Hierarchical Policy Learning for Offline Reinforcement Learning of Complex Tasks
Abstract
Despite the recent remarkable progress of offline reinforcement learning (RL) in simple environments with a single task, it still lacks generalization to new tasks or states, resulting in insufficient performance in complex environments. Recent works attempt to address this challenge, but they either rely on specific information about environment dynamics or require tremendous amounts of demonstration data, which are generally unavailable. In this project, we propose a novel offline RL framework that addresses this challenge by utilizing hierarchical policy design to solve complex tasks. Unlike the existing methods, our framework does not require environment-specific information and can be generalized to new tasks and unseen states with minimal expert demonstration.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 16, 2024
- Source ID
- FA23862314047
Entities
People
- Hyun Oh Song
Organizations
- Air Force Office of Scientific Research
- Seoul National University
- United States Air Force