Abstract
This paper addresses a fundamental challenge in Knowledge Representation (KR): the integration of absolute definitions with "common sense" rules that allow for exceptions. We present a formal, path-based theory for inheritance networks that enables strict links (absolute truths) and defeasible links (defaults) to coexist. By employing a skeptical reasoning approach and a refined definition of preemption (specificity), this framework allows AI systems to navigate complex taxonomies without collapsing into logical contradictions.
Introduction
In the late 1980s, AI researchers were divided between two paradigms for representing knowledge in "IS-A" hierarchies. Strict inheritance systems utilized classical logic, where absolute rules (e.g., "All birds fly") became logically inconsistent upon encountering a single exception (e.g., a penguin). Conversely, defeasible inheritance systems allowed for defaults but struggled to represent definitional truths—for instance, the fact that a square is always a rectangle, not just typically one. This paper proposes a unified theory that weaves these two approaches into a single "mixed" network.
Methods: A Path-Based Approach
The authors utilize a directed graph representation where nodes represent properties or individuals, and edges (links) represent relationships. The framework defines four primary link types:
- Positive Strict ($\Rightarrow$): Absolute inclusion.
- Negative Strict ($\not\Rightarrow$): Absolute exclusion.
- Positive Defeasible ($\rightarrow$): Default inclusion.
- Negative Defeasible ($\not\rightarrow$): Default exclusion.
Skeptical Reasoning and Preemption
To resolve conflicts, such as the famous "Nixon Diamond," the system employs skeptical reasoning: it refuses to draw a conclusion unless one argument is clearly superior. Superiority is determined via preemption (specificity), where information from a more specific node overrides information from a more general ancestor.
Key Insights: Defining Directness
A core contribution is the realization that strict links fundamentally change how "direct" information is defined. The authors argue that a defeasible link followed by a chain of strict links should be treated as a relatively direct argument. This formalization of "permitted paths" provides a blueprint for inference engines to navigate mixed networks while maintaining logical integrity.
Results and Discussion
The paper successfully merges strict and defeasible inheritance into a coherent mathematical framework. By localizing inconsistency within the graph structure, the system avoids the "Principle of Explosion" found in classical logic, where a single contradiction can invalidate an entire knowledge base.
However, limitations remain. The theory is primarily restricted to acyclic networks, as defeasible cycles create significant complications for defining specificity. Furthermore, while the framework handles simple "IS-A" links effectively, integrating it with complex conceptual definitions and ensuring computational efficiency in massive networks are subjects for future research.
Summary updated