The Application Study Of Artifical Neural Networks In The Quantitative Prediction Of Chemical Physical Properties

In this dissertation, Artifical Neural Networks(ANNs) is used to Predict the Chemical Physical Properties(CPP) first time. This means that a kind of technique in one research field is transplanted to another field successfully.Firstly, the developed history of ANNs and its current application situation are reviewed. The idea of the quantitative prediction with ANNs is proposed and the possibility of this idea is analyzed. The features of this technique are summarized.The Error Back Propagation(BP) network is chosen for this work. The activation(or transfer) functions of neurons are discussed. The roles of parameters in activation function, the relationship between the activity of neurons and the derivative of activation function are approached. It is also concerned to select the training patterns and pretreat the signals.According to a lot of experiment and research, the rules of determining network topologic structure for the prediction of CPP are put forward. The thought of network short circuit treatment is raised in order to promote the direct effect of input signals on output signals. This method is feasible by the proof of experiment and helpful to improve the learning effect. The initial weights setting in different cases are tested so that the reasonable range of initial weights value can be given.In this paper, the influence of learning rate η, increment of learning rate Δη and upper limit of learning rate η_max to the training procedure are discussed. The setting rules of these parameters are suggested. The method of increasing learning quality by the attenuation of Δη is first posed. With this method, the better output squared error could be obtained.Another important parametei—momentum factor α is deep studied in this thesis, α is changed from constant to dynamic adjustment in the training procedure. The viewpoint of α peak value is advanced. The training can efficiently carry out with the automatic control system of α peak value. The larger learning strength can be kept and the learning vibration is avoided. The concept of the best learning state is also proposed in study. In addition, the strategies of learning time control, weight disturbance and adjustment of training set scale are developed. With them, the train-...