Research map / Graph learning / Graph algorithms and clustering
Efficient Lookahead Encoding and Abstracted Width for Learning General Policies in Classical Planning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an aggregated-delta encoding that jointly represents an IW(1) lookahead tree in one R-GNN pass, plus Abstracted IW(1) with type-based novelty checks, yielding state-of-the-art general policies on IPC 2023 learning track.
Problem
Learning general policies for classical planning that scale to very large instances (thousands of objects) where IW-lookahead policies are computationally expensive and expressivity-limited by per-transition encodings.
Approach
Encode the whole IW(1) lookahead tree jointly as relational differences anchored to successor state-nodes so an R-GNN scores all successors in one pass, and replace atom-level novelty with abstracted novelty (all but one argument replaced by its type) so novelty scales with object count rather than atom count.
Key results
- AIW-AD achieves 668/900 (74%) coverage on IPC 2023 versus 586/900 for LAMA (65%) and 529/900 for IW-AD (59%).
- AIW-AD fully solves Satellite (100%) and reaches 100% on the broader benchmark's Grid and Satellite, with 1570/1576 total coverage.
- The AD encoding reduces VRAM for IW-policy models from over 24GB to at most 4GB.
Surprising findings
- Abstracted IW(1) outperforms full IW(1) lookahead on IPC 2023 (668 vs 529 coverage for IW-AD), despite more aggressive pruning.
- BAIW (base-type abstraction) fails on Logistics with only 48% coverage due to over-pruning.
Limitations
- AIW loses completeness guarantees of IW(k) for width-k problems.
- Sokoban and Floortile remain largely unsolved; Childsnack scaling limited by enumerating successors.
- Evaluation relies on greedy deterministic policies with a bounded number of choices and timeouts.
Experimental setup
Datasets: IPC 2023 learning track hard-to-learn domains from prior work (beyond C2)
Benchmarks: IPC 2023 learning track Generalized planning benchmark domains beyond C2 (Grid, Logistics, Satellite, Rovers, etc.)
Metrics: coverage test-time solve rate
Baselines: LAMA Distincter WL-GOOSE IW-policies with External encoding Flat AA policies
Builds on: IW-policies Iterated Width lifted hindsight experience replay R-GNN planning policies
Closest papers by meaning
- Learning More Expressive General Policies for Classical Planning Domains (AAAI 2025 · similarity 0.90)
- State Encodings for GNN-Based Lifted Planners (AAAI 2025 · similarity 0.87)
- Learning Domain-Independent Heuristics for Grounded and Lifted Planning (AAAI 2024 · similarity 0.86)
- Graph Neural Network Based Action Ranking for Planning (NeurIPS 2025 · similarity 0.86)
- Graph Learning for Numeric Planning (NeurIPS 2024 · similarity 0.85)
- System 1.x: Learning to Balance Fast and Slow Planning with Language Models (ICLR 2025 · similarity 0.85)
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.