摘要
受限玻尔兹曼机在实际使用中不可避免地遇到设置网络结构的问题,然而对于这一问题并没有有效的方法,因此提出基于粒子群算法优化设计RBM网络结构方法(Particle Swarm Optimization-Restricted Boltzmann Machine,PSO-RBM)。该方法克服了粒子群算法在处理标称型数据时的局限性,采用连续型变量构造个体进行迭代训练,在求解适应度时再转变成标称型变量,做到可见层特征和隐藏层数目的优化选择。在MNIST数据集上实验,结果表明粒子群算法优化后的RBM网络结构在错误率和训练时间上与传统的RBM网络结构相比较,该方法在综合性能上有一定的优势,实现了粒子群算法优化设计RBM网络结构的目的。
Restricted Boltzmann machine inevitably encounters the problem of setting up network structure in practical use,but there is no effective method for this problem,so the optimization design method of RBM network structure based on particle swarm optimization(Particle Swarm Optimization-Restricted Boltzmann Machine,PSO-RBM)is proposed.This method overcomes the limitation of particle swarm optimization algorithm in dealing with nominal data.Continuous variables are used to construct indi⁃viduals for iterative training,and then transformed into nominal variables when solving fitness,so as to optimize the characteristics of visible layer and the number of hidden layers.The experimental results on MNIST dataset show that the optimized RBM network structure has some advantages in comprehensive performance compared with the traditional RBM network structure in error rate and training time,and realizes the purpose of RBM network structure optimization design by particle swarm optimization algorithm.
作者
闻一波
雷菊阳
WEN Yibo;LEI Juyang(School of Mechanical and Automotive Engineering,Shanghai University of Engineering Science,Songjiang 201620)
出处
《计算机与数字工程》
2021年第4期852-855,共4页
Computer & Digital Engineering
关键词
粒子群算法
受限玻尔兹曼机
特征
particle swarm optimization
restricted Boltzmann machine
feature