Research Of Dynamic Fuzzy Neural Network Structure Based On Rough Set

Neural network is one of the core content of Artificial Intelligence. With its unique learning and adaptive capacity, making the neural network has been a hot research field of artificial intelligence. From how to build neural network to how to build highly efficient and stable neural network, the development of neural network has made tremendous progress.The proposed dynamic fuzzy neural network overcomes the shortcomings of neural network of the past. Its structure is not predetermined, but dynamic. Before the start of the study, fuzzy rules do not given in advance, but format in the learning progress. It does not require any prior knowledge or have to consider the layout of network construction and the number of network nodes. However, all the rule-nodes in the dynamic fuzzy neural network are not always active. The original rule-nodes could become inactive and have no contribution to the system with adding new rules. They become redundant nodes finally. Researchers proposed a dynamic fuzzy neural network pruning technique in order to optimize the dynamic fuzzy neural network structure and delete redundant nodes.There are many pruning algorithm, for example, Sensitivity Calculation Algorithm, Weight Subtraction, Competitive Learning, Minimum Output, Singular Value Decomposition, Eigenvalue Decomposition and so on. Although these methods have their own advantages on the different aspect, there are some questions still. Sensitivity Calculation Algorithm has to recalculate sensitivity for the entire network after a pruning. Weight Subtraction can only be used for BP learning algorithm. The output structure of Competitive Learning is not optimal. Eigenvalue Decomposition may lead to pathological solutions. Singular Value Decomposition has no strict formula for determining which rule should be retained.From the study we found that rough set theory has powerful features in mining potential knowledge relationships and dealing with uncertainty knowledge. Therefore, in this dissertation we introduce rough set theory for optimizing the structure of dynamic fuzzy neural network and propose a dynamic fuzzy neural network structure optimization method based on rough set.In this dissertation, we apply dynamic fuzzy neural network in Herminte function approximation, Mackey-Glass time series and nonlinear dynamic system identification through Matlab simulation experiments. The experiment results show that the method in this dissertation is improved markedly in three indicators such as number of rules, training times and the test error.