Research And Application Of Swarm Intelligence Algorithms In ANNs

These years, with the evolutional compute rising, people gradually combine it with artificial neural network (ANN), training ANNs with many evolutional algorithm. Because of evolutional algorithm having stronger global searching ability and robustness, if they're combined, not only ANN's generalization mapping ability can be unleashed, but also the ANN's convergence ability and learning ability can be improved.Theoretical analyses and algorithm improving on intelligence algorithm are mainly discussed in our work, and the theory and network configuration have been also introduced; And in the last, they are combined to apply to instances to research the performance. The main contents of this dissertation are as follows:(1) Research on fundamental principle of Particle Swarm Optimization Algorithm (PSO) and quantum-behaved particle swarm optimization (QPSO) provides theoretical principle for further learning.(2) Aimed at local convergence in QPSO algorithm when solving multimodal problems, the reason for local convergence lies in the collections of swarm diversity decline and the particles lose the ability of searching in a wide space. New methods are raised that are using boundary variation and chaos mechanism in the process of searing. It has been shown that the improved QPSO algorithms have preferable ability in solving the multimodal problems.(3) After understanding some knowledge of Wavelet Neural Network(WNN),the network configuration of WNN and wavelet function, selecting the Legendre WNN as the modeling network configuration; This is because of Legendre Wavelet has the feature that is having subsection expression and is polynomial, so the Legendre WNN has some merits such as simple structure and fast convergence velocity.(4)Training Legendre WNN with QPSO based boundary variation (QPSOB), QPSO and PSO, and using the trained models to runoff predict and exceptional detection. The experiments show that Legendre WNN's model built by QPSOB, has some merits such as high convergence, fast velocity, and has a certain utility value.The research results show that ,QPSOB algorithm, has better performance to QPSO and PSO in precision and velocity of convergence; Training Legendre WNN with them ,and applying the trained models to instants ,the experiments show the similar results.