Research On The Issue Of Feedforward Neural Networks Learning And Its Application To Control Of Table

This dissertation is mainly concerned with three-layer multi-input and one-outputfeedforward neural networks generally accepted in engineering, and the creativeresearch work focuses on the following aspects: Based on the issue of feedforward neural networks learning, the uniformrepresentation for feedforward neural networks output is given, and the mathematicessence of feedforward neural networks is derived in theory. The following issues arethoroughly discussed in this regard: . The mapping performance of input vectors of networks to its desired output inlearning input samples .The redundant and ineffective components of augmented input vectors ofnetworks in learning input samples . The effect of activation function of hidden neurons on learning of networks .The linear correlation and ineffective elements of hidden-layer learningparameters set in learning input samples .The effect of observation noise on dynamic networks learning parametersequation . Expected risk of samples, the contribution of single hidden neuron to the riskand their estimation .The approach to estimating the effect of single hidden neuron on the redundancyof hidden-layer learning parameters The present paper also sets the constraints of mapping of input vectors ofnetworks to its desired output, the redundancy of augmented input vectors, theredundancy of hidden-layer learning parameters set, the efficiency of hidden-layerlearning parameters set and networks structure. The relationship between learningsamples and generalization of networks is further studied and finally the mathematicessence of feedforward neural networks whose activation function of hidden neuronssatisfies mercer condition is analyzed, and the theoretic guiding of networks learningis provided in the present dissertation. Based on the issue of design of feedforward neural networks learning algorithms,a new basic learning rule of feedforward neural networks is proposed, whichexplicitly includes networks parameter learning and implicitly includes networksconstruct learning. Then some offline and online feedforward neural networkslearning algorithms are presented and their computing complexity analyzed. Based on the issue regarding the servo control of the flight simulator of thesynthetical test table for a tactical missile, the present dissertation makes furtheranalysis and points out that the key to the design of the whole servo control system isthe design of the speed servo control subsystem. The approach to feedforward neuralnetworks approximation of optimal speed control vector functions is presented bothin theory and application, followed by the performance analysis of speed feedforwardcompensation controller of feedforward neural networks, as well as the guiding ofdesign of feedback controller. The design of position servo control system under thespeed servo control subsystem with good performance is discussed in the last place,and the approach to design of position feedforward compensation controller andposition feedback controller is presented in theory and application. The reliability and advantage of the above theories and methods on feedforwardneural networks learning are illustrated through a great deal of concretely test, andthat the issue regarding the servo control of the flight simulator of the synthetical testtable for a tactical missile is thus brought to a close by speed feedforwardcompensation controller of feedforward neural networks is also approved by test.