Abstract: In general, the performances of Multilayer Feedforward Neural Network(MFNN), e.g. function approximation accuracy and network efficiency, are not often sufficiently satisfactory, when it is used for modeling of complex non-linear chemical system. In this work, a novel learning approach, which can adaptively construct the optimal architecture of MFNN, is developed for modeling Quantitative Composition-Activity Relationship(QCAR) of Chinese Medicine, and it is successfully applied to predicting the bioactivity of Chuanxiong. The approach taken by the methodology described here is to train an initial neural network that is smaller than necessary and then adaptively increase the number of hidden layer′s neural nodes when the approximation accuracy is insufficient. Once a new hidden node has been added to network, weights of original nodes are frozen. The parameters of the new node, new-added weights as well as the corresponding change of old weights, are determined on the basis of current learning results. It is also verified by simulated experiments that the presented approach can satisfyingly determine the optimal architecture of MFNN, and its learning efficiency is high. As a consequence, the approach can be used for modeling of complex non-linear chemical system.

Key words: Neural network, Network architecture, QCAR modeling, Activity prediction, Chinese medicine