多策略改进MOA的卷积神经网络超参数优化算法  

Hyperparameter optimization of convolutional neural network via multi-strategy improved MOA

作　　者：孙林 刘天诺 徐文静 SUN Lin;LIU Tiannuo;XU Wenjing(College of Artificial Intelligence,Tianjin University of Science&Technology,Tianjin 300457,China)

机构地区：[1]天津科技大学人工智能学院,天津300457

出　　处：《闽南师范大学学报（自然科学版）》2024年第3期42-56,共15页Journal of Minnan Normal University：Natural Science

基　　金：国家自然科学基金资助项目(62076089)。

摘　　要：针对卷积神经网络(convolutional neural network,CNN)在各类任务中性能的优劣受超参数的影响,以及传统蜉蝣优化算法(mayfly optimization algorithm,MOA)的求解精度不高等问题,提出多策略改进MOA的CNN超参数优化算法.首先,利用佳点集理论和Tent映射策略对雄雌种群分别初始化,使种群分布更加均匀多样.然后,研究非线性动态调速策略,对蜉蝣速度更新中重力系数和正吸引系数进行非线性改进;同时引入对数型衰减权重策略,对蜉蝣位置进行更新,构建多策略改进MOA,满足MOA搜索能力、收敛速度和精度的要求.最后,将多策略改进MOA应用于CNN,设计多策略改进MOA的CNN超参数优化算法,旨在提升CNN整体分类性能.在7个基准函数上对改进的MOA进行实验验证改进算法的有效性,在MNIST数据集和Fashion-MNIST数据集上进行图像分类实验.实验结果表明,提出的算法能够有效提升CNN模型的召回率和准确率,具有较强的泛化能力.To address the issues of the impact of hyperparameters on the performance of convolutional neural network(CNN)in various tasks,and to overcome the limited precision of traditional mayfly optimization algorithm(MOA),a hyperparameter optimization method of CNN via multi-strategy improved MOA is proposed.First,population initialization is enhanced by separately initializing the male and female populations using the theory of good point set and the Tent mapping strategy,ensuring a more uniform and diverse distribution of the population.Second,a nonlinear dynamic speed regulation strategy is studied,which non-linearly improves the gravitational coefficient and positive attraction coefficient in the velocity update of the mayfly algorithm weight strategy is used to update the ephemera position,and a multi-strategy improved MOA algorithm is constructed to meet the requirements of the search ability,convergence speed and accuracy of the MOA.Finally,the multistrategy improved MOA is applied into the CNN model to design a hyperparameter optimization algorithm,aiming at enhancing the overall classification performance of the CNN.Experimental results conducted on seven benchmark functions demonstrated the effectiveness of the improved mayfly optimization algorithm.Furthermore,classification experiments are conducted on the MNIST and Fashion-MNIST datasets.The experimental results indicated that the developed algorithm can effectively enhance the recall rate and accuracy of the CNN model,exhibiting strong generalization capability.

关 键 词：蜉蝣优化 佳点集 TENT映射 卷积神经网络 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]