講演抄録/キーワード |
講演名 |
2021-11-19 11:10
Toward Human Cognition-inspired High-Level Decision Making For Hierarchical Reinforcement Learning Agents ○Rousslan Fernand Julien Dossa(Kobe Univ.)・Takashi Matsubara(Osaka Univ.) CCS2021-28 |
抄録 |
(和) |
Hierarchical reinforcement learning (HRL) methods aim to leverage the concept of temporal abstraction to efficiently solve long-horizon, sequential decision-making problems with sparse and delayed rewards.
However, the decision-making process of the agent in most HRL methods is often based directly on low-level observations, while also using fixed temporal abstraction.
We propose the hierarchical world model (HWM), which can capture more flexible high-level, temporally abstract dynamics, as well as low-level dynamics of the system.
We posit such model is a natural extension to the HRL framework toward a decision-making process closer to that of humans. |
(英) |
Hierarchical reinforcement learning (HRL) methods aim to leverage the concept of temporal abstraction to efficiently solve long-horizon, sequential decision-making problems with sparse and delayed rewards.
However, the decision-making process of the agent in most HRL methods is often based directly on low-level observations, while also using fixed temporal abstraction.
We propose the hierarchical world model (HWM), which can capture more flexible high-level, temporally abstract dynamics, as well as low-level dynamics of the system.
We posit such model is a natural extension to the HRL framework toward a decision-making process closer to that of humans. |
キーワード |
(和) |
Reinforcement learning / Hierarchical reinforcement learning / World models / Temporal abstraction / Hierarchically organized behavior / / / |
(英) |
Reinforcement learning / Hierarchical reinforcement learning / World models / Temporal abstraction / Hierarchically organized behavior / / / |
文献情報 |
信学技報, vol. 121, no. 253, CCS2021-28, pp. 61-66, 2021年11月. |
資料番号 |
CCS2021-28 |
発行日 |
2021-11-11 (CCS) |
ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
CCS2021-28 |