华东师范大学学报(自然科学版)

• 计算机科学 • 上一篇    下一篇

一种基于粒子群优化的极限学习过程神经网络

刘志刚[1] , 许少华[2] , 李盼池[1]   

  1. 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
  • 收稿日期:2015-06-12 出版日期:2016-07-25 发布日期:2016-09-29
  • 通讯作者: 刘志刚, 男, 博士研究生, 副教授, 研究方向为过程神经网络、进化算法. E-mail: dqpilzg@163.com.
  • 基金资助:

    国家自然科学基金 (61170132); 中国博士后科学基金 (201003405)

An extreme learning process neural networks based on particle swarm optimization

LIU Zhi-gang[1] , XU Shao-hua[2] , LI Pan-chi[1]   

  1. 1.School of Computer and Information Technology, Northeast Petroleum University, Daqing Heilongjiang 163318, China;
    2.College of Information Science and Engineering, Shandong University of Science and Technology, Qingdao Shandong 266590, China
  • Received:2015-06-12 Online:2016-07-25 Published:2016-09-29

RichHTML

7

PDF

588

摘要：

本文针对过程神经元网络(Process Neural Network, PNN)模型学习参数较多, 正交基展开后的梯度下降算法初值敏感、计算复杂、不易收敛等问题, 结合极限学习机(Extreme Learning Machine, ELM)的快速学习特性, 提出了一种新型的极限学习过程神经元网络. 学习过程中摒弃梯度下降算法的迭代调整策略, 采用 Moore-Penrose 广义逆计算输出权值矩阵. 同时为弥补极限学习机由于随机赋值造成的不足, 利用粒子群算法(Particle Swarm Optimization, PSO)良好的全局搜索能力进行模型参数优化, 获得紧凑的网络结构, 提高了模型泛化能力. 仿真实验以 Henon 混沌时间序列和太阳黑子预测为例, 验证了网络的有效性.

关键词: 过程神经元网络; 极限学习机, 粒子群, Moore-Penrose广义逆, 网络训练

Abstract:

Aiming at the problems that process neural network has more learning parameters, sensitive to initial value, complicated computation and difficult to converge
for the gradient descent algorithm based on orthogonal basis expansion, a new process neural network based on extreme learning machine is presented in this paper. The iterative adjustment strategy is rejected in the trainning process and use Moore-Penrose to calculate the output weight matrix. In order to make up for the lack of random assignment for the extreme learning machine, the particle swarm algorthim is taken and the parameters are optimized with its global search ability. This algorthim can get the more tightly network structure and improve the model generalization ability. The model and algorthim are applied to Henon chaotic time series and sunspot prediction. The simulation results confirm the validity and feasibility of the model and learning algorithm.

Key words: process neural network, extreme learning machine, particle swarm, Moore-Penrose generalized inverse, network training