Research map / Reinforcement learning / Visual and representation RL
Learning to Search and Searching to Learn for Generalization in Planning
Michael Aichmüller, Yannik Hesse, Hector Geffner
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace real-time search and HER with WA* best-first search in the learning loop | GSP greedy policy · IPC 2023 blocksworld | coverage: 98% -> 100% | Improved |
| replace real-time search/HER with WA* search | GSP greedy policy · Sokoban 10x10 4 boxes | solved: 309 -> 681 | Improved |
| use learned Q as WA* heuristic instead of greedy execution | GSP · IPC 2023 Sokoban | coverage: 14% -> 32% | Improved |
| use learned Q as WA* heuristic instead of greedy execution | GSP · IPC 2023 Floortile | coverage: 20% -> 28% | Improved |
| use learned Q as WA* heuristic instead of greedy execution | GSP · IPC 2023 spanner | coverage: 100% -> 27% | Hurt |
| use learned Q as WA* heuristic instead of greedy execution | GSP · IPC 2023 blocksworld | coverage: 100% -> 79% | Hurt |
| use AlphaZero MCTS instead of best-first search as learning driver | relational GNN · IPC 2023 | coverage: blocksworld 31% vs 100% | Hurt |
| use GSP WA* search instead of WA* with MLP heuristics | GSP · Sokoban 10x10 4 boxes | average expansions: 3298 -> 207 | Improved |
| use GSP GBFS instead of GBFS baseline | GSP · Sokoban 10x10 4 boxes | expansions and solved: 5040 -> 564 expansions; 914 -> 998 solved | Improved |
| train GSP on sliding tile 24-puzzle | GSP · 24-puzzle | solved: 0 solved | Hurt |
| add solution lower-bound targets (ablation) | GSP · IPC 2023 planning domains | coverage | Improved |
| add dead-end supervision (ablation) | GSP · IPC 2023 planning domains | coverage | No clear effect |
Closest papers by meaning
- Efficient Lookahead Encoding and Abstracted Width for Learning General Policies in Classical Planning (NeurIPS 2026 · similarity 0.89)
- Learning Domain-Independent Heuristics for Grounded and Lifted Planning (AAAI 2024 · similarity 0.88)
- Horizon Generalization in Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Learning More Expressive General Policies for Classical Planning Domains (AAAI 2025 · similarity 0.87)
- Graph Neural Network Based Action Ranking for Planning (NeurIPS 2025 · similarity 0.87)
- Reinforcement learning with combinatorial actions for coupled restless bandits (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
