The Study Of RBF Networks And Its Application To Complex Chemical Information Processing

This thesis reviewed the RBF network and its learning algorithms. A series of improvement have been made on the RBF network learning algorithms, also model combining RBF network with Multi-variate data processing has been put forward. Satisfying result have been made by applying those methods to chemical information modeling, pattern classification and the prediction of Quantitative Structure-Activity Relationship (QSAR).The main focus of this thesis is on improving the RBF network learning algorithms as well as establishing new model combing the RBF change method from the RBF network with traditional Multi-variate data processing methods.The main contributions have been made as follows:(1) The learning of the RBF network weights by BP algorithm has been put forward. Also BP algorithm with momentum term can make the learning rate increased without oscillation. A method with the help of the extend Kalman filters to improve the BP algorithm is put forward which can advance the performance and learning rate of the RBF network.(2) The Principal Component Regression (PCR) and Partial Least Square Regression (PLSR) methods are applied to determine the weight of the RBF networks with certain good results.(3) Based on the virtue of the RBF change, the RBF-PLS-Bayes model is put forward which combines RBF change and PLSR with the Bayes Discrimination. By applying the model to chemical pattern classification, we got a satisfying result.(4) By applying the Genetic Algorithm (GA) to the RBGF network, the training method of RBF network weights can be conducted.(5) The RBF-PLSR method based on genetic clustering algorithm is put forward. Applying the model to the prediction of Amine toxicity QSAR, we got a excellent result.Finally, a summary of the thesis was given some conclusions were ma獼e through the research. Further research interests in the future were prospected.