Adaptive Power Management of IoT Systems by Reinforcement Learning
Project/Area Number |
18J20946
|
Research Category |
Grant-in-Aid for JSPS Fellows
|
Allocation Type | Single-year Grants |
Section | 国内 |
Research Field |
Computer system
|
Research Institution | The University of Tokyo |
Research Fellow |
SHRESTHAMALI SHASWOT 東京大学, 情報理工学系研究科, 特別研究員(DC1)
|
Project Period (FY) |
2018-04-25 – 2021-03-31
|
Project Status |
Completed (Fiscal Year 2020)
|
Budget Amount *help |
¥2,200,000 (Direct Cost: ¥2,200,000)
Fiscal Year 2020: ¥700,000 (Direct Cost: ¥700,000)
Fiscal Year 2019: ¥700,000 (Direct Cost: ¥700,000)
Fiscal Year 2018: ¥800,000 (Direct Cost: ¥800,000)
|
Keywords | Multi-objective RL / Wireless Sensor Nodes / Reinforcement Learning / Internet of Things / Distributed Learning / Deep Q Networks / Wireless Sensor Networks / Machine Learning / Edge Intelligence / Distributed Reinforcement Learning / Energy Harvesting Wireless Sensor Nodes / Deep Q- Learning |
Outline of Annual Research Achievements |
The research achievement for this year was two fold - 1)developing a Multi-Objective Reinforcement Learning (MORL) approach for Energy Neutral Operation (ENO) of Energy Harvesting Wireless Sensor Nodes (EHWSNs), and 2) preparing the PhD thesis. My previous research efforts concentrated on achieving ENO using a single reward function. However, one cannot optimize energy scheduling between various tasks in EHWSNs with this approach. So, I investigated into MORL methods to optimize energy scheduling over multiple tasks that modern EHWSNs execute. Since previous RL methods could not perform runtime tradeoffs along the Pareto-space and/or require prohibitively large amounts resources, I developed a novel multi-objective RL framework for EHWSNs. This framework can optimize over multiple objectives and tradeoff dynamically. It consumes much less resources compared to direct multi-objective RL methods making them suitable for resource constrained EHWSNs. I remodeled the EHWSN system, developed a multi-objective Markov Decision Process (MDP) and proposed two novel MORL algorithms that enables EHWSNs to learn tradeoff policies in shorter time periods while making lesser mistakes. Our paper on this novel method is currently under review. I also consolidated the different results from my previous research papers to prepare my PhD thesis and defense presentation. My research (supported by this KAKENHI) was also met with high enthusiasm in the research community and I was invited to be the publicity chair of AIChallenge IoT 2020 Workshop that was co-located (virtually) with SenSys 2020.
|
Research Progress Status |
令和2年度が最終年度であるため、記入しない。
|
Strategy for Future Research Activity |
令和2年度が最終年度であるため、記入しない。
|
Report
(3 results)
Research Products
(1 results)