Towards Practical Classical Planning Compilations of Numeric Planning

Luigi Bonassi, Francesco Percassi, Enrico ScalaAAAI 2025 · DOI

TL;DR

This paper explores practical ways to translate complex numeric planning problems into a format that standard classical AI planners can understand. They introduce three compilation methods, with one using axioms proving surprisingly effective, allowing classical planners to compete with specialized numeric planners on certain problem types.

What Problem Does This Paper Address?

Artificial intelligence planning aims to find sequences of actions to achieve goals. While classical planning deals with discrete states and actions, many real-world scenarios involve continuous or quantifiable numeric variables (e.g., fuel levels, temperatures). Numeric planning handles these but is generally undecidable. A known theoretical solution is to compile finite-domain numeric planning problems into classical planning, preserving plan length. However, the practical utility of these compilations has been questionable. This paper investigates whether these theoretical compilations can be made practical, allowing off-the-shelf classical planners to tackle numeric planning tasks effectively.

Methods

The authors propose three compilation methods:

  • ONE-HOT: This method directly translates each possible value of a numeric variable into a separate Boolean variable. This mirrors a common manual approach but can lead to a combinatorial explosion of variables for larger numeric domains.
  • BLAST: This approach uses a binary encoding (two's complement) for numeric variables, representing them with a fixed number of bits. It translates numeric conditions and effects into Boolean logic, including handling arithmetic operations like addition. However, the encoding of arithmetic operations can become complex and computationally expensive as the number of bits increases.
  • BLASTx: This is an extension of BLAST that leverages axioms. Instead of encoding complex arithmetic operations directly into conditional effects, BLASTx uses axioms to define derived variables that represent intermediate calculation steps. This makes the action models more compact and potentially more efficient for classical planners that support axioms.

Datasets

The experiments were conducted on domains from the latest International Planning Competition (IPC), specifically the Simple Numeric Planning (SNP) domains. They excluded the "SETTLERS" domain due to its specific characteristics not fitting their current framework. They also categorized domains as "strongly numeric" (SN) or "mildly numeric" (MN) based on the ratio of numeric to Boolean variables.

Key Insights and Ideas

  • The "overly complex" can be practical: Surprisingly, the theoretically complex logarithmic encoding (BLASTx), when augmented with axioms, proved to be quite effective, outperforming simpler encodings and even specialized numeric planners in some cases.
  • Axioms are powerful for numeric planning: The success of BLASTx highlights the utility of axioms in classical planning for compactly representing and reasoning about numeric computations. This suggests that a hybrid approach, combining classical planning with axiom-based reasoning, can be very powerful.
  • Numericity matters: The paper identifies a clear distinction in performance between strongly numeric (SN) and mildly numeric (MN) domains. Classical planners with compilations perform better on MN domains, suggesting that the choice of planner (classical vs. specialized numeric) should depend on the problem's "numericity."
  • Compilation overhead is real: While compilations enable classical planners to tackle numeric problems, they introduce overhead. This makes them slower than native numeric planners in terms of raw speed, but the ability to leverage powerful classical planners can still be advantageous.

Key Results

  • BLASTx outperforms other compilations: In experimental evaluations, BLASTx achieved the highest coverage (number of solved instances) across various domains, outperforming ONE-HOT and BLAST.
  • Competitive with state-of-the-art numeric planners: For mildly numeric (MN) domains, BLASTx was competitive with, and at times exceeded, the performance of specialized numeric planners like M(3h||3n) and NLM-CUTPLAN.
  • BLASTx excels in specific domains: Notably, BLASTx performed particularly well in the "DELIVERY" domain and also showed strong performance in "SETTLERS" (using its SNP formulation).
  • Classical planners with compilation are slower but effective: While compilation-based approaches were slower than native numeric planners due to compilation time, LAMA-FIRST (a classical planner) with BLASTx was a strong contender, especially in MN domains.

Limitations and Future Work

  • Domain limitations: The current work focuses on Simple Numeric Planning (SNP) and requires numeric variables to be initialized in the initial state. Tasks where numeric variables are undefined and initialized by actions are not yet supported.
  • K selection: The choice of the number of bits (K) for binary encodings and the domain size (D) for ONE-HOT can impact performance and requires careful selection based on the problem instance.
  • Future Work: The authors suggest several avenues for future research:
  • Developing novel numeric heuristics that leverage the compiled representations.
  • Automating the extraction of bounds for numeric variables to improve compilation.
  • Exploring practical compilations for more expressive fragments of numeric planning beyond SNP.

Summary updated

BibTeX
@inproceedings{bonassi2025towards,
  author = {Luigi Bonassi and Francesco Percassi and Enrico Scala},
  title = {{Towards Practical Classical Planning Compilations of Numeric Planning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i25.34847},
}
Figure