potentiality of fuzzy systems constructed by neural network (NN) that using learning methods of NN automatically design and regulate the parameters according to input/output data. Incorporating merits of fuzzy systems into NN, widely applied in many areas such as communication, automatic control, and signal processing, etc., fuzzy neural network (FNN) has become an important method and technology in intelligent simulation. Fuzzy inference network (FIN) and fuzzy associative memory network (FAM) are two most important FNN models. It is difficult that designs the learning algorithms of FIN to make it suitable for on-line operation. In this dissertation, we try to analyze and improve the learning algorithms of FIN and achieve some results. The research on feature and learning algorithms of FAM is an key topic. However, the results in research of FAM are almost all for max-min FAM model. So, generalizing max-min FAM becomes very significant. The dissertation makes a systematic analysis of the convergence, fault-tolerance, and storage capacity for max-min and max-product FAM. Some improved algorithms are proposed, and the results are generalized to max-T norm FAM. The research projects in this dissertation are listed as following. ? Fuzzy Inference NetworkFuzzy clustering algorithms have played an important role in structure identification of fuzzy systems. Fuzzy c-means algorithm (FCM) is sensitive to initial conditions and easily trap into local minima. A hybrid clustering algorithm of combining genetic algorithm and FCM is proposed. It is verified that the proposed algorithm is capable of convergence to global optima at probability.Traditionally, the structure and parameter learning of FIN are done sequentially. So a large amount of data should be collected in advance. Moreover, the independent realization of the structure and parameter learning usually each spends a lot of time so that it is unsuitable for on-line operation. So an adaptive neuro-fuzzy inference network...