改进差分进化在前馈神经网络训练中的应用  

Modified differential evolution in the application of feedforward neural network training

作　　者：沈洪远[1] 李志高[1] 王俊年[1] 吕铭晟[1] 

机构地区：[1]湖南科技大学信息与电气工程学院,湖南湘潭411201

出　　处：《湖南科技大学学报（自然科学版）》2013年第1期70-74,共5页Journal of Hunan University of Science And Technology：Natural Science Edition

基　　金：国家自然科学基金项目(60974048);湖南科技大学博士启动基金项目(E51066)

摘　　要：分析了误差反传算法,将改进差分进化算法和神经网络结合,采用常数变异交叉与自适应变异交叉的混合策略对网络的权值和阈值进行训练.并用2个典型问题对该网络和误差反传网络进行仿真对比.结果表明:经改进差分进化算法训练的神经网络,收敛速度快、泛化性能好.Based on the error back propagation algorithm, the Modified Differential Evolution （MDE） with neural network ,the mixed strategies of constant variation crossover and adaptive crossover mutation was combined were adopted to practice the weights and threshold in the network, and the two typical problems were issued by the simulation contrast of this network and the error back propagation network. The results demonstrate that the neural network with modified differential evolution algorithm training has fast convergence speed and good generalization performance.

关 键 词：差分进化 神经网络 反向传播算法 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]