Abstract
In classical logic, the addition of new information cannot invalidate previously held conclusions—a property known as monotonicity. However, human reasoning is inherently non-monotonic; we frequently make assumptions based on typicality and retract them when faced with specific exceptions. This paper, originally presented by Chern H. Seet in 1987, proposes a framework for default reasoning that treats exceptions as explicit syntactic elements. By utilizing a modal operator and a structured belief revision strategy, the system allows for the refinement of knowledge bases without the loss of general rules, providing a transparent alternative to contemporary methods like Circumscription or Default Logic.
Introduction
The challenge of "jumping to conclusions" is central to artificial intelligence. If a system is told that Tweety is a bird, it should ideally conclude that Tweety can fly. However, if it later learns Tweety is a penguin, a classical logic system faces a contradiction. To resolve this, researchers in the 1980s developed non-monotonic logics. While influential, many of these systems—such as Reiter’s Default Logic—were criticized for hiding the mechanics of exceptions or requiring entirely new logical foundations.
Seet’s approach seeks to bridge this gap by keeping exceptions "first-class citizens" within the syntax of the logic. The goal is to create a system where a standard theorem prover can manage defaults by explicitly tagging and revising exceptions during the reasoning process.
The Formal Framework
The proposed system extends standard first-order logic with a modal operator, $\Box$. This operator is used to mark specific terms as exceptions to general rules. The reasoning agent operates within a sequence of "worlds," transitioning from a set of Current Beliefs ($B_C$) to Future Beliefs ($B_F$) as new data is acquired.
Two Types of Consequence
The framework distinguishes between two modes of inference:
- Strict Consequence ($\vdash$): Standard logical provability. If a fact is explicitly stated or logically follows from the axioms, it is strictly true.
- Default Consequence ($\circ \vdash$): A meta-inference approach. To prove a statement by default, the system attempts to disprove the existence of an exception. If the system fails to prove that an exception (tagged with $\Box$) applies, it assumes the general rule holds.
Belief Revision as Refinement
Unlike traditional belief revision, which often involves deleting contradictory information to maintain consistency, Seet’s strategy focuses on refinement. When a new belief ($\beta$) contradicts an existing belief set ($B_C$), the system transforms the old belief into a weaker version.
For example, if the system believes "All birds fly" and learns "Penguins do not fly," it does not discard the first rule. Instead, it updates the rule to: "Birds fly unless they are penguins." Syntactically, the penguin exception is wrapped in the $\Box$ operator. This ensures that the knowledge base remains consistent while preserving the maximum utility of the original general rule.
Modified Resolution
To automate this, the paper outlines a modified resolution refutation process. When the theorem prover encounters a $\Box$-tagged term, it "spawns" a sub-proof to check if that exception can be refuted. This effectively automates the "unless" clause of default reasoning within a standard proof structure.
Key Insights and Results
The primary contribution of this work is the transparency of its exceptions. By making exceptions syntactically explicit, the knowledge base becomes easier to audit and update.
The Tweety Test and Nested Exceptions
The paper demonstrates the system's efficacy through classic logical puzzles:
- The Tweety Test: The system correctly concludes Tweety flies when known only as a bird, but retracts this conclusion immediately upon learning Tweety is a penguin.
- Nested Exceptions (The "Undo" Property): The framework handles complex hierarchies. If the system learns that "Super-penguins" are a rare breed of penguin that can fly, it can revise the "Penguins don't fly" rule to include "Super-penguins" as an exception. This restores the original "Birds fly" conclusion specifically for Super-penguins, demonstrating the system's ability to handle deep layers of contradictory information.
Limitations and Conclusion
While Seet’s strategy offers significant clarity, it introduces practical challenges. The most notable is formula bloat: as a system encounters more exceptions, its logical formulas grow increasingly long and complex. Furthermore, the computational cost of spawning sub-proofs for every $\Box$-term could be prohibitive for large-scale applications.
Despite these hurdles, the paper provides a robust theoretical foundation for viewing default reasoning not as a deviation from logic, but as a structured process of belief refinement. It remains a significant example of how explicit syntax can be used to manage the inherent uncertainty of real-world knowledge.
Summary updated