基于差分进化算法的神经网络优化设计  

Optimal Neural Network Design Based on the Differential Evolution Algorithm

作　　者：邹东池 白富生[2] ZOU Dongchi;BAI Fusheng(School of Mathematical Sciences,Chongqing Normal University;National Center for Applied Mathematics,Chongqing Normal University,Chongqing 401331,China)

机构地区：[1]重庆师范大学数学科学学院 [2]重庆师范大学重庆国家应用数学中心,重庆401331

出　　处：《重庆师范大学学报（自然科学版）》2022年第1期79-89,共11页Journal of Chongqing Normal University:Natural Science

基　　金：国家自然科学基金(No.11871128);重庆自然科学基金(No.cstc2019jcyj-msxmX0368,No.cstc2021jscx-jbgsX0001)。

摘　　要：【目的】研究解决传统神经网络手动设计网络结构的局限性,并探究差分进化算法对神经网络优化的有效性。【方法】提出了一种基于差分进化算法的多层前馈神经网络的优化设计方案,用以同时完成神经网络的权值空间和网络结构空间的搜索,给出不同场景下的最优网络结构。该算法采用(1+1)-ES二元进化策略,使用一种新的网络结构交叉和变异方法,通过双种群结构共同进化及自适应变异率等策略加快网络结构的搜索以及算法的收敛。【结果】在预测、分类等问题中,基于差分进化算法的神经网络优化设计能够较好地搜索到最优的神经网络结构,并与传统的BP神经网络以及经典的预测分类算法进行比较,实验结果具有较强的鲁棒性。【结论】基于差分进化算法的神经网络优化设计是解决网络结构寻优问题的有效方法。[Purposes]In order to solve the limitation of traditional neural network manual design network structure and explore the effectiveness of differential evolution algorithm for neural network optimization.[Methods]It proposes an optimization design scheme of multi-layer feedforward neural network based on differential evolution algorithm,which is used to complete the search of the weight space and network structure space of multiple neural networks at the same time.The algorithm adopts the(1+1)-ES evolution strategy,and utilizes a new network structure crossover and mutation method,which accelerates the search of the network structure and the convergence of the algorithm through techniques such as co-evolution of two-population structure and adaptive mutation rate.[Findings]In the prediction,classification and other problems,the neural network optimization design based on differential evolution algorithm can better search the optimal neural network structure,and has higher accuracy and stronger robustness than BP neural network and other classical machine learning algorithms.[Conclusions]The proposed optimal neural network design based on differential evolution algorithm is an effective method in finding optimal network structure.

关 键 词：差分进化算法 神经网络 进化策略 

分 类 号：O29[理学—应用数学] TP183[理学—数学]