Research map / Learning theory and algorithms / Combinatorial and search optimization

Optimizing the Optimization of Planning Domains by Automatic Action Schema Splitting

AAAI 2024 · Learning theory and algorithms › Combinatorial and search optimization · approach: Classical and HTN planning · 1 citations (OpenAlex)

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.

Improves the automatic action schema splitting of Areces et al. (2014) for grounding-based classical planners by adding invariant-based dependency graph edges, a priority-list objective function, and a Monte Carlo search for the split.

Problem

Grounding-based classical planners stall in preprocessing when action schemas have impractically many ground instances; existing automatic action splitting is problem-sensitive and hyperparameter-dependent.

Approach

Encode parameter dependencies from monotonicity and schematic invariants into the dependency graph, replace the γ-weighted objective with a prioritized tuple of features governed by a ground-action threshold, and search splits with a pruned Monte Carlo random walk.

Key results

• On hard-to-ground domains, the full approach (EAS) raises LAMA coverage from 262 total (Base) to 240 across hard-to-ground set, and BFWS from 230 to 258 in Table 1 ablations.
• On the broader benchmark (Table 2), EAS gives LAMA 575 solved vs 742 without splitting-total pattern shows limited overhead; Maidu 640 vs 572 with EAS, and BFWS 531 vs 531-range, with EAS competitive with Powerlifted (482 total).
• Areces et al. splitting with gamma=0 adds considerable overhead on IPC domains, while EAS adds minimal overhead and improves coverage on some domains such as Slitherlink.

Measured effects reported in the paper

Surprising findings

• The Areces et al. approach with gamma=0 reduces LAMA coverage on many non-hard-to-ground IPC domains, showing overhead from unnecessary splitting.

Limitations

• Does not support universally quantified effects or conditional effects in the implementation.
• Monte Carlo search is time-limited and heuristic; split optimization is NP-complete.
• Threshold omega=10^6/|A| is a rough estimate of planner capacity.

Experimental setup

Datasets: Pipesworld-Tankage Genome Edit Distance Organic Synthesis (Alkene, MIT, Original) IPC 2014, 2018, 2023 domains

Benchmarks: Pipesworld-Tankage Genome Edit Distance Organic Synthesis IPC 2014/2018/2023 domains

Metrics: coverage (number of solved instances)

Baselines: Areces et al. splitting with gamma=0 Areces et al. splitting with gamma=0.8 Powerlifted no splitting

Builds on: Areces et al. 2014 action splitting Helmert monotonicity invariants Rintanen schematic invariant synthesis

Closest papers by meaning

• An Effective Polynomial Technique for Compiling Conditional Effects Away (AAAI 2024 · similarity 0.86)
• Automatic Selection of Macro-Events for Heuristic-Search Temporal Planning (AAAI 2025 · similarity 0.85)
• Learning Planning Domains from Non-redundant Fully-Observed Traces: Theoretical Foundations and Complexity Analysis (AAAI 2024 · similarity 0.85)
• Differentiable Learning of Lifted Action Schemas for Classical Planning (NeurIPS 2026 · similarity 0.84)
• Planning Anything with Rigor: General-Purpose Zero-Shot Planning with LLM-based Formalized Programming (ICLR 2025 · similarity 0.84)
• Dealing with Numeric and Metric Time Constraints in PDDL3 via Compilation to Numeric Planning (AAAI 2024 · similarity 0.83)