Self-organizing Cascade Neural Network Based on Differential Evolution with Better and Nearest Option for System Modeling

System modeling of engineering problems is an important task, and it's very difficult because most engineering problems are of great nonlinearity and input variables selection is difficult. Self-organizing cascade neural network (SCNN) is a new network which inserts the hidden unit into network layer by layer, while current training methods are still of low efficiency. In addition, most neural networks' inputs units should be provided before training and related input units analysis is of great time-cost. In this paper, a new meta-heuristic algorithm, called as differential evolution with better and nearest option (NbDE), is introduced to SCNN training. In NbDE-SCNN, the orthogonal least square method is applied to evaluate the network contribution of candidate hidden unit and input unit, and NbDE is used to find the best hidden units. Four benchmarks, including the Henon chaotic series prediction, a nonlinear dynamic system, a hydraulic system and a nonlinearity impedance control strategy are used to test the performance of NbDE-SCNN. Simulation and experiment results show that the NbDE-SCNN can select proper input units for system modeling and shows better efficiency in system modeling compared with conventional training methods.