Abstract
Knowledge Representation (KR) fragmentation poses a significant challenge for the validation and verification (V&V) of expert systems. This paper introduces GKR (Generic Knowledge Representation), a unified framework designed to bridge disparate KR formalisms. By integrating static domain knowledge, dynamic execution traces, and design intent, GKR provides a comprehensive model for developing general-purpose V&V tools that maintain the semantic integrity of the original systems.
Introduction
A foundational goal in early artificial intelligence was the separation of domain knowledge (facts) from the procedures (logic) that utilize them. However, as the field matured, researchers developed a wide array of specialized Knowledge Representation formalisms. This "Babel" of frameworks created a significant bottleneck: a Knowledge-Based System (KBS) built in one framework could not easily utilize tools designed for another.
This lack of interoperability is particularly problematic for Validation and Verification (V&V). To ensure an AI system is accurate and reliable, engineers require a common ground. Historically, translating a specific system into a generic format often resulted in the loss of nuanced logic. GKR is proposed as a "universal language" to solve this translation problem, enabling the use of standardized V&V tools across diverse architectures.
The GKR Framework
GKR improves upon previous generic models, such as CCR and CKR, by offering a more holistic representation of a Knowledge-Based System. The model decomposes a KBS into three distinct, interconnected layers:
- The Static Part (The "What"): This layer represents the core domain knowledge or the Knowledge Base. GKR utilizes a hybrid approach, combining frames (structured objects) with rules.
- The Dynamic Part (The "How"): This layer captures execution traces. It serves as a functional log of how the system behaves when processing specific test cases or real-world problems.
- The Design Particularities (The "Why"): This layer represents control information. It explains the underlying design intent and the logic that dictates why static knowledge produces specific dynamic results during execution.
Technically, GKR enhances standard frame-and-rule systems by incorporating features often omitted in simpler models, such as user-defined facets, explicit inheritance rules, and the ability for rules to access and modify any component of the Knowledge Base.
Research Context and Requirements
As a theoretical model for representation, GKR was developed by analyzing the requirements of several major European AI projects:
- VALID: Focused on the validation and verification of expert systems.
- ACKNOWLEDGE: Aimed at creating a comprehensive "Knowledge Engineering Workbench."
- PIVOT: An automatic translation project utilizing an "interlingua" (language-independent) representation.
Key Insights
- The Triple Representation Requirement: A primary insight of GKR is that representing facts (static knowledge) is insufficient for rigorous validation. To truly verify a system, one must also represent its behavior (dynamic) and its intent (control).
- Non-Hierarchical Flexibility: GKR acknowledges that real-world expert knowledge rarely fits into a neat tree structure. It allows for complex, non-hierarchical relations between frames, providing the flexibility needed for sophisticated domains.
- Integrated Reasoning: In GKR, rules are treated as first-class citizens. They are not merely "if-then" appendages but are integrated into the frame system with the authority to modify the entire structure, allowing for more complex reasoning patterns.
Conclusion and Future Work
GKR provides a formal architecture capable of capturing complex structures that previous models could not represent. At the time of its proposal, the model was being implemented by the Validation Group at the Universidad Politécnica de Madrid to establish a universal environment for V&V tools.
While GKR is comprehensive, its complexity presents a challenge for implementation. The authors identify the "translation tax"—the effort required to map a highly optimized expert system into a generic format—as an ongoing hurdle. However, by capturing the "Why" and "How" alongside the "What," GKR stands as a robust candidate for minimizing semantic loss during the translation process.
Summary updated