GPU上基于改进精英策略差分演化的神经网络学习算法  

GPU-based Neural Network Learning Algorithm Based on Differential Evolution with Improved Elite Strategy

作　　者：王晓峰[1,2] 吴志健[1,2] 周新宇[1,2] 郭肇禄[3] 

机构地区：[1]武汉大学软件工程国家重点实验室,武汉430072 [2]武汉大学计算机学院,武汉430072 [3]江西理工大学理学院,江西赣州341000

出　　处：《小型微型计算机系统》2016年第1期129-133,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61364025;61070008)资助

摘　　要：神经网络在诸多领域中取得了极为成功的应用,然而传统基于梯度下降的神经网络学习算法容易陷入局部最优,导致欠拟合,影响了神经网络模型的学习效果.针对该问题,提出GPU上基于改进精英策略差分演化的神经网络学习算法,利用损失函数梯度信息初始化一部分种群个体,同时利用变异种群精英信息指导搜索,并且在GPU上并行演化,在机器学习算法基准测试数据集上与传统基于误差反向传播的神经网络学习算法以及传统基于差分演化的神经网络学习算法进行了对比.同时,也与CPU上改进精英策略差分演化的神经网络学习算法的性能进行了比较,结果表明本文提出的算法训练时间更短,预测精度更高.Neural network has obtained successful application in many fields, but the learning algorithm of traditional neural networkbased on gradient descent is easy to trap in local optimum, thus influencing the learning effect of neural network model. Aiming at thisproblem, this paper proposes a GPU-based Neural Network Learning Algorithm based on Differential Evolution with Improved EliteStrategy （ GMDE）. Use the gradient descent data to initialize a part of initial population, employ the elite individual of the crossoverpopulation to guide the search, and parallel the algorithm on GPU. The benchmark datasets of machine learning algorithm are used tocompare the performance of the algorithm proposed with the traditional BP neural network learning algorithm and the traditional differ-ential evolution learning algorithm. At the mean time, this paper also compares the performance with the learning algorithm based ondifferential evolution with elite strategy running on CPU. The experiment results show that the training time of GMDE is shorter whilethe prediction precision is higher.

关 键 词：差分演化 神经网络 并行算法 GPU 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]