摘要
给出了用模糊感知器学习算法和(ε,δ)准则估计多元线性回归模型回归系数的详细算法,讨论了学习速率、ε和δ的设定;并与经典的回归系数估计方法最小二乘法作比较,发现总体拟合最好的特性对于含异常数据(noisydata)的情况反而会使预测值背离事实更远,而基于模糊感知器的学习算法实现线性回归具有编程简单、对数据无特殊要求而且对数据的容错性较高的优点,可用来实现数据挖掘所需要的预测和异常检测功能。
In this paper a detailed algorithm for multiple linear regression coefficients evaluation is provided, which employs a training algorithm for fuzzy perceptron based on a socalled (ε,δ)criteria . After the setting of the learning rate λ and the fuzzy degree (ε,δ) are addressed, the neural method named FPMLR is compared to the classical least squares method for regression coefficients evaluation. As a result, we conclude that the neural method is superior to the least squares method in robustness aspect due to its feature of fuzziness, which may contribute to omitting noisy data. Finally, we suggest that the neural method be applied to the implementation of prediction and abnormal detection required by data mining.
出处
《南京邮电学院学报（自然科学版）》
2002年第4期33-38,共6页
Journal of Nanjing University of Posts and Telecommunications
