能動学習を用いた機械学習による解析近似器の精度向上に関する研究

抄録

The machine learning approximator of the simulation output has a large error when the training data is small because the training data is discrete. Therefore, we thought that the accuracy could be improved by additional learning based on the concept of active learning. In response to this, we added learning points to the middle points of the neighboring tents of the learning points with large errors, but so far we have only been able to derive points that improve the accuracy by about 70%. In this study, we found that by recognizing the shape of the learned approximator as a distribution of curvature, we can propose an appropriate additional training point with a probability of 90%, which is expected to improve the accuracy.