基于进化深度学习的特征提取算法  

Feature Extraction Algorithm Based on Evolutionary Deep Learning

作　　者：陈珍[1] 夏靖波[1] 柏骏[1] 徐敏[1] 

机构地区：[1]空军工程大学信息与导航学院,西安710077

出　　处：《计算机科学》2015年第11期288-292,共5页Computer Science

基　　金：陕西省自然科学基金项目(2012JZ8005)资助

摘　　要：信息全面与维数灾难的矛盾是大数据时代网络态势感知需要解决的首要难题。特征提取一直是主流的降维方法,但现有算法对高维非线性数据效果不佳;深度学习是一类具有多层非线性映射的学习算法,可以完成复杂函数的逼近,但对隐层相关参数十分敏感。针对上述问题,将进化算法的思想引入深度学习,提出了一种基于进化深度学习的特征提取算法。该算法利用遗传算法及进化策略实现全局搜索及优化的特点,并对深度学习结构及相关参数进行了优化。理论分析及实验结果都证明了该算法的有效性。The contradiction of comprehensive information and dimension curse is the preliminary problem oI network situation awareness in the times of big data. Feature extraction is a mainstream method to dimensionality reduction, but performs not well when solving high-dimension and nonlinear data. Deep learning is a multi-layer and nonlinear algo- rithm which can realize the approximation of complicated function, however, it is sensitive to parameters related to hid- den layer. Based on above analysis, a feature extraction algorithm based on evolutionary deep learning was proposed. The algorithm combines evolutionary algorithm（EA） and deep learning, takes advantage of the characteristics of GA and ES, and optimizes the learning structure and relevant parameters. Theoretical analysis and simulation results both prove the effectiveness of this algorithm.

关 键 词：网络态势感知 特征提取 进化算法 深度学习 波尔兹曼机 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]