LAUSR

In traditional neural trees (NTs), each internal node is designed as a neural network (NN), such as single- or two-layer neural networks, to determine which branch should be followed for an input sample. Because each NN contained in the internal nodes is designed separately, the produced NT does not consider overall effectiveness. Thus, the designed NT is usually not an optimal NT. In this study, the tree-structured multilayer neural network (TSMLNN) is proposed for classification. The TSMLNN is similar to an NT, which is the result of dividing a deep multilayer NN into many small sub-networks. The TSMLNN has the advantages of both a multilayer NN and an NT. In addition, the split method is proposed to determine how to split the network in the TSMLNN. The genetic algorithm is proposed to automatically search for the weights, activation threshold of each node and the proper number of nodes at each layer according to both the computing complexity and classification error rate in the TSMLNN, and the proposed TSMLNN tends to be optimal. A heuristic method is also proposed to help users to decide which TSMLNN is the best within the classification error rate range. Finally, the performance of the proposed TSMLNN is compared with that of state-of-the-art neural networks in experiments.