多任务Kriging变量选择的研究与应用

Research and Application on Variable Selection in Multi-Task Kriging Model

本文研究多任务Kriging模型的变量选择问题,并给出多种稀疏化惩罚下多任务Kriging的变量选择算法。数值模拟及实例分析表明,相比单任务的Kriging变量选择,多任务模式能显著提高计算效率而不失模型拟合的准确性;相比LMC及卷积模型,多任务稀疏化Kriging能有效提取任务间的共性信息,极大节约计算成本同时提高预测精度。

We study the variable selection in multi-task Kriging model and develop the algorithms for com-monly used penalizations. In numerical simulations, our multi-task penalized approach achieves higher computational efficiency without loss of accuracy and stability compared to the single-task approach. In real data application, multi-task penalized Kriging effectively captures shared features among tasks and thus reduces computational burden compared with the LMC and CONV models.