The Study Of Neural Network Structure Optimization Based On PSO

The structure of nerual network is always the difficult and key point to improve the generalize. The traditional method to adjust the structure itself is not adaptive to general conditions. Particle swarm optimization(PSO) is a kind of intelligent optimization algorithms, which is widely concerned at present. The algorithm is simplicity of operator, and has seldom paraneters to set. It has global search ability and faster convergence speed. As the development of the PSO rensent years, the structure of neural network using PSO is feasible and applied prospects.The main content of the thesis includes:(1) In this paper, the background and theory are introduced detailedly. Moreover, several weaknesses and their cause of BP algorithm are analyzed thoroughly. Sevaral improved methods to the structure of feedforward network are summarized;(2) The background, basic theory, maths description of PSO are systematically reviewed. Especially, the cooperative PSO is introduced;(3) The domain of parameter selection to assure the convergence of the PSO algorithm by solving difference equations is determined. In experiment, a wave function is structured to describe the particle trajection embraced the optimal point. So the particle trajections under different initial conditions are compared. The results indicate that the explotation and exploration is effected by the initial conditions;(4) A cooperative PSO with chaos mechanisms is proposed to avoid the particles to be stagnation.The new agorithm is inclued two parts of judgment and dispose. The experimental result manifests that the new agorithm is better than the originate;(5) The cooperative PSO with chaos is applied to the neural network to adjust the structure and perameters of neural network. The Iris classification is used to test the performance of the algorithm. The experimental result manifests that the new agorithm improved the accuracy and generalize.Finally, the works of this dissertation are summarized roundly, and further research directions are indicated.