Research On Applying Particle Swarm Optimization To Neural Network Parameter Optimization

In recent years, researchers have been obtained inspiration from natural phenomena, and put forward many good intelligent algorithms, such as genetic algorithm, ant colony algorithm, simulated annealing algorithm and particle swarm optimization, and so on. Intelligent algorithms have made great progress.Neural network uses BP algorithm to optimize the calculation of connection weights as a learning tool.But BP algorithm searches with the direction of error gradient descent , so it is easy to fall into local minimum and causes the problem of premature convergence. It can achieve the search ergodicity with the help of GA algorithm to optimize the neural network parameters.But it causes the number of iterations to increase, So the convergence speed is slower. Applying the basic PSO algorithm to optimize the neural network parameters can either achieve the ergodicity of searching, but also eliminate the selection, crossover, mutation and other basic operation in GA algorithm. At the same time ,the speed of searching has increased.But the particles of groups update and move to the optimal location in accordance with the same rules. When the particles fall into local minima, the consistency of movement hinders the algorithm out of local minimum.Based on this issue, this paper presents an improved particle swarm algorithm: aggregation– Reset Particle Swarm Optimization (FRPSO). The main idea of FRPSO is to change those particles with the worst adaptive value, when the populations of the original algorithm aggregate. Those particles with the worst adaptive value can be constantly re-initialized to avoid the algorithm from a local minimum.The author has done a lot of research for the improved particle swarm algorithm. The experiments proved that the improved FRPSO can effectively help original algorithm come out of local minima and it has the more improved performance. And apply the improved FRPSO algorithm to the neural-network parameter optimization by comparison with the BP algorithm, GA algorithm and FRPSO algorithm. The results verify the effectiveness, superiority and feasibility of the improved FRPSO algorithm.