ASP-Based Abstractions for Reinforcement Learning
A new paper on arXiv (2605.31444) explores using Answer-Set Programming (ASP) to implement abstractions for Reinforcement Learning (RL). The CARCASS framework, originally developed by Martijn van Otterlo, uses logical representations to model Markov Decision Processes (MDPs) in first-order domains, previously implemented in Prolog. This research replaces Prolog with ASP, a fully declarative modelling language, to create powerful abstractions. The ASP-based implementation was evaluated in two domains: Blocks World and Minigrid. Results indicate that CARCASS with ASP offers a promising approach for constructing abstractions in RL, addressing challenges of large state spaces and generalisation.
Key facts
- Paper arXiv:2605.31444 proposes ASP-based abstractions for RL.
- CARCASS framework by Martijn van Otterlo models MDPs in first-order domains.
- Original CARCASS used Prolog; this work uses ASP.
- ASP is a fully declarative modelling language.
- Evaluation in Blocks World and Minigrid domains.
- Results show promise for constructing RL abstractions.
- Abstraction addresses large state spaces and generalisation.
- Relational Reinforcement Learning (RRL) reasons about objects and relations.
Entities
Artists
- Martijn van Otterlo