Improvement On Particle Swarm Algorithm And Its Application In BP Neural Network

Particle Swarm Optimization algorithm is a type of novel bionic swarm intelligent optimization algorithm. It has attracted the attention of the majority of scholars owing to its simple principle, few parameters to adjust, fast rate of convergence and easy to be realized. However, so far the theoretical analysis and practical application of the particle swarm algorithm is not yet ripe, there are still a lot of issues which require further study. With regards to the problem that particle swarm algorithm is apt to premature and plunge into local minima, the thesis makes improvements about the standard particle swarm algorithm, and then applies the improved particle swarm algorithm to BP neural network. The main tasks of this paper are as follows:This paper makes an overview of the research and development of particle swarm optimization domestic and overseas firstly. It makes a more systematic analysis of the basic principle of particle swarm optimization algorithm, summarizes some common improved particle swarm optimization algorithms. Then it introduces the analysis, the basic process and the application domain of Hooke-Jeeves pattern search method.The thesis improves the standard particle swarm optimization algorithm because it is apt to be premature and thus plunge into the local minima. The initial population defined originally is divided into the same two sub-populations and each subgroup is divided into two subsets respectively based on the thought of domination by the fitness function, namely, Pareto subset and N_Pareto subset. And then the two better Pareto subsets of the two subgroups are combined into a new population. The new population adopts learning factors and the inertia weights which are different from those of the standard particle swarm algorithm, so such new particles have different flight paths in the evolutionary process and can explore the area as far as possible, resulting in improving the global search capability of the algorithm. In order to balance the global optimization ability and local optimization ability and raise the precision and efficiency of particle swarm algorithm, the paper introduces Hooke-Jeeves search algorithm which has strong global convergence capability into the optimization process of the new population and puts forward the IMPSO algorithm. Through analyzing and comparing the test results of the standard benchmark function between the new algorithm and the standard particle swarm optimization algorithm, the simulating results show that the improved particle swarm algorithm is effective.Ultimately the improved particle swarm algorithm is applied into BP neural network. The paper introduces the principle of neural network and multilayer feedforward neural network based on the BP algorithm, and then uses IMPSO algorithm to train BP neural network and meanwhile gives a flow chart of the new particle swarm algorithm training the BP neural network. The BP neural network trained by the improved particle swarm algorithm is used to predict the hardening layer depth during gear heat treatment and diagnose the fault of the cylinder head and the cylinder wall of the diesel engine. And the paper compares the forecasting results and diagnostic results among the IMPSO-BP neural network, BP neural network, PSO-BP neural network trained by the standard particle swarm optimization algorithm. The contrasted result proves that the BP neural network trained by the improved particle swarm optimization algorithm has better optimization performance and learning capacity.