Macro-FF: Improving AI Planning with Automatically Learned Macro-Operators (1109.2154v1)

Abstract: Despite recent progress in AI planning, many benchmarks remain challenging for current planners. In many domains, the performance of a planner can greatly be improved by discovering and exploiting information about the domain structure that is not explicitly encoded in the initial PDDL formulation. In this paper we present and compare two automated methods that learn relevant information from previous experience in a domain and use it to solve new problem instances. Our methods share a common four-step strategy. First, a domain is analyzed and structural information is extracted, then macro-operators are generated based on the previously discovered structure. A filtering and ranking procedure selects the most useful macro-operators. Finally, the selected macros are used to speed up future searches. We have successfully used such an approach in the fourth international planning competition IPC-4. Our system, Macro-FF, extends Hoffmanns state-of-the-art planner FF 2.3 with support for two kinds of macro-operators, and with engineering enhancements. We demonstrate the effectiveness of our ideas on benchmarks from international planning competitions. Our results indicate a large reduction in search effort in those complex domains where structural information can be inferred.

• The paper introduces two novel strategies, CA-ED and SOL-EP, that automatically generate macro-operators to enhance AI planning efficiency.
• It details a four-stage framework including domain analysis, macro generation, filtering, and application to simplify complex planning tasks.
• Experimental results on benchmarks like Satellite and PSR show notable reductions in search space and CPU time, demonstrating practical benefits.

Improving AI Planning with Automatically Learned Macro-Operators

The paper "Macro-FF: Improving AI Planning with Automatically Learned Macro-Operators" presents a methodological advancement in the domain of automated AI planning. The authors, Botea et al., propose an approach that enhances AI planners by leveraging macro-operators to simplify and expedite plan generation. This technique is particularly effective in addressing the challenges posed by complex and demanding benchmarks that have historically been significant obstacles for planners.

Overview

The authors describe two distinct methods for generating macro-operators: Component Abstraction Enhanced Domain (CA-ED) and Solution-Enhanced Planner (SOL-EP). Both employ a shared four-stage framework: domain analysis, macro-operator generation, filtering, and practical application in planning. The key differentiator between CA-ED and SOL-EP is how they interact with domain definitions and problem instances—CA-ED modifies the problem domain itself, while SOL-EP enhances the planning process using the extracted macros.

Detailed Methodology

Component Abstraction - Enhanced Domain (CA-ED): This strategy generates macro-operators through analysis of the structural components of domain problems. It primarily focuses on extracting reusable structural elements from static domain features, thereby creating a compact set of macro-operators. This process is characterized by the transformation of the initial Problem Domain Description Language (PDDL) domain into an enhanced domain where macros are integrated as new operators. The CA-ED approach is limited to STRIPS domains with static facts due to the dependence on static predicates for component abstraction.

Solution-Enhanced Planner (SOL-EP): Addressing the constraints of CA-ED, SOL-EP extracts macros directly from the solution paths of training problem instances. This approach does not modify the domain but instead augments the planner with code that accommodates macros, facilitating their use during planning. The SOL-EP method demonstrates versatility by supporting more complex domains, including those described using a subset of the Action Description Language (ADL).

Numerical Results and Experimental Analysis

In competitive evaluations, the Macro-FF system, built upon the FF planner, demonstrated significant reductions in search complexity across various international planning competition domains. Notable results were observed in domains like Satellite, PSR, and Promela Optical Telegraph, where the use of macro-operators reduced both the search space and CPU time.

The implications of this work extend to both theoretical and practical facets of AI planning. Theoretically, the research enhances understanding of domain structure exploitation as a mechanism for accelerating AI planning. Practically, it offers a robust framework that planners can implement to improve efficiency.

Implications and Future Developments

While existing methods, such as Hierarchical Task Networks (HTNs) and temporal logic control rules, rely on extensive human intervention and domain-specific input, Macro-FF provides an automatic and adaptive alternative. The potential future trajectory of this research could explore the translation of similar methodologies to other complex AI domains, including resource allocation or multi-agent systems. Additionally, extending these learning techniques to generate hierarchical control structures like HTNs presents a valuable opportunity for future research.

In conclusion, the paper delivers an important contribution to AI planning, offering a methodology that intelligently leverages past experience to enhance planner efficiency. Continued investigation into automated domain learning and macro-operator integration holds promise for further optimizing AI systems in tackling increasingly sophisticated planning tasks.