CAJ | 학술논문

小波神经网络结合了小波变换和神经网络的优点,具有很强的非线性映射能力和自适应、自学习能力,特别适合于入侵检测系统.但小波神经网络的也有易于陷入局部极小值、收敛速度慢的弱点.对此,本文引入遗传算法来优化产生小波神经网络的初始权值与阈值等,确定一个较好的搜索空间,从而克服小波神经网络易于陷入局部极小值的缺点;同时引入了阻尼牛顿算法,在遗传算法所确定了的搜索空间中对网络进行快速训练,解决传统小波神经网络收敛速度慢的问题,两者构成阻尼牛顿-遗传-小波神经网络.仿真结果表明该方法可行,使神经网络的逼近能力和泛化能力得到了显著提高.
소파신경망락결합료소파변환화신경망락적우점,구유흔강적비선성영사능력화자괄응、자학습능력,특별괄합우입침검측계통.단소파신경망락적야유역우함입국부겁소치、수렴속도만적약점.대차,본문인입유전산법래우화산생소파신경망락적초시권치여역치등,학정일개교호적수색공간,종이극복소파신경망락역우함입국부겁소치적결점;동시인입료조니우돈산법,재유전산법소학정료적수색공간중대망락진행쾌속훈련,해결전통소파신경망락수렴속도만적문제,량자구성조니우돈-유전-소파신경망락.방진결과표명해방법가행,사신경망락적핍근능력화범화능력득도료현저제고.
The wavelet neural network (WNN) combines both advantages of the wavelet transform and the neural network, hence being of strong nonlinear mapping, adaptive and self-learning capabilities, and fairly suitable to the intrusion detection. However, it has some weakness in computing, such as easy convergence to local minimums and a slow convergence rate. To improve WNNs performance first the genetic algorithm (GA) is introduced to optimize WNNs initial weights and thresholds etc. for getting a better solution space to avoid local minimums;then the Levenberg-Marquardt (LM) algorithm is used to speed up the convergence rate, thus leading to an algorithm-hybrid neural network, namely the GALM-WNN. The simulation results show that such a hybrid treatment makes WNNs approximation and generalization capability be significantly enhanced.