Interpretability of Machine Learning-Assisted Control Methods through Stochastic Controllability Analysis

2021 Fiscal Year Final Research Report

• Project/Area Number: 18H01461
• Research Category: Grant-in-Aid for Scientific Research (B)
• Allocation Type: Single-year Grants
• Section: 一般
• Review Section: Basic Section 21040:Control and system engineering-related
• Research Institution: Kyoto University
• Principal Investigator: Kashima Kenji 京都大学, 情報学研究科, 准教授 (60401551)
• Co-Investigator(Kenkyū-buntansha): 辻野 博文 大阪大学, 薬学研究科, 助教 (10707144) ; 山下 沢 武庫川女子大学, 薬学部, 准教授 (70398246)
• Project Period (FY): 2018-04-01 – 2022-03-31
• Keywords: 制御工学 / 機械学習 / 情報通信工学 / 薬学 / 確率統計

Outline of Final Research Achievements

While control systems that utilize large amounts of data and machine learning methods are expected to have nonlinearity and the ability to flexibly respond to changes in the environment, the lack of theoretical performance guarantees that have been ensured using mathematical models so far makes it difficult to dispel vague concerns when it comes to industrial applications. In this study, we developed the data-driven model reduction theory proposed by the principal investigator and derived theoretical results for de-black-boxing machine learning-assisted methods by giving them a stochastic control theoretical interpretation.

Free Research Field

制御工学

Academic Significance and Societal Importance of the Research Achievements

学術的には、機械学習的手法により抽出した特徴量を状態変数とする低次元モデル構築や、差分プライバシー解析、深層強化学習にもとづく自己駆動型制御システム設計、軌道の位相的性質を活用する動的システム学習とモデル予測制御への応用など、統計的学習理論とシステム制御理論の融合を推進した。社会的には、工学応用期に移行しつつある機械学習技術の信頼性向上及び実応用検証による有効性の確認をおこなった。