Accelerating multi-step sparse reward reinforcement learning

Yang, Xintong ORCID: https://orcid.org/0000-0002-7612-614X and Ji, Ze ORCID: https://orcid.org/0000-0002-8968-9902 2024. Accelerating multi-step sparse reward reinforcement learning. Presented at: Cardiff University Engineering Research Conference 2023, Cardiff, UK, 12-14 July 2023. Published in: Spezi, Emiliano and Bray, Michaela eds. Proceedings of the Cardiff University Engineering Research Conference 2023. Cardiff: Cardiff University Press, pp. 86-90. 10.18573/conf1.u

Preview

PDF - Published Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (538kB) | Preview

Abstract

After the great successes of deep reinforcement learning (DRL) in recent years, developing methods to speed up DRL algorithms for more complex tasks closer to those in the real world has become increasingly important. In particular, there is a lack of research on long-horizon tasks that contain multiple subtasks or intermediate steps and can only provide sparse rewards at task completion point. This paper suggests to 1) use human priors to decompose a task and provide abstract demonstrations – the correct sequences of steps to guide exploration and learning, and 2) adjust the exploration parameters adaptively according to the online performances of the policy. The proposed ideas are implemented on three popular DRL algorithms, and experimental results on gridworld and manipulation tasks prove the concept and effectiveness of the proposed techniques.

Item Type:	Conference or Workshop Item (Paper)
Date Type:	Publication
Status:	Published
Schools:	Engineering
Subjects:	B Philosophy. Psychology. Religion > BF Psychology L Education > LB Theory and practice of education L Education > LC Special aspects of education L Education > LC Special aspects of education > LC5201 Education extension. Adult education. Continuing education T Technology > TA Engineering (General). Civil engineering (General)
Additional Information:	Contents are extended abstracts of papers, not full papers
Publisher:	Cardiff University Press
ISBN:	978-1-9116-5349-3
Funders:	China Scholarship Council
Date of First Compliant Deposit:	10 June 2024
Date of Acceptance:	2024
Last Modified:	29 Jul 2024 14:58
URI:	https://orca.cardiff.ac.uk/id/eprint/169685

Actions (repository staff only)

Edit Item