Technical Summary: The Energy Impact of Domain Model Design in Classical Planning

Problem Statement

While AI research has traditionally prioritized algorithmic performance metrics such as accuracy and runtime, the emerging paradigm of "Green AI" challenges this by recognizing energy consumption as a critical dimension. Despite the high computational demands of automated planning, its energy efficiency has received little attention. This gap is particularly significant given the modular architecture of classical planning, where planning algorithms are decoupled from application-specific knowledge encoded in domain models (typically in PDDL).

Domain modelers possess considerable freedom in specification, making design decisions regarding action granularity, predicate structure, ordering, and encoding style. While prior work has established that these syntactic and semantic choices dramatically influence planner runtime and solution quality, the impact of such domain model configurations on energy consumption remains unexplored. This paper addresses this gap by investigating whether and how specific domain model design decisions affect the energy efficiency of classical planners.

Methodology

The authors introduce a Domain Model Configuration Framework designed to systematically vary and test aspects of domain models to understand their impact on planner behavior. The framework organizes domain features into three categories:

1. Syntactic Structure Choices (SSC): Reordering domain elements (e.g., predicate declarations, action definitions, preconditions, and effects) based on criteria such as usage frequency, alphabetical order, or literal counts. These changes preserve semantics but alter syntax.
2. Modelling Redundancy Choices (MRC): Introducing semantically inert constructs, such as dummy objects, redundant predicates, increased arity, or duplicated inapplicable actions, to inflate representational overhead without changing solvability.
3. Task Design Choices (TDC): Modifying task solvability or difficulty, such as introducing dead-end states or removing goal-achieving effects, which alter the solution space and search structure.

Experimental Setup:

• Planners: Five state-of-the-art planners were evaluated: three based on the Fast Downward (FD) framework (FDSSA, CA, DALAIA) and two based on LAPKT (ANST, FBNS).
• Domains: Five benchmark domains from the International Planning Competition (IPC): Barman, Blocks World, Gripper, Thoughtful, and Ricochet Robots.
• Variants: For each domain, 32 variants were generated (1 original + 20 SSC, 7 MRC, and 4 TDC variants).
• Execution: Experiments were conducted on a dedicated machine (Intel Core i7-1065G7, 32 GB RAM) running Ubuntu 24.04 LTS in a minimal environment. Energy consumption was measured at the CPU package level using Intel RAPL counters via pyRAPL within isolated containers.
• Constraints: Runs were restricted to a single CPU core, 8 GB memory, and a 5-minute wall-clock timeout. Each configuration was executed 30 times to ensure statistical robustness. Only runs terminating within the timeout were analyzed.

Key Contributions

1. Framework Development: A systematic framework for characterizing and transforming domain model configurations to produce controlled variants for empirical evaluation.
2. Empirical Evidence: The first empirical study linking domain model configuration directly to energy consumption in AI planning.
3. Cross-Planner Analysis: A comprehensive evaluation across two dominant planning frameworks (FD and LAPKT) and five diverse benchmark domains.

Results

The study reveals that energy consumption in classical planners is generally stable under minor representational changes but exhibits significant sensitivity to specific modeling and task-level mechanisms. Crucially, energy consumption does not always correlate with runtime.

• Syntactic Structure Choices (SSC):

  • FD-based planners showed remarkable stability, with mean energy deviations below 1% and near-perfect Pearson correlations across variants.
  • LAPKT-based planners displayed domain-dependent sensitivity. In domains like Barman and Thoughtful, reordering actions or effects led to large energy deviations (e.g., a 3.8-fold difference in Barman) and correlations dropping to near-zero or negative values. In Thoughtful, action ordering variants caused immediate timeouts, while predicate ordering remained stable.

• Modelling Redundancy Choices (MRC):

  • Most redundancy mechanisms produced modest effects. However, MRC-ROA (adding an unused dummy parameter to all action definitions) consistently increased energy consumption across nearly all planner-domain combinations.
  • The impact of MRC-ROA ranged from a 2x increase to over 12x the baseline energy, with high variability (standard deviations up to 1400 J). In Blocks World, energy increased more than 11-fold, and some instances that solved in the original domain failed under this configuration.
  • FD-based planners were generally more resilient to redundancy than LAPKT-based planners, though MRC-ROA remained the primary source of inefficiency for both.

• Task Design Choices (TDC):

  • Effects were highly domain- and planner-dependent. While some variants remained close to baseline, others caused extreme deviations.
  • In Blocks World, TDC-DEF (removing a goal-achieving effect) increased energy consumption more than 30-fold for FDSSA. Conversely, in Thoughtful, the same variant reduced energy from 260 J to 168 J.
  • Dead-end mechanisms (TDC-RPD, TDC-COP) often amplified grounding complexity or altered search trajectories, leading to significant energy spikes in Barman and Ricochet Robots for specific planners.

Significance and Claims

The paper claims that energy efficiency in classical planning is a structural property arising from the interaction among domain model configuration, planner design, and domain topology, rather than a function of runtime alone.

• Architectural Vulnerabilities: The results highlight architecture-specific vulnerabilities. FD-based planners, which front-load computation through grounding and preprocessing, are particularly sensitive to redundancy and dead ends. LAPKT-based planners are more resilient to redundancy but can be sensitive to structural ordering due to order-dependent data structures in novelty evaluation.
• Domain Topology: The complexity of the domain's relational structure (e.g., Blocks World and Barman) magnifies the impact of modeling choices, whereas compact domains (e.g., Gripper, Ricochet Robots) maintain predictable energy use.
• Practical Implications: For domain modelers, representational efficiency is critical; avoiding redundant action parameters and carefully validating task design can substantially reduce energy consumption without changing the planner. For planner developers, the findings suggest opportunities for energy-aware preprocessing and the use of order-independent data structures.

The authors conclude that energy emerges as a performance metric that exposes inefficiencies invisible through runtime analysis alone. They propose that the Domain Model Configuration Framework provides a systematic means to characterize these sensitivities, supporting future research into predictive models for energy costs based on static domain properties.