摘要
基于多群体协同进化粒子群算法,提出一种用于BP神经网络训练的新型学习算法.将网络中需要调整权值与偏差组成的矢量看成MCPSO算法中粒子,通过粒子间的竞争与合作,完成网络训练过程.将基于MCPSO训练的BP网络分别应用于函数逼近和模式分类问题.结果表明,基于MCPSO的神经网络学习算法在收敛速度和学习效率等方面优于其他方法.
A novel training algorithm based on MCPSO was proposed to train the BP neural networks was proposed. The free parameters including the weights and bias were regarded as the particles in MCPSO, and the networks were trained by competition and collaboration of the individuals in MCPSO. The designed evolutionary networks were applied to function approximation problems and pattern classification problems. The experimental results demonstrate that the MCPSO based training algorithm is superior to other training algorithms in terms of convergence rate and learning efficiency.
出处
《深圳大学学报(理工版)》
EI
CAS
北大核心
2009年第2期147-150,共4页
Journal of Shenzhen University(Science and Engineering)
基金
深港创新园资助项目(SG20081022137A)
关键词
粒子群
多群体
神经网络
函数逼近
模式分类
particle swarm
multi-swarm
neural network
function approximation
pattern classification