摘要
分析了利用遗传规划进行复杂非线性系统建模时容易出现过学习现象问题的原因,提出了一个基于插值函数保护法、一个评价函数光滑度的准则和基于多目标非支配排序的改进的遗传规划方法.利用非支配排序的思想结合传统的遗传规划来实现对于模型的精确度、复杂度和光滑度的平衡从而提高学习结果的泛化能力.将该方法应用于工业阿维菌素发酵过程中的菌丝浓度估计,取得了良好的效果.
To eliminate the overfitting phenomenon in genetic programming （GP）, a new protected approach based on interpolation for some mathematical functions and a method to estimate the smoothness of a GP tree are proposed. Furthermore, to get more accuracy model without losing generalization, the evenness,the complexity and the training errors of the model should all be considered. However, their effects on the training process are difficult to be decided so a non-dominated sorting based multi-objective GP is proposed to make the evaluation of these effects impersonally. Finally, to test the capability of the new approach, a real modeling problem arising from industrial Avermectin process and numerical examples were used and the results show that the improved GP has more chance to find an acceptable solution with higher generalization and without losing accuracy.
出处
《传感技术学报》
CAS
CSCD
北大核心
2007年第9期2029-2034,共6页
Chinese Journal of Sensors and Actuators
基金
浙江省科技计划项目资助(2006C11066
2006C31051)
浙江省院士基金项目资助(2005A1001-13)
安徽省教育厅科研资助项目(2005jq1033)
安徽省教育厅科研资助项目(2005jq1033)
作者简介
吴燕玲（1971-），女，博士生，从事进化计算、过程建模、智能控制等研究，ylwu@iipc．zju．edu．on,
卢建刚（1968-），男，副教授，从事复杂工程系统的建模、控制与优化的研究，jglu@iipc．zju.edu．cn
孙优贤（1940-），男，中国工程院院士，教授，博士生导师，从事鲁棒控制与容错控制的研究．
