Research On The Design Of RBF Network Based On The Rough Sets

The rough set theory, putted forward by Polish scientist Z. Pawlak in1982, is a new valid mathematical tool to deal with imprecise, uncertain and incomplete data. RBF neural network is a new developed feedforward neural network in recent years. Because of its simple structure, swift training and the capabilities of approaching any nonlinear functions precisely, it has become a new tool used in system modeling. But for complicated system, it is difficult to guarantee the integrality and validity of knowledge in the initial training data when setting up the model by neural networks. There are often redundancy and noise in the training data, which make the structured network very big and the computation required may be too heavy. Thus it is difficult to reach our anticipant precision. There is fine complementarity between rough set theory and neural networks, so integrating both of them can offer a powerful way for processing uncertain, incomplete information, which has payed more attention by more and more domestic and foreign scholars.In this dissertation, the design of RBF network based on the rough sets theory is researched mainly. The major innovations in this article are as follows:(1) The learning algorithm to select RBF networks center is researched, we analyse the advantages and disadvantages about the learning algorithm, and finally we approximate the Hermite polynomial better though using the OLS algorithm to train the RBF network;(2) we put forward a kind of reduction algorithm about attribute values based on the rough sets. In this paper, we introduce the algorithm by example, the results show that it can not only get the best decision rules, and can greatly reduce the information system of the storage space, this algorithm can solve all the concerned problems;(3) A design method of RBF neural network construction based on rough sets and orthogonal least squares (OLS) is proposed in this paper. First, the rough sets knowledge expression system is built taking use of the large number of system sample data. The impact of input variables on output variables is analyzed through computing an accuracy measure of input space knowledge on output space knowledge with reducing the input space of the knowledge expression system. System order and the input layer nodes of neural network can be decided by this way. Secondly, the hide layer nodes and the weights of output layer of the RBF network can be obtained using OLS algorithm. Finally, the research of simulations on a nonlinear system are carried out using this method in this paper, the research results show the method is effective and feasible.