基于改进遗传算法的自适应卷积神经网络  

Adaptive Convolutional Neural Network Based on Modified Genetic Algorithm

作　　者：高新成 李强[2] 王莉利[2] 杜功鑫 柯璇[3] GAO Xin-cheng;LI Qiang;WANG Li-li;DU Gong-xin;KE Xuan(Modern Education Technology Center,Northeast Petroleum University,Daqing 163318,China;School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China;School of Earth Science,Northeast Petroleum University,Daqing 163318,China)

机构地区：[1]东北石油大学现代教育技术中心,黑龙江大庆163318 [2]东北石油大学计算机与信息技术学院,黑龙江大庆163318 [3]东北石油大学地球科学学院,黑龙江大庆163318

出　　处：《计算机技术与发展》2022年第10期132-136,142,共6页Computer Technology and Development

基　　金：国家自然科学基金项目(41804133);东北石油大学引导性创新基金(2020YDL-03)。

摘　　要：传统卷积神经网络模型的构建具有过度依赖经验知识、不可预知性、训练难度大等缺点,导致对网络结构和参数的设置需要耗费大量的时间进行调优测试。针对上述问题,提出基于改进遗传算法的自适应卷积神经网络算法。改进遗传算法通过对卷积神经网络进行编码处理,将分类误差和结构复杂度作为适应度函数,针对选择、交叉和变异策略进行改进,在保证遗传算法种群多样性的同时提高收敛速度,避免算法陷入局部最优解。利用改进遗传算法全局寻优的特性,对神经网络体系结构和重要参数进行优化,实现卷积神经网络的自适应构建,以提高神经网络分类准确率。在MNIST、Fashion-MNIST和CIFAR-10数据集上的实验表明,该算法优化后的卷积神经网络在分类精度、参数设置等方面均取得了良好的效果,与其他神经网络相比,改进的遗传算法具有成功优化卷积神经网络的潜力,对不同分类任务的研究具有重要意义。The construction of traditional convolutional neural network models has disadvantages such as excessive reliance on empirical knowledge,unpredictability,and difficulty in training.As a result,it takes a lot of time to optimize and test the network structure and parameter settings.Aiming at the above problems,an adaptive convolutional neural network algorithm based on modified genetic algorithm is proposed.The modified genetic algorithm encodes the convolutional neural network,uses classification error and structural complexity as fitness functions,and modifies selection,crossover,and mutation strategies to ensure the diversity of genetic algorithm populations,to avoid the algorithm falling into a local optimal solution and increase the convergence speed at the same time.According to the characteristics of modified genetic algorithm for global optimization,the neural network architecture and important parameters are optimized,and the adaptive construction of convolutional neural network is realized to improve the accuracy of neural network classification.Experiments on MNIST,Fashion-MNIST and CIFAR-10 datasets show that the convolutional neural network optimized by such algorithm has achieved good results in classification accuracy,parameter settings,etc.Compared with other neural networks,the modified genetic algorithms have the potential to successfully optimize convolutional neural networks,which are of great significance to the study of different classification tasks.

关 键 词：卷积神经网络 改进遗传算法 自适应 结构优化 参数优化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]