GONDOR: Memory-Efficient Greedy Best-First Search for Planning
A team of researchers has unveiled GONDOR (Greedy Online Navigation with Dynamic Outpost-based Re-search), an enhanced version of Greedy Best-First Search (GBFS) that is tailored for edge devices with limited memory. This innovative approach periodically compresses the search tree while maintaining a minimal collection of anchor states, allowing for path reconstruction through re-searching these states once the goal is reached. The study investigates various algorithmic variants based on different outpost selection strategies and examines the application of Bloom filters for efficient duplicate detection in the closed list. Results from experiments in numeric planning domains and various heuristic setups highlight its effectiveness. This research is documented in arXiv:2605.28454.
Key facts
- GONDOR extends Greedy Best-First Search for low-memory environments.
- It compresses the search tree periodically, retaining sparse anchor states.
- Path reconstruction occurs by re-searching between anchor states.
- Bloom filters are used for duplicate detection.
- Multiple outpost selection policies are discussed.
- Experiments cover numeric planning domains and heuristic configurations.
- The paper is available on arXiv with ID 2605.28454.
- Target application includes planning on edge devices.
Entities
Institutions
- arXiv