改进遗传模拟退火算法优化BP算法研究  

Rsearch on Improved BP Algorithm for Genetic Simulated Annealing Algorithm

作　　者：郭彩杏 郭晓金[2] 柏林江 GUO Cai-xing;GUO Xiao-jin;BAI Lin-jiang(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065 ,China;Institute of Broadband Networks and Information Processing,Chongqing University of Posts and Telecommunications,Chongqing 400065 ,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]重庆邮电大学宽带网络及信息处理研究所,重庆400065

出　　处：《小型微型计算机系统》2019年第10期2063-2067,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61671094)资助;重庆市科委项目(CSTC2015JCYJA40032)资助

摘　　要：针对传统BP神经网络在函数拟合中收敛速度慢、精度低的缺点,提出一种改进遗传模拟退火算法优化的BP神经网络算法(IGSAA-BP).该算法首先根据进化中种群适应度的集中分散程度改进了自适应遗传算法的交叉和变异概率公式,使算法能够更加有效地避免陷入局部最优;然后根据旧种群和新种群中每个对应个体的进化程度提出一种改进的Metropolis准则,分情况修正新种群中的所有个体,增加种群个体的多样性,提高了算法的全局寻优能力.利用改进遗传模拟退火算法初始化BP神经网络的权阈值,并与GA-BP、IAGA-BP网络对比.实验表明,IGSAA不仅提高了BP网络的收敛速度,还有效地提高了网络的拟合能力,拟合精度提高了5%.Targeted at the traditional BP neural network has the disadvantages of slow convergence and low precision in function fitting,an improved BP neural network algorithm( IGSAA-BP) based on improved genetic simulated annealing algorithm is proposed.The algorithm firstly improves the crossover and mutation probability formulas of the adaptive genetic algorithm according to the degree of concentration and dispersion of population fitness in evolution,so that the algorithm can more effectively avoid falling into the local optimum;Then according to the evolution degree of each corresponding individual in the old population and the new population,an improved Metropolis criterion is proposed. The criterion modifies all individuals in the new population in different situations to increase the diversity of individual populations and improve the global optimization ability of the algorithm. The weighting threshold of BP neural network is initialized with improved genetic simulated annealing algorithm and compared with GA-BP and AGA-BP networks. The experiment show s that IGSAA not only improves the convergence speed of BP network,but also effectively improves the fitting ability of the network,and 5% increase in fitting accuracy.

关 键 词：神经网络 BP算法 自适应遗传算法 模拟退火算法 METROPOLIS准则 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]