基于自编码器和填充函数的深度学习优化算法  

Depth learning optimization algorithm based on the autoencoder and filling function

作　　者：薛南 吕柏权[1] 倪陈龙 Xue Nan;Lv Baiquan;Ni Chenlong(College of Mechatronic Engineering and Automation,Shanghai University,Shanghai 200072,China)

机构地区：[1]上海大学机电工程与自动化学院

出　　处：《电子测量技术》2019年第23期79-84,共6页Electronic Measurement Technology

摘　　要：针对粒子群算法容易陷入局部极小值的问题,提出了一种基于自编码器和填充函数的深度学习优化算法。深度学习采用的是分层的训练机制,它通过最小化误差函数进行分层的无监督训练。首先使用自编码器在无标签的训练样本上分层预训练得到各层参数,然后使用监督学习的方法对网络参数进行微调,从而减少陷入局部极小点的可能性。另外,引进了填充函数法,使之能够跳出局部最小值,继续迭代至更小的极值点,得到精度更高的全局最优点。本文选择了4个比较典型的基准测试函数进行仿真,通过分析基于自编码器和填充函数的深度学习优化算法对每个基准函数的收敛速度和搜寻精度评价该算法的优劣。结果表明优化算法成功收敛到全局最优解,验证了改进后的算法满足函数优化要求。Aiming at the problem that particle swarm optimization is easy to fall into local minimum value,this paper proposes a depth learning optimization algorithm based on Autoencoder and filling function.Depth learning adopts a layered training mechanism,which carries out layered unsupervised training by minimizing error function.In the first step,unlabeled training samples are used to conduct stratified pre-training for each layer of parameters;in the second step,labeled training samples are used to conduct supervised fine-tuning for the entire network parameters.This helps to reduce the possibility of falling into local minima.In addition,the method of filling function is introduced,which makes it possible to jump out of the local minimum and continue to iterate to the smaller extreme point,thus obtaining the global optimum with higher precision.In this paper,four typical benchmark functions are selected for simulation.The depth and learning accuracy of each benchmark function is evaluated by analyzing the deep learning optimization algorithm based on self-encoder and fill function.The results show that the optimization algorithm successfully converges to the global optimal solution,and verifies that the improved algorithm satisfies the function optimization requirements.

关 键 词：自编码器 填充函数 全局最优值 深度学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]