基于自助法的遗传编程泛化能力

Improving generalization ability in genetic programming with boostrapping

为提高遗传编程算法的泛化能力,提出基于自助法的遗传编程算法BGP。应用自助法从训练集生成自助数据集,计算训练集的训练误差和自助数据集的方差。设计融合训练误差和自助数据集方差的双层锦标赛选择策略,用于选择个体参与复制遗传算子,提高算法的泛化能力。实验结果表明,应用自助法能有效提高遗传编程算法的泛化能力,没有充分的证据说明膨胀和泛化能力存在必然联系。

To improve the generalization ability of genetic programming,a design algorithm called BGP which based on boostrapping was presented.The boostrapping datasets were generated from the training set,the training error of the training set and the variance of the boostrapping dataset were calculated.Based on the training error and the boostrapping dataset variance,a twolayer tournament was proposed to select individuals to participate in the elitism genetic operator.The results show that,for the undertaken problems,the genetic programming based on boostrapping generalizes better than the standard genetic programming,and the results also show that there is no necessary relationship between bloat and generalization in genetic programming.