基于QPSO和极限学习的离散过程神经网络及学习算法  

Discrete process neural networks and learning algorithms base on QPSO and extreme learning machine

作　　者：刘志刚[1] 许少华[1,2] 李盼池[1] 

机构地区：[1]东北石油大学计算机与信息技术学院,黑龙江大庆163318 [2]山东科技大学信息科学与工程学院,山东青岛266590

出　　处：《控制与决策》2016年第12期2241-2247,共7页Control and Decision

基　　金：国家自然科学基金项目(61170132);黑龙江省教育厅科学技术研究项目(12521369)

摘　　要：连续过程神经元网络在权函数正交基展开时,基函数个数无法有效确定,因此逼近精度不高.针对该问题,提出一种离散过程神经元网络,使用三次样条数值积分处理离散样本和权值的时域聚合运算.模型训练采用双链量子粒子群完成输入权值的全局寻优,通过量子旋转门和非门完成种群进化.局部使用极限学习,通过Moore-Penrose广义逆计算输出权值.以时间序列预测为例进行仿真实验,结果验证了模型的有效性,且训练收敛能力和逼近能力都有一定程度的提高.When the weight functions of the continuous process neural network are expanded by orthogonal basis, the number of the basis function can not be determined effectively. The continuous process neural network has lower approach accuracy. Therefore, a discrete process neural network is presented. The three spline numerical integration is applied to deal with the aggregation of discrete samples and weights in time-domain. The double chain quantum particle swarm algorithm is used to the global optimization of model parameters. The evolution of the population is executed by the quantum rotation gate and quantum not gate. The extreme learning algorithm is applied to the local search and the output weights are computed by the Moore-Penrose inverse. The results of the simulation experiment based on the time sequence prediction verify the effectiveness of the proposed model, and show that the capability of training convergence and accurate approximation are improved at a certain degree.

关 键 词：过程神经元网络 极限学习 量子粒子群 MOORE-PENROSE广义逆 网络训练 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]