摘要
将神经网络总体平均误差作为目标函数 ,以待求的神经网络权值和阈值作为设计变量 ,通过设计变量合理排序与分配 ,提出多隐层多层神经网络权值和阈值计算的高精度真实共轭梯度最优化算法·与BP算法和梯度优化算法相比 ,既能实现每步迭代在搜索方向上获得最优步长保证目标函数递减 ,又能克服在目标点附近的振荡现象·编制出神经网络权值和阈值计算的通用程序 ,给出神经网络合理结构选择的基本原理·通过足球机器人位置分析算例的神经网络分析和模式识别 ,表明所提出算法的有效性和实际应用价值·
Defining network average error as optimum objective function, weights and thresholds as design variables,a new kind of real conjugate gradient optimum algorithm were studied. The method overcomes the oscillation phenomenon. The best step length can be aquired per compution. The objective function decreases gradually. A computing program about weights and threshold, based on accurate conjugate gradient optimum algorithm of multi layer neural network, was put forward. The selecting method of rational construct was pointed out. Through analyzing neural network of status of soccer robots and pattern recognition, its validity and applying prospect was showed.
出处
《东北大学学报(自然科学版)》
EI
CAS
CSCD
北大核心
2002年第1期20-23,共4页
Journal of Northeastern University(Natural Science)
基金
国家自然科学基金资助项目 (68975 0 0 3 )
辽宁省博士起动基金资助项目 (2 0 0 110 2 0 17)
关键词
最优化方法
共轭梯度法
多层神经网络
权值
阈值
网络合理结构
模式识别
optimum algorithm
conjugate gradient optimum algorithm
multi layer neural network
weights and threshold
network rational construct
pattern recognition