Learning Mechanism Resrearch In Control And Identification Problems

Under the base of literature review, this paper studies on-line learning and its application in control, off-line learning and its application in identification.The second chapter, focus on learning control, proposes a switching periodic adaptive control approach for a class of nonlinear systems with periodic parametric uncertainties whose period and bound are not known. In the learning control algorithm, a fully saturated periodic adaptation law is utilized to estimate the unknown parameter vector; a logic switching based algorithm is provided to tune the unknown period and bound of the parameter vector online. By virtue of Lyapunov energy function, asymptotic convergence can be ensured for the tracking error and all the signals in the system is guaranteed bounded. A simulation to a one-link robotic manipulator is carried out to demonstrate the effectiveness of the switching learning control algorithm.The third chapter, focus on learning identification, studies "to be expert in one aspect and good at many" subnet learning algorithm; in MNN training procedure, we give experimental study of main and auxiliary objective are all MSE (mean square error), and we discuss influence of distance measure and membership degree to system identification performance.The fourth chapter mainly aims at reducing relative learning complexity of subnets originate from "be Expert in Multiple aspects and Good at Many"(EMGM) modular neural network (MNN). Firstly, monotone construction of original sub-net training criterion is built to obtain a new criterion, with which a very efficient subnet training algorithm is designed. This algorithm contains no iteration. Secondly, a learning condition which is equivalent to original cost function is find for every subnet, and this condition still satisfies the condition needed for the efficient training algorithm designed before. Ten identification problems have been used to test the effectiveness of the new framework. Both theoretical and experimental result show the new algorithm will reduce relative learning complexity of every subnet. Bias/Variance analysis shows maximum ability of generalization performance improvement of EMGM MNN may exist and the improvement comes from the improvement of Bias estimation. Furthermore, inspired by MNN, we designed super NN which has the ability to remember all training samples.