Constrained-storage variable-branch neural tree for classification

摘要

In this study, the constrained-storage variable-branch neural tree (CSVBNT) is proposed for pattern classification. In the CSVBNT, each internal node is designed as a single-layer neural network (SLNN) that is used to classify the input samples. The genetic algorithm is proposed to search for the proper number of output nodes in the output layer of the SLNN. Furthermore, the growing method is proposed to determine which node has the highest priority to split in the CSVBNT because of storage constraint. The growing method selects a node to split in the CSVBNT according to the classification error rate and computing complexity of the CSVBNT. In the experiments, CSVBNT has lower classification error rate than other NTs when they have the same computing time.