An Hybrid Neural Network Based On Self-Growing Method Study

Neural networks are one of the hotspots of current research on artificial intelligence. As a simulation of biological nervous system, pre-designed structure will often make neural network has some limitations in its functionality. Studying on self-growing structure neural network is not only to explore the origin of intelligence, but also to streamline the structure and to improve the generalization ability. Currently, self-growing structure neural network has made a lot of progress in many areas. However, most of the self-growing structure neural network has problems in high complexity, slow-growing, Producing useless neurons and making bad use of the advantages of self-growth. In this paper, based on previous research, a new self-growing structure neural network including mixing operator is proposed to improve the efficiency through a guided particle swarm algorithm. The content is shown as follows:(1) A new particle swarm optimization is proposed based on guiding particle technique. This technique can avoid decrease the diversity of the particle swarm population, through the quantization and the immune feature selection, and the guiding particle is proposed to guide the particle swarm to escape from local minima. Compared with some similar types of algorithms to solve classic function-optimization problems, the proposed algorithm shows good results;(2) improved the traditional neural networks algorithm accuracy and generalization by guiding PSO;(3) a new self-growing neural network based on hybrid neurons is proposed for high accuracy of nonlinear identification. Hybrid neural network achieved the characteristics of rapid growth, accurate results and less neuron, though the hybrid hidden layer consists by addition part and multiplication part. A variant of quantum particle swarm optimizer called Guiding Quantum Particle Swarm Optimizer incorporating Immune algorithm (GQPSOI) is also introduced to guide the growth of the neural network structure and weights updated. Through the fuel cell modeling and comparative analysis, it proves the Hybrid neural network is effectiveness and has a good application prospect...