- 【Updated on May 12, 2025】 Integration of CiNii Dissertations and CiNii Books into CiNii Research
- Trial version of CiNii Research Knowledge Graph Search feature is available on CiNii Labs
- Suspension and deletion of data provided by Nikkei BP
- Regarding the recording of “Research Data” and “Evidence Data”
Weighted Average Composition of Deep Reinforcement Learning Agents in Discrete Action Problems
-
- SATO Kenichiro
- Tokyo Institute of Technology
-
- KOHJIMA Masahiro
- NTT Corporation
-
- MATSUBAYASHI Tatsushi
- NTT Corporation
-
- TODA Hiroyuki
- NTT Corporation
Bibliographic Information
- Other Title
-
- 深層強化学習Agentの離散行動空間タスクにおける重み付き結合
Search this article
Description
Composition of pre-trained agents is gathering attention in the field of reinforcement learning since this approach allows us to construct an agent that solves a new task by combining multiple pre-trained agents that solve different tasks. In this study, we extend an existing method that composes pre-trained agents with simple average and propose a new method that composes pre-trained agents with a weighted average. The proposed method enables us to solve a new task whose reward function is expressed as the linear combination of base tasks. We verify the effectiveness of the proposed method by CartPole control and traffic signal control problems.
Journal
-
- 電子情報通信学会論文誌D 情報・システム
-
電子情報通信学会論文誌D 情報・システム J103-D (5), 403-414, 2020-05-01
The Institute of Electronics, Information and Communication Engineers
- Tweet
Keywords
Details 詳細情報について
-
- CRID
- 1390285300154149632
-
- ISSN
- 18810225
- 18804535
-
- Text Lang
- ja
-
- Data Source
-
- JaLC
-
- Abstract License Flag
- Disallowed