Research map / Graph learning / Graph algorithms and clustering
Graph Learning for Numeric 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 GOOSE, a framework that encodes numeric planning tasks as graphs (νILG), extends WL kernel to continuous and categorical attributes (CCWL), and trains linear or GNN heuristics with a new sparse ranking formulation for greedy best-first search.
Problem
Learning heuristic functions for numeric planning, which requires reasoning over both logical and arithmetic state variables, in a data-efficient and interpretable way.
Approach
A numeric-aware graph encoding (νILG) and a CCWL graph kernel that handles continuous node attributes via order-invariant pooling per colour, combined with a sparse LP-based ranking formulation that avoids misclassifying siblings with equal h* values.
Key results
- The CCWL ranking model (hCCWLF rank) reaches total coverage 445 at L=1 (476 best over learners) versus 369 for the best domain-independent satisficing planner M(3h||3n).
- CCWL models are at least an order of magnitude faster than GNN models for heuristic evaluation (Fig. 8).
- Learned CCWL ranking heuristic achieves 90/90 on Childsnack, where other learners and planners fail to generalise.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CCWL features rather than GNN representations for cost-to-go estimation | Numeric planning GBFS · All domains | coverage: 319 -> 379 | Helped |
| Use CCWL features rather than GNN representations for ranking estimation | Numeric planning GBFS · All domains | coverage: 326 -> 445 | Helped |
| Use ranking instead of cost-to-go estimation with CCWL | CCWL numeric planning model · All domains | coverage: 379 -> 445 | Helped |
| Use ranking instead of cost-to-go estimation with GNN | GNN numeric planning model · All domains | coverage: 319 -> 326 | Helped |
| Use CCWL ranking rather than the strongest domain-independent baseline | Numeric planning GBFS · Childsnack | coverage: 53 -> 90 | Helped |
| Use CCWL ranking rather than the strongest domain-independent baseline | Numeric planning GBFS · Rovers | coverage: 34 -> 23 | Hurt |
| Use CCWL ranking rather than the strongest domain-independent baseline | Numeric planning GBFS · Satellite | coverage: 38 -> 16 | Hurt |
| Use CCWL ranking rather than the strongest domain-independent baseline | Numeric planning GBFS · Spanner | coverage: 76 -> 90 | Helped |
| Set GNN cost-to-go depth to L=0 rather than L=1 (ablation) | hGNN cost · All domains | coverage: 319 -> 336 | Helped |
| Set GNN ranking depth to L=0 rather than L=1 (ablation) | hGNN rank · All domains | coverage: 326 -> 338 | Helped |
| Set CCWL cost-to-go iterations to L=0 rather than L=1 (ablation) | hCCWLF cost · All domains | coverage: 379 -> 407 | Helped |
| Set CCWL ranking iterations to L=1 rather than L=0 (ablation) | hCCWLF rank · All domains | coverage: 386 -> 445 | Helped |
Surprising findings
- For most domains L=0 or L=1 gives the best coverage; increasing CCWL/GNN depth rarely helps, suggesting heuristic evaluation time matters more than expressivity.
- Non-linear kernels generalise worse than linear models in this setting due to overfitting to the range of training targets.
Limitations
- Assumes problems are explicitly representable in PDDL, so training data can be generated with domain-independent planners.
- Expressivity is limited for domains requiring more complex reasoning, e.g., Rovers, Satellite, and Transport, where learners fall behind.
- Results are from single runs with no error bars.
Experimental setup
Datasets: IPC 2023 Learning Track domains (converted to numeric)
Benchmarks: Capacity-constrained Blocksworld Numeric Childsnack Numeric Ferry Numeric Miconic Numeric Rovers Numeric Satellite Numeric Spanner Numeric Transport
Metrics: coverage expansions plan length heuristic evaluation time
Baselines: hLMCUT Metric-FF hADD hMRP hMRP+hj M(3h||3n) GNN cost-to-go GNN ranking
Builds on: Weisfeiler-Lehman kernel Relational Graph Convolution Network Metric-FF Learning for planning ranking formulation (Garrett et al. 2016)
Closest papers by meaning
- Symbolic Numeric Planning with Patterns (AAAI 2024 · similarity 0.88)
- PDBs Go Numeric: Pattern-Database Heuristics for Simple Numeric Planning (AAAI 2025 · similarity 0.87)
- Learning Domain-Independent Heuristics for Grounded and Lifted Planning (AAAI 2024 · similarity 0.87)
- State Encodings for GNN-Based Lifted Planners (AAAI 2025 · similarity 0.86)
- Towards Practical Classical Planning Compilations of Numeric Planning (AAAI 2025 · similarity 0.86)
- Graph Neural Network Based Action Ranking for Planning (NeurIPS 2025 · similarity 0.86)
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.