Abstraction for Offline Goal-Conditioned Reinforcement Learning
A new paper on arXiv proposes that hierarchical policies in offline goal-conditioned reinforcement learning (GCRL) enable not only temporal abstraction but also absolute abstraction, allowing agents to reuse experience across similar state-space contexts. The authors introduce relativised options and distinct representations for different hierarchy levels, along with two simple algorithms for learning these options and abstracting from the absolute frame of reference. Experiments show that these inductive biases significantly improve performance in offline GCRL.
Key facts
- Paper titled 'Abstraction for Offline Goal-Conditioned Reinforcement Learning'
- Submitted to arXiv on May 27, 2025
- Focuses on Markov Decision Processes (MDPs) in Goal-Conditioned Reinforcement Learning (GCRL)
- Demonstrates that hierarchy enables absolute abstraction beyond temporal abstraction
- Introduces relativised options and distinct representations for different hierarchy levels
- Proposes two simple algorithms for learning relativised options and abstracting from absolute frame
- Experiments show significant performance improvement in offline GCRL
- arXiv ID: 2605.22711
Entities
Institutions
- arXiv