Introduction
This paper introduces Belief-Function Logic (BFL), a hybrid framework that integrates quantified degrees of belief—derived from Dempster-Shafer theory—with classical first-order logic (FOL). BFL provides a formal bridge between rigorous logical reasoning and the mathematical handling of uncertain, incomplete, or partially inconsistent information.
The Challenge of Categorical Logic
Traditional first-order logic is inherently "categorical": a statement is either true or false. However, real-world knowledge is rarely so binary. While various models for uncertainty exist—such as probability theory and the Dempster-Shafer (D-S) theory of belief functions—they are often disconnected from the expressive power of logic.
Most existing "uncertain logics" either treat uncertainty as a meta-level property or restrict themselves to simple propositional statements. BFL aims to be a coherent hybrid that allows for partial belief and ignorance while maintaining the full expressive power of FOL, including quantifiers like "for all" ($\forall$) and "there exists" ($\exists$).
The BFL Framework: Syntax and Semantics
BFL is defined through three main pillars: syntax, semantics, and deduction.
Syntax
A BFL formula is expressed as $F:[a,b]$, where $F$ is a standard first-order sentence. The interval $[a,b]$ represents the belief: $a$ is the strength of belief in $F$ being true, and $(1-b)$ is the strength of belief in $F$ being false. The gap $(b-a)$ represents "ignorance" or uncommitted belief.
Semantics
The framework uses "hyper-interpretations"—sets of possible worlds—to model incompleteness. A belief function is defined over these sets, allowing the logic to satisfy properties like monotonicity and a graded form of deductive closure. Logical sentences serve to partition the space of interpretations into regions where specific beliefs can be assigned.
Automated Reasoning and Deduction
To make the logic practical, the paper introduces D-models, which are "least informative" models based on Dempster’s rule of combination. For automated reasoning, BFL translates logical formulas into "uncertainty networks" or valuation systems. This allows belief values to be computed through local propagation, often utilizing resolution-based techniques to derive belief intervals for complex queries.
Key Insights: Beyond Classical Logic
- Graded Logical Properties: BFL generalizes classical logical rules. For instance, it supports a "graded" modus ponens: if you believe $P \to Q$ to degree $a$ and $P$ to degree $c$, you can believe $Q$ to at least degree $a \times c$.
- Logical Interpretation of Dempster-Shafer: A significant insight is that Dempster’s combination rule emerges naturally as the mechanism for building the "least informative" model that satisfies a set of constraints.
- Handling Inconsistency: Unlike standard FOL, which becomes useless if it encounters a contradiction, BFL can handle "$\alpha$-inconsistency." It treats minor contradictions as noise, allowing the system to remain functional.
- Epistemic Focus: BFL models epistemic uncertainty (partial belief in a complete truth) rather than statistical frequency.
Results and Future Directions
The paper proves that BFL is a conservative extension of FOL; if intervals are restricted to $[1,1]$ (true) or $[0,0]$ (false), it behaves exactly like classical logic. It also provides a sound and complete deduction procedure for "D-consistent" sets of formulas, making belief-function reasoning computationally feasible in complex domains.
While BFL is currently a monotonic logic and cannot handle default reasoning where new information retracts previous conclusions, it establishes a robust foundation. Future work may extend this framework to non-monotonic reasoning or other uncertainty formalisms like possibility theory.
Summary updated