[2J1-GS-2] Machine learning: Gaussian process model

[2J1-GS-2-01] Bayesian Active Learning for Inverse Problem of Structured-Output

〇Kota Matsui¹, Shunya Kusakawa², Keisuke Ando³, Kentaro Kutsukake¹, Toru Ujihara^3,4, Ichiro Takeuchi^2,1,5 (1. RIKEN Center for Advanced Intelligence Project, 2. Nagoya Institute of Technology, 3. Nagoya University, 4. AIST, 5. NIMS)

Keywords:active learning, inverse problem, Gaussian process model

We propose an active learning method for the inverse problem of finding input parameters that achieve the desired structured-output. Here, the structured-output refers to a multidimensional vector in which each element has a correlation. Specifically, we propose three acquisition functions to minimize the squared error between the desired structured-output and the prediction by the model by explicitly incorporating the correlation between output elements for a black-box vector-valued objective function into a Gaussian process model. We apply the proposed method to the search problem of growth rate distribution using actual data of silicon carbide (SiC) crystal growth modeling, and verify its effectiveness.