[Paper] The Energy Impact of Domain Model Design in Classical Planning
Published: (January 29, 2026 at 11:46 AM EST)
3 min read
Source: arXiv
Source: arXiv - 2601.21967v1
Overview
The paper The Energy Impact of Domain Model Design in Classical Planning examines a rarely‑studied angle of AI planning: how the way we write the planning domain model influences the energy a planner consumes. While most research optimises for speed or solution quality, the authors show that even modest tweaks to the model can lead to measurable energy savings—sometimes without any runtime benefit.
Key Contributions
- Energy‑focused analysis of classical planning, extending the “Green AI” agenda to the planning community.
- Domain model configuration framework that systematically varies model attributes (e.g., predicate ordering, action arity, presence of dead‑ends).
- Empirical study across 5 benchmark domains and 5 state‑of‑the‑art planners, evaluating 32 model variants per benchmark (≈ 800 planner runs).
- Evidence that energy consumption is not always proportional to runtime, highlighting a new performance trade‑off.
- Guidelines for energy‑aware domain engineering, offering concrete recommendations for practitioners.
Methodology
- Domain Variants Generation – The authors built a lightweight tool that takes a canonical PDDL (Planning Domain Definition Language) file and produces controlled variants by:
- Re‑ordering predicates and actions.
- Changing the arity (number of arguments) of actions.
- Injecting or removing dead‑end states (states from which no goal is reachable).
- Benchmarks & Planners – Five widely used planning domains (e.g., Blocks‑World, Logistics, Satellite) were paired with five planners representing different algorithmic families (heuristic search, SAT‑based, etc.).
- Measurement Setup – Each planner‑domain pair was executed on identical hardware while logging:
- Runtime (seconds).
- Energy consumption (via CPU power counters, measured in Joules).
- Statistical Analysis – The authors applied ANOVA and pairwise comparisons to isolate the effect of each model attribute on energy and runtime, controlling for planner and domain differences.
Results & Findings
| Aspect | What the data showed |
|---|---|
| Action arity | Increasing the number of parameters per action raised energy use by up to 18 %, even when runtime stayed flat. |
| Predicate ordering | Simple re‑ordering could cut energy by 5‑12 % for some planners, with negligible runtime impact. |
| Dead‑end states | Adding unreachable states inflated energy consumption dramatically (up to 30 %) while sometimes reducing runtime due to early pruning. |
| Planner‑specific trends | SAT‑based planners were more sensitive to predicate ordering, whereas heuristic search planners reacted more to dead‑ends. |
| Energy vs. Runtime | Correlation coefficients ranged from 0.2 (weak) to 0.7 (moderate) across planner‑domain combos, confirming that faster does not always mean greener. |
Overall, the study demonstrates that domain‑level design choices are a low‑cost lever for reducing the carbon footprint of planning systems.
Practical Implications
- Energy‑aware domain engineering: Teams building planning‑based services (e.g., logistics optimization, robot task sequencing) can adopt the paper’s configuration guidelines to shave off energy costs without rewriting the planner itself.
- Cost savings in cloud environments: Since many providers charge by CPU time and energy usage, even a 10 % reduction translates into lower operational expenses at scale.
- Sustainable AI compliance: Companies aiming for ESG (Environmental, Social, Governance) targets can report concrete energy‑reduction metrics derived from domain model refactoring.
- Tooling opportunities: The configuration framework could be integrated into CI pipelines to automatically test new domain revisions for energy impact, similar to performance regression testing.
- Algorithm‑planner selection: Knowing that certain planners are more energy‑sensitive to specific model traits helps architects pick the right planner‑model combo for green deployments.
Limitations & Future Work
- Hardware specificity – Energy measurements were taken on a single CPU architecture; results may differ on GPUs, ARM chips, or heterogeneous clusters.
- Scope of domains – Only five classic benchmarks were examined; industrial domains with richer predicates could exhibit different sensitivities.
- Static analysis – The study focuses on static model properties; dynamic runtime adaptations (e.g., on‑the‑fly pruning) were not explored.
- Future directions suggested by the authors include: extending the framework to probabilistic and temporal planning, evaluating energy on distributed planning pipelines, and developing automated “energy‑aware” domain refactoring tools that suggest optimal model layouts.
Authors
- Ilche Georgievski
- Serhat Tekin
- Marco Aiello
Paper Information
- arXiv ID: 2601.21967v1
- Categories: cs.AI, cs.SE
- Published: January 29, 2026
- PDF: Download PDF