Investigation Of AND-OR Fuzzy Neural Networks With Application To Ship Control

Recently, research effort into the direction of combining the fuzzy logical control methods with neural networks has become a fruitful area in the domain of intelligent control fields. Since advantages of both techniques are not only complementary but also supplementary each other. The result of combination of the two techniques forming one form of networks is called fuzzy neural networks. However, the fuzzy neural networks may be not work well without a completed connecting all neurons, which has become the main part of current research activities, and consequently a great deal of achievement has been obtained. This thesis proposes a new AND-OR fuzzy neural networks based on AND, OR neurons, which are made up of T norm and S norm, whereas the proposition of such a new neural network can't use the former theories without a further improvement. In order to analyze and verify the new neural networks, this thesis has done research into this kind of fuzzy neural networks. The main contributions are as follows.After the analysis of the inner structure of AND, OR neurons, consisting of T norm and S norm, a result has been obtained, that is, AND-OR fuzzy neural networks are equaled to AND gate and OR gate in digital circuits, and it has the ability to reduce the input space. As a new feed forward multilayer neural networks, AND-OR fuzzy neural networks is designed based on some neuron definitions, which are the in-degree, out-degree of neurons, the relation degree and connectivity of layer. Zadeh's operators are employed in order to infer the symbolic expression of every layer.Some important results are also obtained based on studying the AND-OR fuzzy neural networks. First, one equivalent relation is illustrated between the architecture of AND-OR fuzzy neural networks and the complex rule base. Second, the other equivalent relation is proved between the computation process of AND-OR fuzzy neural networks and the fuzzy weighted Mamdani inference process. At last, the local approximate capability of AND-OR FNN is proved based on Weierstrass theorem.