| Fuzzy Neural Network(FNN) is a kind of network combining artificial neural network and fuzzy logic system, can deal with abstract information, and it is an active branch in the field of intelligent control of theoretical research. The typical fuzzy neural network structure is known as Fuzzy Multi-layer Perceptron's fuzzy neural network structure. This kind of network is based on the structure of fuzzy system to determine the structural equivalent of neural network, that is to say, making each layer of neural network, each node correspond one part of fuzzy system.The number of nodes in FNN's rule layer determines the whole network's size and performance. Containing too many redundant rule nodes will lead to the hugeness of the network structure and the slow response from input to output; increase the computing complexity of network, as well as difficult understanding to the complicated fuzzy rule which the nodes represent, have a significant impact on the actual control. When learning efficiency is not high or it is easy to get into local extremum, changing the structure of the controller can improve system performance; make FNN accomplish the learning of network structure, in order to get the best network structure in the sense of satisfying the system performance requirement. So the optimization to FNN' s structure is necessary. This thesis focuses on this problem to make following improvementsabout FNN.This thesis is divided into a total of five chapters, Chapter I Introduction FNN basic concepts and development of the status quo; And the main study of this thesis.Chapter II list current more common several fuzzy neural network optimization algorithms, including exhaustive, growth, pruning method and evolutionary algorithm.Chapter III for FNN control structure with the choice of general experience, redundant nodes often occupy a large proportion problem, the use of fuzzy growth of neural network algorithm so that the network from a simple small scale has begun to grow, reaching performance targets so far . And to improve the growth law, must step in every training after the error rate of decay and increased Calculation nodes judgement, to avoid because of the continuous growth led to an instant expansion of the network, but also reduce a certain amount of computation.Chapter IV of FNN pruning algorithm based on the correlation of two nodes on a joint nodes, the nodes associated with the merger of the node to adjust the threshold, and then compute node after the merger FNN each of the nodes scattered , The small nodes scattered through a deletion, at the same time to adjust node threshold, and ensure network performance.Chapter V for growth, pruning of their respective advantages and disadvantages. This chapter presents a viable solution, will be pruning method and the use of growth: from the initial small-scale network has begun to grow, reaching performance targets and then through mergers and pruning, to the computing overhead for the global optimum.By an ordinary structure, growth, and the pruning of the joint algorithm proposed in this thesis are the four fuzzy neural network the large number of comparative experiments show that the algorithm of the rules of ordinary fuzzy nodes only 25 percent of neural networks, training and shorten the period of 61.5 % And in the same degree of optimization algorithm joint training time than pure growth, pruning method to shorten the respective 72.9% and 40.6%... |
PDF Full Text Request