메뉴 건너뛰기
소속 기관 / 학교 인증
인증하면 논문, 학술자료 등을  무료로 열람할 수 있어요.
한국대학교, 누리자동차, 시립도서관 등 나의 기관을 확인해보세요
(국내 대학 90% 이상 구독 중)
고객센터 ENG
주제분류

논문 기본 정보

저자정보
저널정보
대한전자공학회 대한전자공학회 학술대회 2021년도 대한전자공학회 하계종합학술대회 논문집
오류 신고하기

피인용 0

검색

    초록·키워드

    The main challenge of reinforcement learning from image-based observation is sample inefficient. Learning a good general feature representation on upstream task which then fine-tune to multiple downstream tasks is a good strategy but haven’t well explore. Using this strategy, this paper proposes a Self-supervised Pretraining Framework for Learning feature Representation (SPF) in visual observation reinforcement learning. SPR pre-train an general feature encoder on proxy task follow forward dynamic model signal using contrastive learning loss instead of relying on specific task reward. Then, this general feature encoder can be reuse on multiple downstream tasks to learn reinforcement policy. Experimental results on a diverse set of image-based continuous control tasks from Deep Mind Control Suite shows that SPF significantly improve sample efficiency compare to recent advance approaches.

    최근 본 자료 전체보기