Towards efficient deep reinforcement learning in complex and dynamic tasks /
Files
(Published version)
Date
2024
Authors
Chen, Bin
Editors
Advisors
Journal Title
Journal ISSN
Volume Title
Type:
thesis
Citation
Statement of Responsibility
Conference Name
Abstract
As an essential means of achieving generalised AI, deep reinforcement learning (DRL) achieved beyond human performance in many decision-making environments, such as competitive games and navigation tasks, by combining the representational power of deep learning with the decision-making power of reinforcement learning. However, the DRL faces sampling efficiency problems with sparse rewards in complex and dynamic environments because it requires massive interactions with environments under partial observability. Due to this inefficient data sampling issue, it isn’t easy to train a DRL agent to achieve optimal performance even after millions of steps. In addition to the instability of the complex and dynamic environment, the problem of training efficiency due to the increasing agent number is even more severe in multiagent DRL (MARL) learning. This thesis focuses on sparse reward issues in complex tasks, sampling efficiency issues in dynamic tasks and state space explosion issues in large-scale multiagent tasks.
School/Discipline
University of South Australia. UniSA STEM.
UniSA STEM
UniSA STEM
Dissertation Note
Thesis (PhD(Computer and Information Science))--University of South Australia, 2024.
Provenance
Copyright 2024 Bin Chen.
Description
1 ethesis (xii, 150 pages) :
colour illustrations.
Includes bibliographical references (pages 128-142)
colour illustrations.
Includes bibliographical references (pages 128-142)
Access Status
506 0#$fstar $2Unrestricted online access