A New Fuzzy Neural Network And Its Applications To Fuzzy Control

Fuzzy inference systems are well-suited for representing the imprecise nature of knowledge and reasoning processes(fuzzy if-then rules and fuzzy reasoning) derived from human expertise, but it has no adaptive capability (learning from examples) to take advantage of a desired input-output data set. At the other end, Neural networks represent a totally different paradigm with learning capability that adapts its parameters based on desired input-output pairs, but neither can it accommodate a priori knowledge from human experts, nor can we transform network configurations and connection weights into a meaningful representation to account for structured knowledge. These two modeling approaches differ completely in the way the knowledge is acquired and encoded internally; thus their advantages and disadvantages are complementary. Fuzzy neural network system results from the fusion of neural networks and fuzzy logic. This kind of system brings the learning abilities of neural networks to the fuzzy decision system.The aim of this dissertation is to provide an integrate framework capable of subsuming both neural networks and fuzzy inference systems. We propose an"Average"fuzzy inference model with"Average"operator. We transform the"Average"fuzzy inference system into its equivalent adaptive network architecture, generally called FMLP (Fuzzy Multilayer Perceptrons), thus introducing the learning capability into fuzzy inference system. By employing a gradient descent learning procedure, the proposed architecture can set up reasonable initial membership functions and start the learning process to generate a set of fuzzy if-then rules to approximate a desired data set, as shown in the simulation examples of nonlinear function modeling.The performance of Fuzzy controllers is substantially restricted by the soundness of knowledge acquisition techniques and the availability of domain (human) experts. Due to the adaptive capability of FMLP, we propose a self-learning control strategy which utilizes FMLP as a controller. Simulations show the effectiveness of the proposed control scheme and the self-learning capability.