Abstract
This paper, originally presented at AAAI-2021, introduces Disjunctive Rational Closure (DRC), a novel framework for non-monotonic reasoning. While standard Rational Closure (RC) relies on the often-criticized Rational Monotonicity property, DRC utilizes interval-based typicality models to satisfy the more intuitive Disjunctive Rationality (DR) rule. We demonstrate that DRC provides a robust, base-driven approach to conditional inference that maintains computational feasibility while addressing the limitations of traditional ranked semantics.
Introduction
In Artificial Intelligence, defeasible reasoning allows systems to draw "normal" conclusions that can be retracted upon receiving new information. For instance, a system might assume birds fly unless it learns the bird in question is a penguin. The current "gold standard" for such reasoning is Rational Closure (RC). However, RC is built upon Rational Monotonicity (RM), a property that can be overly "venturous," forcing systems to maintain conclusions even when irrelevant new information is introduced.
The authors propose a shift toward Disjunctive Rationality (DR). DR states that if a conclusion follows from a disjunction (e.g., "A or B"), it should logically follow from at least one of the individual components. While DR is widely accepted, a standard "simplest" inference method based solely on it has been missing until now.
Methodology: Interval-Based Typicality
To overcome the limitations of RC, the authors move away from "ranked semantics," where every possible world is assigned a single typicality rank. Instead, they introduce interval-based interpretations. In this model, each world is assigned an interval $[L, U]$ representing a range of typicality.
The construction of DRC involves two primary components:
- Lower Rank ($L$): Derived from the standard Rational Closure.
- Upper Rank ($U$): Calculated based on the specific rules in the Knowledge Base (KB) that a world "verifies."
This "base-driven" approach means the specific syntax of the KB matters; explicitly stated rules exert more influence than derived ones. Preference between worlds is only established if one interval is strictly lower than another, allowing the system to represent "incomparability" when evidence is insufficient.
Theoretical Insights and Trade-offs
A key contribution of the paper is an "impossibility theorem" proving that no inference method can satisfy all desirable logical properties simultaneously. Specifically, one must choose between properties like Vacuity and Cautious Monotonicity. The authors prioritize Cautious Monotonicity—ensuring that learning expected information doesn't disrupt existing conclusions—over Vacuity.
The use of intervals allows the system to remain conservative. By treating overlapping intervals as incomparable, DRC avoids the unwarranted inferences sometimes produced by the strict ordering of Rational Closure.
Results and Examples
The paper evaluates DRC using illustrative toy examples, such as a scenario involving COVID-19 safety measures (masks and social distancing).
The authors demonstrate that DRC satisfies core desiderata, including Inclusion, DR, and Cautious Monotonicity. Furthermore, while the construction process involves exponential growth relative to input size, its complexity remains within the same class as standard RC, ensuring it is viable for offline knowledge base compilation.
Conclusion
Disjunctive Rational Closure offers a flexible alternative to traditional Rational Closure. By embracing incomparability through interval-based models, it provides a more nuanced approach to handling uncertainty in complex logical systems. Future work will focus on refining syntax-independent versions of the framework and defining optimal preference relations within the space of interval models.
Summary updated