基于相关分析的多目标优化Pareto优劣性预测

发布时间：2019-01-24 07:59

【摘要】：昂贵多目标进化算法中,目标向量评估所需计算时间或实验成本高昂,大量昂贵评估必然导致成本灾难.本文根据多目标优化Pareto优劣性取决于各目标分量的序关系这一关键性质,提出一种序拟合方法进行Pareto优劣性预测.在分析样本数据决策空间与目标空间序相关性的基础上,通过线性相关的假设条件,建立低成本的序关系预测方程,并用预测的序关系确定Pareto优劣性.然后对典型多目标优化问题进行Pareto优劣性预测对比实验,结果表明所提方法显著提高了Pareto优劣性的预测精度.最后,将该预测方法集成到NSGA-II算法中,可以避免进化过程中的模型重构,有效减少昂贵目标向量的评估次数.
[Abstract]:In the expensive multi-objective evolutionary algorithm, the computation time or the experimental cost of the objective vector evaluation is high, and a large amount of expensive evaluation will inevitably lead to the cost disaster. In this paper, according to the key property that the quality of multi-objective optimization Pareto depends on the order relation of each target component, an ordered fitting method is proposed to predict the superiority and inferiority of Pareto. Based on the analysis of the order correlation between the decision space and the target space of the sample data, the low cost order relation prediction equation is established by the assumption of linear correlation, and the superiority and inferiority of Pareto are determined by the order relation of the prediction. Then the comparison experiment of Pareto prediction for typical multi-objective optimization problems is carried out. The results show that the proposed method can improve the prediction accuracy of Pareto significantly. Finally, integrating the prediction method into NSGA-II algorithm can avoid the model reconfiguration in the evolution process and effectively reduce the evaluation times of expensive target vectors.
【作者单位】：中南大学信息科学与工程学院;湖南理工学院信息与通信工程学院;
【基金】：国家自然科学基金(No.60975049,No.61174132) 湖南省省教育厅科学研究重点项目(No.15A079) 湖南省高校科技创新团队支持计划资助
【分类号】：TP18

【相似文献】