A Fast Learning Algorithm For WNN And Its Application In Control

WNN is a new type of NN whose structure based on wavelet theory .it has advantage of determining structure of networks and initializing parameters by wavelet theory. Besides, the weight of output layer in WNN is linear function, the LS algorithm may be used to training networks, which is convenient to on-line use in control fields..The developments of wavelet networks and wavelet theory are reviewed. The characters, learning methods and application in control fields are discussed in detail in this paper. From the classify and comparison in NN, we can find the merits of WNN while using in modeling and control fields of nonlinear dynamic system.The discrete orthogonal WNN and continuous parameters WNN are analyzed and summarized in this paper ,as well as summarizing and conclusion learning algorithms ,author introduce SGD method ( BP ) ,Gauss-Newton method ( RPE ) ,VMM method (BFGS) ,GA and their applications in WNN. Besides, the method to optimization of WNN's structure is presented. Based on the analyzing of advantages of all methods, a new algorithms ~ -BFGS+LS method is proposed, which is hybrid parameters learning algorithm. It has the advantages of rapid convergence speed and high approximation capability.Based on simulation and analyzing of PID self-tuning control and model reference adaptive control theories, the identification and control methods for the complicated nonlinear dynamic systems are proposed. Then ,the author put forward MRAC methods in CSTR .In comparison with PI control ,MRAC control use the information of plant and has the excellent effect, overcoming disadvantages of PI control.