Recently, there have been considerable researches about the fusion of fuzzy logic and neural networks. The purpose of these researches is to combine the advantage of both. In fuzzy logic, it is easy to represent with fuzzy rules the expert's linguistic knowledge about target system. But it is difficult for human to fine-tune the parameters of fuzzy rules. In neural networks, it is possible to adjust the parameters of neural network automatically with its learning ability. But it is difficult to utilize a priori knowledge and provide high-level interpretation for its nodes and connections.
In this thesis, a new fuzzy neural network model is proposed. The proposed models embody the fuzzy model of which fuzzy rules have fuzzy sets in the antecendent part and consequent part. It also provides the fuzzy inferencing. And it has the multi-layered architecture similar to multi-layer perceptron.
The characteristics of the proposed model are as follows: The parameters of fuzzy model can be adjusted by learning. A priori knowledge can be utilized when constructing a fuzzy neural network. It can embody the fuzzy model of which fuzzy rules have fuzzy sets in their antecedent and consequent part. Fuzzy inference can be done fast by the inherent parallelism of neural networks. The improved fuzzy model can be extracted from the tuned fuzzy neural network model.
To show the applicability of the proposed fuzzy neural network, the proposed model is applied to approximation of several 1-dimensional and 2-dimensional functions. In the experiments, we got the desirable results. The proposed model can make good approximation of target function. In comparison with other researches, fewer rules are needed for the approximation.
