高维进化策略调整神经网络结构和参数  

High Dimensional Evolutionary Strategy for Tuning the Structure and Parameters of a Neural Network

作　　者：肖赤心[1,2] 蔡自兴[1] 王勇[1] 

机构地区：[1]中南大学信息科学与工程学院,湖南长沙410083 [2]湘潭大学信息工程学院,湖南湘潭411105

出　　处：《小型微型计算机系统》2008年第12期2313-2318,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金(60234030,60404021)资助;国家基础研究项目(A1420060159)资助;湖南省院士基金项目(07IJY3035)资助

摘　　要：提出一种基于佳点集原理的进化策略用于神经网络结构和参数的调整.为了克服正交设计法的一些不足来处理高维最优化问题,本文采用分步交叉框架,将佳点集技术引入实数域交叉算子增强高维空间的搜索能力.前馈神经网络的隐含节点与连接边数从小逐步递增直至学习效果足够好.通过调整能得到一个部分连接的前馈网络,减少了网络实现的耗费.最后,佳点集进化策略有效应用于生成预测太阳黑子的演化神经网络.实验结果证明了新方法的有效性.In this paper, a Good Node Set-evolutionary strategy （GNSES） is applied to solve the problem of tuning both network structure and parameters of a feedforward neural network. To overcome the deficiency of orthogonal design to handle high dimensional optimization problems, this paper presents a muti-steps crossover frame incorporated with GNS principle for real domain to enhance the exploitation capabilities of the evolution strategy in high dimensional space. The numbers of hidden nodes and the links of the feedforward neural network are chosen by increasing them from small numbers until the learning performance is good enough. As a result, a partially connected feedforward neural network can be obtained after tuning. The cost of implementation of the neural network can be reduced. The presented GNSES approach is effectively applied to generating the evolving neural network for forecasting the sunspot numbers. Experiment results show the efficiency of our methods.

关 键 词：演化神经网络 进化策略 佳点集 预测太阳黑子 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]