[4M1-GS-2a] Machine learning

10:00 AM - 10:15 AM JST(1:00 AM - 1:15 AM UTC)

[4M1-GS-2a-05] Level-Set Estimation Based on Observation Noise for Tuning Parameters of Machine Tools

〇Renshi Nagasawa¹, Hideyuki Masui¹, Koki Nakane¹, Yu Inatsu², Masayuki Karasuyama² (1. Mitsubishi Electric Corporation, 2. Nagoya Institute of Technology)

Keywords:

Level Set Estimation,Active Learning,Gaussian Process Regression,Bayesian Optimization

Identifying the level set of parameters for which a quality metric exceeds a predefined threshold is crucial for maintaining safety margins in machine-tool parameter tuning. Since the parameter–quality relationship is typically treated as a black box and on-machine trials are costly, active learning is required to estimate the level set accurately with as few trials as possible. In our previous work, we used Gaussian process regression for level-set estimation, employing a termination criterion based on misclassification risk and an acquisition function to reduce the number of trials. However, this criterion proved overly conservative on noisy real-world data, limiting practical use. Therefore, we propose a new termination criterion that appropriately relaxes the uncertainty evaluation used for termination, according to the estimated measurement-noise level. Experiments on real machining data demonstrate that the proposed method achieves sufficient accuracy and terminates within a practical number of trials.