高斯過程被廣泛地運用在電腦實驗的建模上，但是隨著資訊爆炸的時代來臨，資料量成長的速度遠遠超過硬體設備的進步，使得高斯過程的預測能力在樣本數很大時急遽下降。常見的解決方法是透過改進模型或是實驗設計降低樣本數。本篇論文提出一個座標交換演算法，在給定樣本中迭代找出最具有預測力（predictive）的子樣本（subdata），且透過數值模擬來展示只需使用這些子樣本就可以達到準確的預測結果。;Gaussian processes are widely used for emulating expensive computer simulators. However, their use is limited under large-scale data due to numerical problems. Common solutions to this difficulty are either to modify the model or to design suitable computer experiments. In this paper, we propose an exchange algorithm to search for predictive subdata given a full dataset. We demonstrate our method on several examples to show the performance between our method and other competitive methods. Furthermore, we show that the predictive subdata selected by our method achieves high prediction accuracy even though the full dataset is not space-filling in the experimental region.
