Research map / Graph learning / Graph algorithms and clustering
Graph Neural Network Based Action Ranking for 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 GABAR, a GNN with GRU decoder over an action-centric graph that learns to rank grounded actions for classical planning, trained on small instances and tested on much larger ones.
Problem
Learning generalizable relational policies for classical planning that scale to much larger instances than those used for training, where planners are too slow and value-function learning is hard to generalize.
Approach
Instead of learning a globally consistent value function, learn a locally consistent ranking over applicable actions using an action-centric graph (action nodes linked to objects) and a GRU-based autoregressive decoder that selects an action schema and then its parameters conditioned on prior choices.
Key results
- GABAR reaches 95.5% coverage on easy, 92.2% on medium and 89.2% on hard test problems averaged across eight domains.
- On hard instances GPL drops to 6.5% coverage and GRAPL to 22.1%, while ASNets reaches 48.5%.
- OpenAI-O3 and Gemini-2.5-Pro reach only 0.4% and 1.5% average coverage on hard problems.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add explicit action nodes and action-object edges (ablation) | GABAR · Combined Hard | coverage: 7.4 -> 89.2 | Helped |
| use conditional decoding for action parameters (ablation) | GABAR · Combined Hard | coverage: 60.0 -> 89.2 | Helped |
| rank actions instead of learning a value function (ablation) | GABAR · Combined Hard | coverage: 12.1 -> 89.2 | Helped |
| add the global node (ablation) | GABAR · Combined Hard | coverage: 42.5 -> 89.2 | Helped |
| include both action representation and conditional decoding (ablation) | GABAR · Combined Hard | coverage: 2 -> 89.2 | Helped |
Surprising findings
- Value-function-based GPL and the GABAR-RANK ablation fall to near-zero coverage on hard Logistics and Visitall instances, while the action-ranking variant remains robust.
- Removing action nodes (GABAR-ACT) drops easy-set coverage from 95.5% to 33.5%.
Limitations
- Graph representation grows quickly with high-arity actions or predicates.
- Greedy/beam decoding and a 1000-step limit with cycle-avoidance rely on visited-state history.
- Reference plans come from a satisficing planner (FD lama-first) rather than optimal plans.
Experimental setup
Datasets: PDDL benchmark domains generated with openly available PDDL generators (Blocks, Gripper, Miconic, Spanner, Logistics, Ro
Benchmarks: Blocks Gripper Miconic Spanner Logistics Rovers Visitall Grid
Metrics: coverage within 1000 steps plan quality ratio vs Fast Downward (fd-lama-first) plan length
Baselines: GPL ASNets GRAPL OpenAI-O3 Gemini-2.5-Pro GABAR-ACT GABAR-CD GABAR-RANK
Builds on: GPL ASNets GRAPL
Closest papers by meaning
- State Encodings for GNN-Based Lifted Planners (AAAI 2025 · similarity 0.89)
- Learning More Expressive General Policies for Classical Planning Domains (AAAI 2025 · similarity 0.88)
- Learning Domain-Independent Heuristics for Grounded and Lifted Planning (AAAI 2024 · similarity 0.87)
- Differentiable Learning of Lifted Action Schemas for Classical Planning (NeurIPS 2026 · similarity 0.86)
- Efficient Lookahead Encoding and Abstracted Width for Learning General Policies in Classical Planning (NeurIPS 2026 · similarity 0.86)
- Graph Learning for Numeric Planning (NeurIPS 2024 · 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.