Research Of Optimization Algorithm On Artificial Neural NetWork

Optimization problem is an age-old question, but it also has the most practical significance. Optimization problem of BP algorithm and PSO algorithm is mainly discussed in the whole process base on neural network. The characteristics of these two algorithms is analysed in the first part. Then, the BP algorithm is improved on the basis of which the momentum coefficient is added in the adjusting formula of weight. Besides, formulae derivation is specified. Finally, an improved PSO algorithm is introduced, including composition and steps.In the traditional BP algorithm, the connection weights w is only adjusted assume that all the transfer function of the neuron is the same and is not changeable. That is to say the information is stored in the weights w. Although the momentum coefficient is added, the adjustment of the transfer function is ignored. Thus, the traditional BP algorithm has a series of problem such as slow convergence problems. Therefore, an improved BP algorithm is proposed, which adjust connection weights and transfer function at the same time. It improves convergence rate and enhances the ability to store information of neural node, and makes a better nonlinear mapping ability of neural network.In research of PSO algorithm, an improved algorithm is proposed based on the directional evolution operator. The improved algorithm composes directional evolution operator and standard PSO algorithm. The directional evolution operator is constructed by selecting some best particles, then, other particles will move to the evolution operator. In the construction of DEO, migration operation is added which increases the diversity of the searching space. Practice shows that the improved PSO algorithm enhanced the ability to find the optimal solution of standard PSO algorithm.