摘要
讨论了多层神经网络算法缺陷,提出了一种基于改进反向传播(Back Propagation,BP)的快速入侵检测算法——IBP算法:在BP算法中的梯度下降算式中,加入一个动量项α[ω(t)-ω(t-1)],改善计算神经元j到神经元i的级联权值;采用学习速率可变的策略;算法训练网络时采用批处理的样本输入方式。改进后的算法选取较大的学习速率η=0.5和η=0.65,并采用3层神经网络的结构,输入、输出样本是16维和15维,各进行100次独立仿真实验,结果证明可加快算法收敛速度,另外,仿真实验还证明:改进后的算法对初始权值的敏感性、网络所表现出的稳定性等都比传统算法性能优越。
This paper discusses the limitation of multi - ply feedback neural network algorithm, presents a fast intrusion detection algorithm, i.e. IBP algorithm, based on the modified BP algorithm： adding a momentum in decreasing gradient formula to calculate the cascade weight value of neuron j to neuron i, adopting alterable learning rate strategy, choosing batch processing sample input while training the neural network. Bigger learning rates η= 0. 5 and η= 0. 65 are selected in the improved algorithm and the structure of three -ply neural network is adopted. The simples of input and output are of fifteen dimension and sixteen dimension. The simulation with computer shows that the modified algorithm is superior to the traditional algorithm in constringency speed, susceptiveness to the initial weight value, stabilization in network.
出处
《空军工程大学学报（自然科学版）》
CSCD
北大核心
2009年第4期53-57,共5页
Journal of Air Force Engineering University(Natural Science Edition)
基金
陕西省自然科学基础研究计划资助项目(DG070302)
空军工程大学电讯工程学院博士启动基金资助项目
关键词
BP算法
收敛速度
入侵检测
BP algorithm
constringency speed
intrusion detection
