基于MAPSO算法的小波神经网络训练方法研究

为提高小波神经网络(Wavelet Neural Network,WNN)的建模质量,针对标准粒子群(Particle Swarm Optimization,PSO)算法优化WNN存在的早熟和局部收敛问题,提出一种基于多粒子信息共享(Multi-particle information share)和自适应惯性权重(Adaptive inertia weight)策略的PSO方法(MAPSO)用于WNN训练。多粒子信息共享采用多粒子信息来修正各粒子下一次的行动策略,以降低粒子陷入局部最优的可能性;惯性权重自适应调整根据群体早熟收敛程度,按个体适应度自适应调整惯性权重,以使陷入局部最优粒子跳出。同时,给出了算法实现的基本流程。仿真结果表明MAPSO算法既具有PSO算法的简捷性,又能够提高WNN学习速度和精度及全局搜索能力,是小波网络的有效训练方法。

Research on WNN Training Algorithm Based on MAPSO Algorithm

In order to improve the quality of WNN modeling,aiming at disadvantages of SPSO algorithm such as prematurity convergence and local minima in WNN training,MAPSO algorithm was proposed based on multi-particle information share and self-adaptive inertia weight adjustment.The multi-particle information share adopted multi-particle information to modify the action strategy of each particle at next time for reducing the probability of local minima.The inertia weight adjustment used individual fitness to adjust the inertia weight according to prematurity convergence degree of swarm for the particle under the local minima dapping.The processes of training WNN by this algorithm were presented.The simulation results show it is an effective algorithm,which not only is simple and efficient like PSO algorithm but also can increase the learning speed and precision,especially global search ability.