Research map / Learning theory and algorithms / Combinatorial and search optimization
Differentiable Learning of Lifted Action Schemas for 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 DIAS, a GNN-plus-Sinkhorn differentiable architecture that learns lifted STRIPS action schemas (including unobserved arguments) from fully observed state transitions by gradient descent, reaching perfect recovery on 8 of 13 IPC domains in the action-names-only setting.
Problem
Learning lifted STRIPS/STRIPS+ action schemas from traces with fully observed states where action arguments are unobserved (or only action names are given), a combinatorially hard isomorphism-like problem usually handled by SAT calls.
Approach
Jointly learn argument selection via an RGCN over state graphs with slot queries and differentiable Sinkhorn assignment, and learnable soft STRIPS precondition/add/delete probabilities over slots, trained end-to-end with BCE losses plus a PCGrad-projected auxiliary loss for minimal effects and maximal preconditions.
Key results
- With full STRIPS action labels, all 13 domains are learned perfectly in all 10 runs.
- With action names only, 8 of 13 domains are learned perfectly in every run; Satellite fails completely (0 of 10 sound-and-complete runs).
- Under observation noise with action names only, most domains degrade at 2 flips per transition and no domain is learned correctly at 8 flips.
Surprising findings
- The MLP-effects variant (RIM-like) learns STRIPS-like effects reliably in several domains but fails completely on Delivery, Hanoi, Satellite, Sokoban and Sokoban-pull.
- Satellite with action names only fails in all runs, mainly because 75% of take_image transitions leave the state unchanged.
Limitations
- Sokoban, Hanoi, Driverlog and Logistics fail in some runs due to dead ends, poor random-walk coverage, or preconditions on variables absent from effects.
- Performance degrades sharply under observation noise and the approach is limited to unary and binary predicates.
- State-dependent precondition MLP is not used because it admits a trivial solution without negative samples.
Experimental setup
Datasets: PDDL Generators IPC domains (13 domains)
Benchmarks: 13 IPC domains (Blocks-3, Delivery, Driverlog, Gripper, Hanoi, Logistics, Miconic, N-Puzzle, Satellite, Sokoban, Sokoban
Metrics: precision recall N_s N_c N_sc
Baselines: L1 (Balyo 2024) RIM-like MLP-effects variant
Builds on: STRIPS+ RGCN Sinkhorn algorithm PCGrad RIM
Closest papers by meaning
- Learning Planning Domains from Non-redundant Fully-Observed Traces: Theoretical Foundations and Complexity Analysis (AAAI 2024 · similarity 0.89)
- PcLast: Discovering Plannable Continuous Latent States (ICML 2024 · similarity 0.87)
- Learning Planning Abstractions from Language (ICLR 2024 · similarity 0.86)
- Graph Neural Network Based Action Ranking for Planning (NeurIPS 2025 · similarity 0.86)
- PlanLLM: Video Procedure Planning with Refinable Large Language Models (AAAI 2025 · similarity 0.86)
- Embodied Task Planning via Graph-Informed Action Generation with Large Language Models (ICML 2026 · 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.