TL;DR
This paper addresses the lack of well-developed generic inference models for knowledge-based planning within the CommonKADS methodology. By deriving a model from the established O-Plan system, the authors provide a reusable framework for knowledge acquisition and verification in planning tasks, demonstrated through its application to Royal Air Force Search and Rescue operations.
What Problem Does This Paper Address?
Building knowledge-based systems (KBS) often relies on structured methodologies like CommonKADS, which provides a library of generic inference models to guide the process. While CommonKADS has robust models for tasks like diagnosis, it lacks well-developed generic models for knowledge-based planning. This is a significant gap, as planning is a crucial and commercially important application of Artificial Intelligence. Existing generic planning models are often derived from existing AI systems, making them proven but potentially tied to specific architectures. The authors aim to bridge this gap by developing a generic inference model for planning tasks that is reusable and can be verified against real-world problems.
Methods
The core of this paper is the development of a CommonKADS generic inference model for planning, directly inspired by the Open Planning Architecture (O-Plan). O-Plan is a well-established, domain-independent planning system with an agent-oriented architecture. The authors map O-Plan's components and functionalities to the CommonKADS expertise model, specifically focusing on the inference knowledge level.
Key steps include:
- Mapping O-Plan Components to CommonKADS: The paper details how O-Plan's "domain information," "plan states," "knowledge sources," and "controller" map to CommonKADS' domain knowledge, inference knowledge (knowledge roles), and task knowledge.
- Deriving Inference Structures: The authors extract generic inference structures from O-Plan's "knowledge sources," which represent how O-Plan resolves planning issues. These structures are visualized and described, including:
- Resolving an issue by condition satisfaction (Figure 2).
- Resolving an issue by "backward chaining" (Figure 3).
- Resolving an issue by expansion into sub-issues (Figure 4).
- Verification: The derived generic models were verified against a real-life planning task: planning for Royal Air Force (RAF) Search and Rescue (SAR) operations. This involved comparing the generic models with a domain-specific inference structure developed for the SAR system.
Datasets
Not discussed in the paper. The paper focuses on deriving models from an existing system (O-Plan) and verifying them against a real-world application domain (RAF Search and Rescue operations), rather than training or evaluating on specific datasets.
Key Insights and Ideas
- O-Plan as a Blueprint for Generic Planning: The paper demonstrates that a mature and well-understood planning system like O-Plan can serve as an excellent basis for deriving generic inference models within a structured methodology like CommonKADS.
- Bridging the Gap in CommonKADS: The work directly addresses a recognized deficiency in the CommonKADS library, providing much-needed generic models for the important domain of knowledge-based planning.
- Verification through Real-World Application: Applying the derived generic models to a complex real-world problem like RAF SAR operations provides strong evidence of their utility and helps identify areas where generic models can be enriched by domain-specific adaptations.
- Generic Models as Completeness Checks: The authors highlight that using generic inference models can act as a valuable completeness check for acquired procedural knowledge, prompting knowledge engineers to consider aspects of planning that might have been overlooked.
Key Results
The paper successfully derives a set of generic inference structures for knowledge-based planning based on O-Plan. These models were then applied to the RAF Search and Rescue planning task. The verification process revealed that:
- The generic models provided a richer representation of techniques for matching issues to activities than the initial domain-specific model.
- Important knowledge roles (resource constraints, library of possible activities) were explicitly identified by the generic models, which were not fully represented in the domain-derived structure.
- The domain-derived structure highlighted the importance of an "outline plan template" for the SAR domain, which was not present in the generic models, suggesting a need for domain-specific adaptations.
- The implemented KBS for RAF SAR planning incorporated the strengths of both the generic inference structure and the domain-specific template, demonstrating a successful integration.
Limitations and Future Work
- Scope of Genericity: The authors acknowledge that the derived models are based on O-Plan's specific approach to planning. They suggest that knowledge-based planning is a broad field with diverse approaches, and these models may not cover every aspect or every type of planning task.
- Need for Domain-Specific Adaptations: While generic models are valuable, they often require adaptation to specific domains, as seen with the "outline plan template" for SAR.
- Future Work: The authors propose a "top-down approach to classifying planning tasks" that identifies the characteristics of different planning approaches. This would allow for the creation of more truly "generic" models that specify types of operations, which can then be combined with "domain-derived" models (like those presented in this paper) for verification and enrichment. They also suggest incorporating other proposed generic planning models into a common framework for rational selection.
Summary updated