基于改进遗传算法的广度架构搜索算法  

Wide architecture search algorithm based on improved genetic algorithm

作　　者：林东凤 黄汉明[1,2,3] 沈俏 LIN Dong-feng;HUANG Han-ming;SHEN Qiao(School of Computer Science and Engineering/School of Software,Guangxi Normal University,Guilin 541004,China;Guangxi Key Laboratory of Multi-Source Information Mining and Security,Guangxi Normal University,Guilin 541004,China;Key Lab of Education Blockchain and Intelligent Technology of Ministry of Education,Guangxi Normal University,Guilin 541004,China)

机构地区：[1]广西师范大学计算机科学与工程学院/软件学院,广西桂林541004 [2]广西师范大学、广西多源信息挖掘与安全重点实验室,广西桂林541004 [3]广西师范大学教育区块链与智能技术教育部重点实验室,广西桂林541004

出　　处：《计算机工程与设计》2024年第12期3667-3673,共7页Computer Engineering and Design

基　　金：广西重点研发计划基金项目(桂科AB18126045);广西多源信息挖掘与安全重点实验室系统性研究课题基金项目。

摘　　要：为扩大遗传算法产生的子代种群和亲代种群间的差异,提出一种搜索算法,即广度单路径架构搜索算法。该方法将搜索过程分为两个阶段,第一阶段为扩张,使用一种新的交叉算子以及停滞检测算法增大子代种群和亲代种群间的差距,扩大搜索范围;第二阶段为收缩,使用前一阶段获得的若干个体,采用单点交叉做搜索,保证搜索的稳定性,得到最终的结果。在4个数据集上的实验结果表明,该算法搜索出的最优网络与手工设计的神经网络和基于传统遗传算法的神经架构搜索方法相比,能获得有竞争力的结果。To expand the difference between the offspring population and the parental population generated using genetic algorithm,a search algorithm,called wide single-path architecture search algorithm,was proposed.The method divided the search process into two stages.In the first stage,a new crossover operator and the stagnation detection algorithm were used to increase the gap between the offspring population and the parental population,and the search scope was expanded.The second stage was contraction,in which several individuals obtained in the previous stage were used and the single-point crossover was used to search,ensuring the stability of the search and obtaining the final result.Experimental results on four datasets show that the optimal network searched using the proposed algorithm can obtain competitive results compared with hand-designed neural networks and neural architecture search methods based on traditional genetic algorithm.

关 键 词：神经架构搜索 遗传算法 进化计算 均匀训练 卷积神经网络 停滞检测 图像分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]