利用单形进化优化的BP神经网络学习算法

在BP神经网络训练算法中,针对权值的优化学习容易陷入局部极值点、收敛速度慢等问题,很多研究引入智能优化算法对其进行改进,但传统的智能优化算法通常有多个控制参数,若不能正确选取参数,或者没有适当选择初始点位置,则很难搜索到最优的神经网络权值。为了解决这些问题,提出一种基于单形进化的BP神经网络学习算法,它通过全随机搜索减少算法的控制参数,利用群体的多角色态保持粒子的多样性,避免算法陷入局部的极值点,减少了对初始值的依赖。在应用中,将该算法应用于神经网络的训练算法中,通过对UCI数据集和人脸图像的测试,实验结果表明,上校算法训练的神经网络有效提高了识别率与训练效率。

BP Neural Network Learning Algorithm Using Surface-Simplex Swarm Evolution

For the BP neural network training algorithm,the optimization of weights is easy to fall into the local extreme points and the convergence rate is slow.In order to solve above problems,many scholars introduce the intelligent optimization algorithm,However,There are multiple control parameters in traditional intelligent optimization algorithms usually.And if the selection of these parameters or the initial point position is not suitable,it is difficult to search for the optimal neural network weights.To overcome these problems,in this paper,we propose a BP neural network learning algorithm based on surface-simplex swarm evolution which reduces the control parameters of the algorithm by all random search.And in this method,the diversity of the particles was maintained by the polygonal state of the population,the algorithm was prevented from falling into the local extreme points,and the dependencies of initial value can be reduced.In the application,the algorithm was applied to the neural network training algorithm.By testing the face images and UCI data,the experimental results show that the neural network trained by the algorithm can improve the recognition rate effectively,increased by 14%and 10%points respectively.