摘要
利用PSO粒子群算法对神经网络的权值和阈值,隐藏层中神经元的传递函数系数进行优化。针对网络训练效果好,而泛化能力很差的情况,将训练样本均方差和权值的平方和结合作为PSO算法的目标函数。实验表明,该方法比惯性权值PSO-BP算法和基本梯度下降法好,不但稳定性好,而且预测精度高,泛化能力得到明显加强。
This paper employs the PSO algorithm to update the weights,the biases and the transfer function's coefficients of the hidden layer in the neural network.As to the phenomena of good approximation and bad generalization,the MSE of the training set and the MSW of the weights are integrated into the fitness goal.In the experiment,the GPSO-BP algorithm which optimizes the coefficients of the transfer function and has the small weights and thresholds is better than the BP algorithm and the PSO- BP algorithm in terms of the mean correct recognition and the stability.
出处
《计算机工程与应用》
CSCD
北大核心
2009年第29期34-36,67,共4页
Computer Engineering and Applications
基金
国家自然科学基金(No.60674054)~~
关键词
BP网络
PSO粒子群算法
传递函数
逼近
泛化
Back Propagation (BP) neural network
Particle Swarm Optimization (Pso) algorithm
transfer function
approximation
generalization