Because of its well capability in expressing the experts' knowledge and simulating the inference process of human being, fuzzy inference system is applied widely. However, people found soon that fuzzy inference system is not good at learning from data. Then the learnable neural network is combined with the fuzzy inference system to build the "Neuro-fuzzy System". There are some disadvantages in the traditional neuro-fuzzy systems, and to cop with those disadvantages, this paper provides a new neuro-fuzzy system. The new system using Mamdani type fuzzy rules avoids the disadvantages of being hard to incorporate experts' knowledge and hard to understand. In the second place, the new system uses discrete membership functions at the consequent parts of the IF-THEN rules. So, any original membership function can be approximated by a discrete membership function and the shape of the membership function can be adjusted in training procedure in turn. The new neuro-fuzzy system includes a fuzzy inference system and its one to one mapping neural network, and this paper also provides the RBF neural network based structural learning algorithm and the gradient-based parameters learning algorithm for the new system. This paper also successfully applied the new neuro-fuzzy system in several important problems of Intelligent Transportation Systems. With the development of control theory, the implementation of control algorithms more and more depends on the powerful computational systems. On the other hand, people are asking for smaller and simpler controller. Then the conflict of complex computation and simple implementation appears. To solve the conflict, "Local Simple Remote Complex" control principle is a pod choice. Two correspondent parts at local and remote fields are integrant parts to implement the "Local Simple Remote Complex" control principle. The local part should be simple and no need to be learnable, but the remote part should have well capability of learning and optimizing parameters used at the local part. This paper utilizes the one to one mapping characteristic of the new neuro-fuzzy system and puts the simple fuzzy inference system at the local field while puts the learnable complex neural network at the remote field. Remote neural network gets the training data from the local field and optimizes the parameters, which in turn, are sent back to the local fuzzy inference to optimize its performance. In the long run, the "Local Simple Remote Complex" control principle is implemented successfully
修改评论