Study on a Database-Driven GPR Design based on Just-In-Time Approach

カテゴリ: 部門大会

論文No: SS2-5

グループ名: 【C】2023年電気学会電子・情報・システム部門大会

発行日: 2023/08/23

タイトル(英語): Study on a Database-Driven GPR Design based on Just-In-Time Approach

著者名: Kunishio Shuhei（Hiroshima University）,Wakitani Shin（Hiroshima University）

著者名(英語): Shuhei Kunishio (Hiroshima University),Shin Wakitani (Hiroshima University)

キーワード: Just-In-Time (JIT)|Gaussian Process Regression (GPR)|Online learningOnline learning|Just-In-Time (JIT)|Gaussian Process Regression (GPR)|Online learning

要約(日本語): Gaussian Process Regression (GPR) is a modeling method that is effective for systems that are difficult to model. GPR is also characterized by its ability to quantify the reliability of predictions by expressing the predicted output for input as the mean and variance of the predictive distribution. However, because GPR calculations are based on matrix operations, the computational load increases with the number of data holdings. Therefore, reducing the computational load in situations where the number of held data increases significantly is necessary. In this presentation, we present a method for online updating of GPR models by applying the Just-In-Time (JIT) approach to GPR to reduce the computational load and verification the effectiveness of the method using numerical simulations.