摘要
针对BP神经网络结构由于特征维数增多变得复杂,以及网络易陷入局部极值点,提出了粗糙集和改进遗传算法结合共同优化神经网络的方法。首先利用粗糙集对样本空间进行属性约简,降低特征维数,进而简化BP神经网络的结构;然后训练过程中先用改进的遗传算法全局搜索网络的权值和阀值,再使用BP算法局部搜索细化,避免网络过早收敛。试验分析证明优化后BP神经网络比传统BP网络的预测精度得到了极大提高,泛化能力得到了增强,说明了该方法的可行性、有效性。
Considering that the BP neural network became complex due to the increase of the sample dimension and it fell easily into local maximums or minimums, we combined genetic algorithm and rough set to optimize the BP neural network. Sections 1 through 3 explain our backpropagation algorithm mentioned in the title, which we be- lieve is effective and whose core consists of： （ 1 ） rough set was applied to simplify the network by reducing the at- tribute dimension; （2） modified genetic algorithm was used to globally search the weights and bios and, further, the BP algorithm was to locally optimize them to avoid the network failing into the local extremes. Simulation re- suits, presented in Fig. 1 and Table 2 in subsection 3.4, and their analysis indicated preliminarily that prediction accuracy was increased greatly over that of the traditional BP neural network and that generalization was enhanced, thus showing that our backpropagation algorithm is indeed effective.
出处
《西北工业大学学报》
EI
CAS
CSCD
北大核心
2012年第4期601-606,共6页
Journal of Northwestern Polytechnical University
基金
陕西省科技攻关项目(2011K06-25)资助
关键词
BP神经网络
粗糙集
遗传算法
属性约简
局部极值
权值和阀值
backpropagation algorithms, decision making, efficiency, errors, genetic algorithms, mathematical models, neural networks, optimization, rough set theory
reduction of attribute dimension, simula- tion, weights and bios
