Neurosymbolic HRL with Incremental Knowledge Improves Sample Efficiency
A new research paper on arXiv (2608.02993) proposes a neurosymbolic approach to Hierarchical Reinforcement Learning (HRL) that incorporates incremental knowledge to improve sample efficiency in sparse-reward, long-horizon environments. The method, termed Incremental Knowledge (InK), uses symbolic high-level components for planning (e.g., with D* algorithm) on an updatable representation of current knowledge, while low-level goal-conditioned neural modules learn motion primitives via reward shaping. The authors argue that standard HRL encodes knowledge in a fixed, non-updatable form, limiting reasoning with incremental knowledge and leading to poor sample efficiency. Experiments on navigation tasks demonstrate the effectiveness of the proposed approach. The paper is announced as a new submission and is available at https://arxiv.org/abs/2608.02993.
Key facts
- Paper arXiv:2608.02993 proposes neurosymbolic HRL with Incremental Knowledge (InK).
- InK uses symbolic planning (e.g., D*) on an updatable knowledge representation.
- Low-level modules learn motion primitives through experience and reward shaping.
- Standard HRL uses fixed, non-updatable knowledge, limiting sample efficiency.
- Experiments on navigation tasks show improved sample efficiency.
- The paper is a new announcement on arXiv.
- The approach addresses sparse rewards and long-horizon reasoning challenges.
- The paper is available at https://arxiv.org/abs/2608.02993.
Entities
Institutions
- arXiv