Data assimilation based reinforcement learning

2015 Fiscal Year Final Research Report

• Project/Area Number: 25730135
• Research Category: Grant-in-Aid for Young Scientists (B)
• Allocation Type: Multi-year Fund
• Research Field: Intelligent informatics
• Research Institution: The University of Tokyo (2015); Osaka University (2013-2014)
• Principal Investigator: Ueno Tsuyoshi 東京大学, 新領域創成科学研究科, 特任研究員 (90615824)
• Project Period (FY): 2013-04-01 – 2016-03-31
• Keywords: 強化学習 / データ同化 / 機械学習 / 人工知能

Outline of Final Research Achievements

Learning action strategies from computer simulations has a potential to achieving the drastic productivity increases because it has no necessity to perform an expensive process, i.e., real experiments for collecting the data. However, the behavior of simulations often differ from that of actual environments; thus, it is not rare that the action strategy obtained from the simulation makes no sense in practical applications. In this project, we developed a new framework, so-called data assimilation reinforcement learning (DARL) which incorporates data assimilation and reinforcement learning. DARL can provide the good action strategy in the small number of experiments by learning not only the action strategy but also the computer simulation simultaneously. We have also applied DARL to material design and drug discovery problems and confirmed its effectiveness compared with current methods.

Free Research Field

人工知能