Study On Adaptive Structure-Optimized Neural Network Control

As an active branch of the intelligent control theory, the Fuzzy Neural Network (FNN) is a combination of the neural network with the fuzzy logic systems. The FNN gets rid of the shortcoming of the pure fuzzy systems in the learning capabilities, and makes the neural network with"black box"attribution transparent and interpretable.The Wavelet Neural Network (WNN) is a novel network model that combines the wavelet analysis with the artificial neural network. Because the WNN has the self-learning ability of neural network and the time-frequency localization of wavelet analysis at the same time, as well as more powerful fault tolerance and approximation abilities. When dealing with the multi-variales, nonlinear and uncertain systems, the WNN are superior to the traditional feed-forward neural network in the convergence speed, fault tolerance and forecasting effects. So the FNN and WNN theories play the important roles in the intelligent control development.The clustering algorithm is introduced into the FNN to extract the systems'characteristics and optimize the input and output spaces, and therefore the initial and cursory fuzzy rules are set up.On the basis of analyzing the disadvantages of the Fuzzy C-means (FCM) clustering algorithm, an improved FCM clustering algorithm is proposed in allusion to the confirmation of the centers number and initialization of cluster centers. According to the results of the improved clustering algorithm, the number of rules and the initial parameters are obtained. So the initial structure of the FNN is determined accordingly.During the learning of FNN, the error back-propagation learning algorithm is used to correct the parameters, and the sensitivity pruning algorithm used to optimize the network structure with the purpose of adjusting the structure and parameters of FNN self-adaptively and obtaining the best fuzzy rules. Finally, with the function approximation as an example, the performance of proposed algorithm is verified. The conclusion is that the new algorithm has advantages in the adaptation, modeling accuracy and other aspects. It can be used effectively to solve the fuzzy modeling and other control problems.In the article, the author constructs an adaptive WNN by combining the characteristics of wavelet function with that of RBF neural network structures. A multi-staged adaptive WNN model is set up and the dimension disaster problem is solved to the certain extent. It is the effective method for the practical application of the WNN.The genetic algorithm and the WNN have their respective advantages. Their union provides an effective solution to the determination of WNN initial parameters. To solve the"precocious"and low-efficient optimization problems of the genetic algorithm, the genetic algorithm is improved, therefore its search efficiency and overall astringency are raised greatly.Considering the weakness of error back-propagation algorithm in the WNN in easily trapping into local minimum point, as well as too high demand to the initial values of parameters, an effective learning path is put forward by combining the global searching ability of genetic algorithm with the good local performance in both time and frequency localization in the wavelet network. The method is used to determine the initial values of network parameters firstly using genetic algorithm, then to transfer to the pure WNN for training, thus its convergence is speeded up greatly. Moreover, an improved conjugate gradient algorithm is used to train the network and to overcome the shortcoming of easily trapping into local minimum points. Finally, the proposed method is simulated and realized in the control system of double inverted pendulum, and its effectiveness is proved.