[3D5-GS-2] Machine learning

[3D5-GS-2-04] Regression Modeling and Uncertainty Estimation Based on Target Prior Knowledge

〇Daisuke Azuma¹, Takashi Washio¹ (1. Osaka university)

Keywords:Physical Regression Modeling, Data Assimilation, Estimating Parameter Distribution

In this study, we propose a method to obtain a regression model having a structure based on our prior knowledge about a target. The method estimates the distribution of the model parameters and objective quantity using stochastic approximation. It enables to estimate the posterior distribution of model parameters and objective quantities with clear meaning in light of the prior knowledge, with excellent extrapolation ability under the prior knowledge. Specifically, using a physical model of the target as a regression model, we estimate the maximum posterior probability of the model parameters from the observed data of the target, and estimate the parameter posterior distribution under the Laplace approximation. Furthermore, the posterior distribution of the objective quantity is estimated. We applied this method to physical systems such as a double pendulum system, and confirmed that a more accurate regression model than Gaussian process regression modeling can be obtained.