Study on Effective Learning on Multi-objective Sequential Optimization and its Applications

2019 Fiscal Year Final Research Report

• Project/Area Number: 16K01269
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Social systems engineering/Safety system
• Research Institution: Kansai University
• Principal Investigator: YUN YEBOON 関西大学, 環境都市工学部, 教授 (10325326)
• Project Period (FY): 2016-04-01 – 2020-03-31
• Keywords: 逐次近似最適化 / 機械学習 / モデル予測 / メター学習

Outline of Final Research Achievements

In optimization with high cost objective function, an approximate function is used as a surrogate. For generating an approximate function based on some sample points, machine learning such as Radial basis networks and Support vector machines is effective. The precision of approximate functions depends on hyper-parameters used in basis and kernel functions, which is deeply related to learning. The new methods of meta-learning in SVM and RBF networks were proposed with the aim of generating an approximate function of high accuracy with small number of function evaluations. Furthermore, for multi-objective optimal control problems based on meta-model under a dynamic environment, this study suggested the method of combining machine learning methods and predetermined linear model in order to construct more accurate and stable model prediction, Finally, the effectiveness of the proposed methods in this research was validated through some numerical examples and engineering design problems.

Free Research Field

社会システム工学

Academic Significance and Societal Importance of the Research Achievements

計算知能や多目的最適化法に関し、そのものを対象とした研究はすでに多く存在する。しかし、基礎的研究にとどまることが多く、実際の応用という観点からの検討が不十分であるか、逆に理論的な根拠は希薄であるが、これまでの経験に基づく方法による試行錯誤的な研究も多いというのが現状である。さらに、GAやPSOの進化的アルゴリズムを用いたパレート解の生成法に関する研究は活発であるが、既存の方法では多くの計算回数を要する。これらのことを総合的に踏まえたうえで、多角な観点から、理論のみならず実問題へ適用性も考慮した成果であり、学術的にも実用的にも有意義であると考える。